<query>
<top>
<paper_num> 1 </paper_num>
<paper_title>   A Semi-Supervised Feature Clustering Algorithm with Application to Word Sense Disambiguation.  </paper_title>
<paper_abstract>   In this paper we investigate an application of feature clustering for word sense disambiguation, and propose a semisupervised feature clustering algorithm. Compared with other feature clustering methods (ex. supervised feature clustering), it can infer the distribution of class labels over (unseen) features unavailable in training data (labeled data) by the use of the distribution of class labels over (seen) features available in training data. Thus, it can deal with both seen and unseen features in feature clustering process. Our experimental results show that feature clustering can aggressively reduce the dimensionality of feature space, while still maintaining state of the art sense disambiguation accuracy. Furthermore, when combined with a semi-supervised WSD algorithm, semi-supervised feature clustering outperforms other dimensionality reduction techniques, which indicates that using unlabeled data in learning process helps to improve the performance of feature clustering and sense disambiguation. 1  </paper_abstract>
<query_num> 101 </query_num>
<text>   =-=e National University of Singapore 3 Science Drive 2 117543 Singapore tancl@comp.nus.edu.sg Yarowsky, 1995-=-), unsupervised learning algorithms (or word sense discrimination) (=-=Pedersen and Bruce, 1997; Schütze, 1998-=-), and knowledge based algorithms (=-=Lesk, 1986; McCarthy et al., 2004-=-). In general, the most common approaches start by evaluating the co-occurrence matrix of features versus contexts of instances of a and =-=Ng, 2002-=-), but feature selection does not improve SVM and Adaboost over SENSEVAL-1 and SENSEVAL-2 data (=-=Lee and Ng, 2002-=-) for word sense disambiguation. Latent semantic indexing (LSI) studied in (=-=Schütze, 1998-=-) improves the performance of sense discrimination, while unsupervised feature selection also improves the performance of word sense discrimination (=-=Niu et al., 2004-=-). But little work is done on using   </text>
<query_num> 102 </query_num>
<text>   =-=nd Li, 2004; Mihalcea, 2004; Niu et al., 2005; Park et al., 2000; Chew Lim Tan Department of Computer Science National University of Singapore 3 Science Drive 2 117543 Singapore tancl@comp.nus.edu.sg Yarowsky, 199-=-5), unsupervised learning algorithms (or word sense discrimination) (=-=Pedersen and Bruce, 1997; Schütze, 1998-=-), and knowledge based algorithms (=-=Lesk, 1986; McCarthy et al., 2004-=-). In general, the most c   </text>
<query_num> 103 </query_num>
<text>   based statistical methods have been proposed to solve this problem, including supervised learning algorithms (=-=Leacock et al., 1998; Towel and Voorheest, 1998-=-), weakly supervised learning algorithms (=-=Dagan and Itai, 1994; Li and Li, 2004; Mihalcea, 2004; Niu et al., 2005; Park et al., 2000; Chew Lim Tan Department of Computer Science National University of Singapore 3 Science Drive 2 117543 Singapore tancl@comp.nus.e-=-   </text>
<query_num> 104 </query_num>
<text>   compress the feature space much more aggressively while still maintaining state of the art classification accuracy. In the context of document clustering, unsupervised feature clustering algorithms (=-=Dhillon, 2001; Dhillon et al., 2002; Dhillon et al., 2003; El-Yaniv and Souroujon, 2001; Slonim and Tishby, 2000-=-) perform word clustering by the use of word-document co-occurrence matrix, which can improve the per  and Tishby (2001) goes further to show that when the training sample is small, word clusters can yield significant improvement in classification accuracy. Unsupervised feature clustering algorithms (=-=Dhillon, 2001; Dhillon et al., 2002; Dhillon et al., 2003; El-Yaniv and Souroujon, 2001; Slonim and Tishby, 2000-=-) perform word clustering by the use of word-document co-occurrence matrix, which do not utilize clas   </text>
<query_num> 105 </query_num>
<text>   eature space much more aggressively while still maintaining state of the art classification accuracy. In the context of document clustering, unsupervised feature clustering algorithms (=-=Dhillon, 2001; Dhillon et al., 2002; Dhillon et al., 2003; El-Yaniv and Souroujon, 2001; Slonim and Tishby, 2000-=-) perform word clustering by the use of word-document co-occurrence matrix, which can improve the performance of document c 01) goes further to show that when the training sample is small, word clusters can yield significant improvement in classification accuracy. Unsupervised feature clustering algorithms (=-=Dhillon, 2001; Dhillon et al., 2002; Dhillon et al., 2003; El-Yaniv and Souroujon, 2001; Slonim and Tishby, 2000-=-) perform word clustering by the use of word-document co-occurrence matrix, which do not utilize class labels to guide clus   </text>
<query_num> 106 </query_num>
<text>   hods have been proposed to solve this problem, including supervised learning algorithms (=-=Leacock et al., 1998; Towel and Voorheest, 1998-=-), weakly supervised learning algorithms (=-=Dagan and Itai, 1994; Li and Li, 2004; Mihalcea, 2004; Niu et al., 2005; Park et al., 2000; Chew Lim Tan Department of Computer Science National University of Singapore 3 Science Drive 2 117543 Singapore tancl@comp.nus.edu.sg Yarowsky, 1-=-  </text>
<query_num> 107 </query_num>
<text>   ill maintaining state of the art classification accuracy. In the context of document clustering, unsupervised feature clustering algorithms (=-=Dhillon, 2001; Dhillon et al., 2002; Dhillon et al., 2003; El-Yaniv and Souroujon, 2001; Slonim and Tishby, 2000-=-) perform word clustering by the use of word-document co-occurrence matrix, which can improve the performance of document clustering by clustering documents over word clusters : supervised feature clustering (SuFC) (=-=Baker and McCallum, 1998; Bekkerman et al., 2003; Slonim 1 Available at http://www.senseval.org/senseval3sand Tishby, 2001-=-), iterative double clustering (IDC) (=-=El-Yaniv and Souroujon, 2001-=-), semi-supervised feature clustering (SemiFC) (our algorithm), supervised feature selection (SuFS) (=-=Forman, 2003-=-), and latent semantic indexing (LSI) (=-=Deerwester et. al., 1990-=-) 2 . We used sIB algori �X clusters, where X is represented as distribution over � F . Subsequent IDC iterations use all the unlabeled data. This IDC algorithm can result in two clustering solutions: � F and � X. Following (=-=El-Yaniv and Souroujon, 2001-=-), the number of iterations is set as 15, and N �X = |S| (the number of senses of target word) in our re-implementation of IDC. After performing IDC, examples can be represented as vectors over featur ing sample is small, word clusters can yield significant improvement in classification accuracy. Unsupervised feature clustering algorithms (=-=Dhillon, 2001; Dhillon et al., 2002; Dhillon et al., 2003; El-Yaniv and Souroujon, 2001; Slonim and Tishby, 2000-=-) perform word clustering by the use of word-document co-occurrence matrix, which do not utilize class labels to guide clustering process. Slonim and Tishby (2000), El-Yaniv a   </text>
<query_num> 108 </query_num>
<text>   indexing, and unsupervised feature clustering when only unlabeled data is available. Supervised feature selection improves the performance of an examplar based learning algorithm over SENSEVAL2 data (=-=Mihalcea, 2002-=-), Naive Bayes and decision tree over SENSEVAL-1 and SENSEVAL-2 data (=-=Lee and Ng, 2002-=-), but feature selection does not improve SVM and Adaboost over SENSEVAL-1 and SENSEVAL-2 data (=-=Lee and Ng, 2002-=-)   </text>
<query_num> 109 </query_num>
<text>   m, the test set in SENSEVAL-3 data was also used as unlabeled data in tranductive learning process. We investigated two distance measures for LP: cosine similarity and Jensen-Shannon (JS) divergence (=-=Lin, 1991-=-). Cosine similarity measures the angle between two feature vectors, while JS divergence measures the distance between two probability distributions if each feature vector is considered as probability   </text>
<query_num> 110 </query_num>
<text>   rt classification accuracy. In the context of document clustering, unsupervised feature clustering algorithms (=-=Dhillon, 2001; Dhillon et al., 2002; Dhillon et al., 2003; El-Yaniv and Souroujon, 2001;Slonim and Tishby, 2000-=-) perform word clustering by the use of word-document co-occurrence matrix, which can improve the performance of document clustering by clustering documents over word clusters. Supervised feature clus S = (I − αL) −1 Y F,S hard . 8. Obtain the feature clustering solution � F by clustering the rows of � Y F,S i into N�F groups. In this paper we use sequential information bottleneck (sIB) algorithm (=-=Slonim and Tishby, 2000-=-) to perform clustering analysis. End Step 3 ∼ 5 are the process to obtain hard labels for features in F , while the operation in step 6 and 7 is a local and global consistency based semisupervised le s our re-implementation of supervised feature clustering. After feature clustering, examples can be represented as vectors over feature clusters. IDC is an extension of double clustering method (DC) (=-=Slonim and Tishby, 2000-=-), which performs iterations of DC. In the transductive version of IDC, they cluster features in F as distributions over class labels (given by the labeled data) during the first stage of the IDC firs ters can yield significant improvement in classification accuracy. Unsupervised feature clustering algorithms (=-=Dhillon, 2001; Dhillon et al., 2002; Dhillon et al., 2003; El-Yaniv and Souroujon, 2001;Slonim and Tishby, 2000-=-) perform word clustering by the use of word-document co-occurrence matrix, which do not utilize class labels to guide clustering process. Slonim and Tishby (2000), El-Yaniv and Souroujon (2001) and D word clusters can improve the performance of document clustering. El-Yaniv and Souroujon (2001) present an iterative double clustering (IDC) algorithm, which performs iterations of double clustering (=-=Slonim and Tishby, 2000-=-). Furthermore, they extend IDC algorithm for semi-supervised learning when given both labeled and unlabeled data. Our algorithm belongs to the family of semisupervised feature clustering techniques,   </text>
<query_num> 111 </query_num>
<text>   t semantic indexing (LSI) studied in (=-=Schütze, 1998-=-) improves the performance of sense discrimination, while unsupervised feature selection also improves the performance of word sense discrimination (=-=Niu et al., 2004-=-). But little work is done on using feature clustering to conduct dimensionality reduction for WSD. This paper will describe an application of featuresclustering technique to WSD task. Feature cluster   </text>
<query_num> 112 </query_num>
<text>   task. Feature clustering has been extensively studied for the benefit of text categorization and document clustering. In the context of text categorization, supervised feature clustering algorithms (=-=Baker and McCallum, 1998; Bekkerman et al., 2003; Slonim and Tishby, 2001-=-) usually cluster words into groups based on the distribution of class labels over features, which can compress the feature space much more aggressivel mensionality reduction algorithms on the data in English lexical sample (ELS) task of SENSEVAL-3 (=-=Mihalcea et al., 2004-=-)(including all the 57 English words ) 1 : supervised feature clustering (SuFC) (=-=Baker and McCallum, 1998; Bekkerman et al., 2003; Slonim 1 Available at http://www.senseval.org/senseval3sand Tishby, 2001-=-), iterative double clustering (IDC) (=-=El-Yaniv and Souroujon, 2001-=-), semi-supervised feature clusterin supervised factor analysis technique based on Singular Value Decomposition of a |X| × |F | example-feature matrix. The underlying technique for LSI is to find an orthogonal basis for the 2 Following (=-=Baker and McCallum, 1998-=-), we use LSI as a representative method for unsupervised dimensionality reduction. 3Available at http://www.cs.huji.ac.il/∼noamm/ feature-example space for which the axes lie along the dimensions of   and document clustering, which can be categorized as supervised feature clustering, semi-supervised feature clustering, and unsupervised feature clustering. Supervised feature clustering algorithms (=-=Baker and McCallum, 1998; Bekkerman et al., 2003; Slonim and Tishby, 2001-=-) usually cluster words into groups based on the distribution of class labels over features. Baker and McCallum (1998) apply supervised feature cluster   </text>
<query_num> 113 </query_num>
<text>   to assign an appropriate sense to an occurrence of a word in a given context. Many corpus based statistical methods have been proposed to solve this problem, including supervised learning algorithms (=-=Leacock et al., 1998; Towel and Voorheest, 1998-=-), weakly supervised learning algorithms (=-=Dagan and Itai, 1994; Li and Li, 2004; Mihalcea, 2004; Niu et al., 2005; Park et al., 2000; Chew Lim Tan Department of Computer Sci-=-   </text>
</top>
<top>
<paper_num> 2 </paper_num>
<paper_title>   Dynamic Local Search for the Maximum Clique Problem  </paper_title>
<paper_abstract>   In this paper, we introduce DLS-MC, a new stochastic local search algorithm for the maximum clique problem. DLS-MC alternates between phases of iterative improvement, during which suitable vertices are added to the current clique, and plateau search, during which vertices of the current clique are swapped with vertices not contained in the current clique. The selection of vertices is solely based on vertex penalties that are dynamically adjusted during the search, and a perturbation mechanism is used to overcome search stagnation. The behaviour of DLS-MC is controlled by a single parameter, penalty delay, which controls the frequency at which vertex penalties are reduced. We show empirically that DLS-MC achieves substantial performance improvements over state-of-the-art algorithms for the maximum clique problem over a large range of the commonly used DIMACS benchmark instances. 1.  </paper_abstract>
<query_num> 201 </query_num>
<text>   applications, for example, information retrieval, experimental design, signal transmission and computer vision (=-=Balus &amp; Yu, 1986-=-). More recently, applications in bioinformatics have become important (=-=Pevzner &amp; Sze, 2000 Ji, Xu, &amp; Stormo, 2004;-=-). The search variant of MAX-CLIQUE can be stated as follows: Given an undirected graph G = (V,E), where V is the set of all vertices and E the set of all edges, find a maximum   </text>
<query_num> 202 </query_num>
<text>   ations divided by two times the maximum clique size, for representative runs of DLS-MC on instances C1000.9 and brock800 1 (this mobility measure is closely related to those used in previous studies (=-=Schuurmans &amp; Southey, 2000-=-)). As can be seen from Figure 11, there is a large difference in mobility between the two variants of the perturbation mechanism for pd = 1 and pd &amp;gt; 1; the former restarts the search from a randomly   </text>
<query_num> 203 </query_num>
<text>   d cliques of size |V | 1−ǫ for any ǫ &amp;gt; 0, unless N P = ZPP (=-=H˚astad, 1999-=-). 1 The best polynomial-time approximation algorithm for MAX-CLIQUE achieves an approximation ratio of O(|V |/(log |V |) 2 ) (=-=Boppana &amp; Halldórsson, 1992-=-). Therefore, large and hard instances of MAX-CLIQUE are typically solved using heuristic approaches, in particular, 1. ZPP is the class of problems that can be solved in expected polynomial time by a   </text>
<query_num> 204 </query_num>
<text>   he figure at a standard confidence level of α = 0.05 with p-values between 0.16 and 0.62). This observation is consistent with similar results for other highperformance SLS algorithms, e.g., for SAT (=-=Hoos &amp; Stützle, 2000-=-) and scheduling problems (=-=Watson, Whitley, &amp; Howe, 2005-=-). As a consequence, performing multiple independent runs of DLS-MC in parallel will result in close-to-optimal parallelisation speedup (=-=Hoos &amp; -=-  </text>
<query_num> 205 </query_num>
<text>   is inspired by the dynamic weights in DAGS and, more generally, by current state-of-the-art Dynamic Local Search (DLS) algorithms for other well-known combinatorial problems, such as SAT and MAX-SAT (=-=Hutter, Tompkins, &amp; Hoos, 2002; Tompkins &amp; Hoos, 2003; Thornton, Pham, Bain, &amp; Ferreira, 2004; Pullan &amp; Zhao, 2004-=-); for a general introduction to DLS, see also the work of (=-=Hoos &amp; Stützle, 2004-=-). Our results therefore provide fur   </text>
<query_num> 206 </query_num>
<text>   rior to QUALEX-MS for most of the MAX-CLIQUE instances from the DIMACS benchmark sets, but for some hard instances it does not reach the performance of RLS (=-=Grosso et al., 2004-=-). The k-opt algorithm (=-=Katayama, Hamamoto, &amp; Narihisa, 2004-=-) is based on a conceptually simple variable depth search procedure that uses elementary search steps in which a vertex is added to or removed from the current clique; while there is some evidence tha ementary search steps in which a vertex is added to or removed from the current clique; while there is some evidence that it performs better than RLS on many instances from the DIMACS benchmark sets (=-=Katayama et al., 2004-=-), its performance relative to DAGS is unclear. Finally, Edge-AC+LS (=-=Solnon &amp; Fenet, 2004-=-), a recent ant colony optimisation algorithm for MAX-CLIQUE that uses an elitist subsidiary local search proce ther NI(C) is no longer empty or when NL(C) becomes empty. Also, in order to reduce the incidence of unproductive plateau search phases, DLS-MC implements the plateau search termination condition of (=-=Katayama et al., 2004-=-) by recording the current clique (C ′ ) at the start of the plateau search phase and terminating plateauSearch when there is no overlap between the recorded clique C ′ and the current clique C. At th n particular, we compared DLS-MC with the following MAX-CLIQUE algorithms: DAGS (=-=Grosso et al., 2004-=-), GRASP (=-=Resende, Feo, &amp; Smith, 1998-=-) (using the results contained in =-=Grosso et al., 2004-=-), k-opt (=-=Katayama et al., 2004-=-), RLS (=-=Battiti &amp; Protasi, 2001-=-), GENE (=-=Marchiori, 2002-=-), ITER (=-=Marchiori, 2002-=-) and QUALEX-MS (=-=Busygin, 2002-=-). To rank the performance of MAX-CLIQUE algorithms and to determine the dominant algorithm sults of this experiment, shown in Table 5, DLS-MC dominates DAGS on all but one instance (the exception being san1000). Table 6 shows performance results for DLS-MC as compared to results for k-opt (=-=Katayama et al., 2004-=-), GENE (=-=Marchiori, 2002-=-), ITER (=-=Marchiori, 2002-=-) and RLS (=-=Battiti &amp; Protasi, 2001-=-) from the literature. To roughly compensate for differences in CPU speed, we scaled the CPU times for k-opt, GENE and   </text>
<query_num> 207 </query_num>
<text>   seems that there are five heuristic MAX-CLIQUE algorithms that achieve state-of-the-art performance. Reactive Local Search (RLS) (=-=Battiti &amp; Protasi, 2001-=-) has been derived from Reactive Tabu Search (=-=Battiti &amp; Tecchiolli, 1994-=-), an advanced and general tabu search method that automatically adapts the tabu tenure parameter (which controls the amount of diversification) during the search process; RLS also uses a dynamic rest  adjusting DLS-MC’s penalty delay parameter during the search, similar to the scheme used for dynamically adapting the tabu tenure parameter in RLS (=-=Battiti &amp; Protasi, 2001-=-) and Reactive Tabu Search (=-=Battiti &amp; Tecchiolli, 1994-=-), or the mechanism used for controlling the noise parameter in Adaptive Novelty + (=-=Hoos, 2002-=-). Finally, given the excellent performance of DLS-MC on standard MAX-CLIQUE instances reported here sugge   </text>
<query_num> 208 </query_num>
<text>   ting the tabu tenure parameter in RLS (=-=Battiti &amp; Protasi, 2001-=-) and Reactive Tabu Search (=-=Battiti &amp; Tecchiolli, 1994-=-), or the mechanism used for controlling the noise parameter in Adaptive Novelty + (=-=Hoos, 2002-=-). Finally, given the excellent performance of DLS-MC on standard MAX-CLIQUE instances reported here suggests that the underlying dynamic local search method has substantial potential to provide the b   </text>
<query_num> 209 </query_num>
<text>   ts in DAGS and, more generally, by current state-of-the-art Dynamic Local Search (DLS) algorithms for other well-known combinatorial problems, such as SAT and MAX-SAT (=-=Hutter, Tompkins, &amp; Hoos, 2002; Tompkins &amp; Hoos, 2003; Thornton, Pham, Bain, &amp; Ferreira, 2004; Pullan &amp; Zhao, 2004-=-); for a general introduction to DLS, see also the work of (=-=Hoos &amp; Stützle, 2004-=-). Our results therefore provide further evidence for the e   </text>
<query_num> 210 </query_num>
<text>   ults reported by Grosso et al. (=-=Grosso, Locatelli, &amp; Croce, 2004-=-), it seems that there are five heuristic MAX-CLIQUE algorithms that achieve state-of-the-art performance. Reactive Local Search (RLS) (=-=Battiti &amp; Protasi, 2001-=-) has been derived from Reactive Tabu Search (=-=Battiti &amp; Tecchiolli, 1994-=-), an advanced and general tabu search method that automatically adapts the tabu tenure parameter (which controls the amount of  -MC with the following MAX-CLIQUE algorithms: DAGS (=-=Grosso et al., 2004-=-), GRASP (=-=Resende, Feo, &amp; Smith, 1998-=-) (using the results contained in =-=Grosso et al., 2004-=-), k-opt (=-=Katayama et al., 2004-=-), RLS (=-=Battiti &amp; Protasi, 2001-=-), GENE (=-=Marchiori, 2002-=-), ITER (=-=Marchiori, 2002-=-) and QUALEX-MS (=-=Busygin, 2002-=-). To rank the performance of MAX-CLIQUE algorithms and to determine the dominant algorithm for each of our benchmark inst  instance (the exception being san1000). Table 6 shows performance results for DLS-MC as compared to results for k-opt (=-=Katayama et al., 2004-=-), GENE (=-=Marchiori, 2002-=-), ITER (=-=Marchiori, 2002-=-) and RLS (=-=Battiti &amp; Protasi, 2001-=-) from the literature. To roughly compensate for differences in CPU speed, we scaled the CPU times for k-opt, GENE and ITER by a factor of 0.91 (these had been obtained on a 2.0 GHz Pentium IV) and th research is to develop mechanisms for automatically adjusting DLS-MC’s penalty delay parameter during the search, similar to the scheme used for dynamically adapting the tabu tenure parameter in RLS (=-=Battiti &amp; Protasi, 2001-=-) and Reactive Tabu Search (=-=Battiti &amp; Tecchiolli, 1994-=-), or the mechanism used for controlling the noise parameter in Adaptive Novelty + (=-=Hoos, 2002-=-). Finally, given the excellent performance of DLS-M   </text>
</top>
<top>
<paper_num> 3 </paper_num>
<paper_title>   Abstract Behavior Types: a foundation model for components and their composition.  </paper_title>
<paper_abstract>   CWI is a founding member of ERCIM, the European Research Consortium for Informatics and Mathematics. CWI&amp;apos;s research has a theme-oriented structure and is grouped into four clusters. Listed below are the names of the clusters and in parentheses their acronyms.  </paper_abstract>
<query_num> 301 </query_num>
<text>   concrete instance of the application of the ABT model, we describe Reo: an exogenous coordination model wherein complex coordinators, called “connectors” are compositionally built out of simpler ones =-=[3, 4]-=-. Reo can be used as a glue language for compositional construction of connectors that orchestrate component instances in a component based system. We demonstrate the surprisingly expressive power of  e entities. Channel composition in Reo is a very powerful mechanism for construction of connectors. The expressive power of connector composition in Reo has been demonstrated through many examples in =-=[3, 4, 8]-=-. For instance, exogenous coordination patterns that can be expressed as (meta-level) regular expressions over I/O operations performed by component instances can be composed in Reo out of a small set e in Reo, especially because a channel can also have only two source ends or only two sink ends. A few examples of some such exotic channels appear in Section 9.3; even more examples are presented in =-=[3, 7, 4]-=-. Strictly speaking, Reo itself neither provides nor assumes the availability of any specific set 15sof channel types; it simply assumes that an appropriate assortment of channel types, each with its   〈β, b〉 ≡ a = b 9.4 Coordinating Glue Code To demonstrate the expressive power of connector composition, in this section we describe a number of examples in Reo. More examples are presented elsewhere =-=[3, 7, 8, 4]-=-. a o b,e,c f d a g b,e,c h,f,i a b c d 9.4.1 Write-Cue Regulator a d j b c b c a Sequencer a a c b b c Sequencer e f g Figure 3: Examples of connectors in Reo Consider the connector in Figure 3.a, co c consist of the first data item written to �a, followed by the first data item written to �b, followed by the second data item written to �a, followed by the second data item written to �b, etc. See =-=[3, 4]-=- for more detail and [8] for a formal treatment of this connector. The coordination pattern imposed by our connector can be summarized as c = (ab)∗, meaning the sequence of values that appear through   </text>
<query_num> 302 </query_num>
<text>   ditions and invariants. In general, this notion of components requires enhanced specification and verification techniques, as also observed by Hennicker and Wirsing [57, 24]. Our notion of components =-=[9, 6, 17]-=- uses channels as the basic inter-component communication mechanism. A channel is a point-to-point medium of communication with its own unique identity and two distinct ends. A channel supports transf   </text>
<query_num> 303 </query_num>
<text>   eo allows dynamic reconfiguration of connectors, even while they are being used by component instances. In this respect, Reo resembles dynamically reconfigurable generalized Kahn networks, as in IWIM =-=[1]-=- and Manifold [12], and its dataflow nature is also related to Broy’s timed dataflow model, although Reo is more general and more expressive that these and similar models. Much as Reo supports physica   </text>
<query_num> 304 </query_num>
<text>   es, and their consequent harmonious combination of synchrony and asynchrony are unique. For instance, these features of our model are in sharp contrast with the use of channels in the Ptolemy project =-=[15, 34, 33]-=- which ascribes a single interpretation for its connecting channels in each context. Asynchronous channels form the basis of the dataflow architecture for networks of components as proposed and formal   </text>
<query_num> 305 </query_num>
<text>   ges, modal logic, transition systems, hybrid systems, infinite data types, the control of discrete event systems, formal power series, etc. (=-=see for instance [53], [41], [42], [50], [51], [52], [22], [27]-=-). Coalgebras have also been used as models for various programming paradigms, notably for objects and classes (see, e.g., [47], [28], and [26]). One of the first applications of coalgebras to compone   </text>
<query_num> 306 </query_num>
<text>   languages, modal logic, transition systems, hybrid systems, infinite data types, the control of discrete event systems, formal power series, etc. (=-=see for instance [53], [41], [42], [50], [51], [52], [22], [27]-=-). Coalgebras have also been used as models for various programming paradigms, notably for objects and classes (see, e.g., [47], [28], and [26]). One of the first applications of coalgebras to c   </text>
<query_num> 307 </query_num>
<text>   n for its connecting channels in each context. Asynchronous channels form the basis of the dataflow architecture for networks of components as proposed and formally investigated by Broy and his group =-=[13, 25]-=-. In this architectural model, large systems can be realized, allowing programmers to easily understand the input/output behavior of a system as the composition of the behavior of its individual compo   </text>
<query_num> 308 </query_num>
<text>   reconfiguration of connectors, even while they are being used by component instances. In this respect, Reo resembles dynamically reconfigurable generalized Kahn networks, as in IWIM [1] and Manifold =-=[12]-=-, and its dataflow nature is also related to Broy’s timed dataflow model, although Reo is more general and more expressive that these and similar models. Much as Reo supports physical mobility through   </text>
<query_num> 309 </query_num>
<text>   y. The connector in Figure 3.f is another connector for the coordination pattern c = (ab)∗, although there is a subtle difference between the behavior of this connector and the one in Figure 3.d. See =-=[3, 4]-=- for more detail. It takes little effort to see that the connector in Figure 3.g corresponds to the meta-regular expression c = (aab)∗. Figures 3.f and g show how easily we can construct connectors th   </text>
</top>
<top>
<paper_num> 4 </paper_num>
<paper_title>   Impact analysis of database schema changes.  </paper_title>
<paper_abstract>   We propose static program analysis techniques for identifying the impact of relational database schema changes upon object-oriented applications. We use dataflow analysis to extract all possible database interactions that an application may make. We then use this information to predict the effects of schema change. We evaluate our approach with a case-study of a commercially available content management system, where we investigated 62 versions of between 70k-127k LoC and a schema size of up to 101 tables and 568 stored procedures. We demonstrate that the program analysis must be more precise, in terms of context-sensitivity than related work. However, increasing the precision of this analysis increases the computational cost. We use program slicing to reduce the size of the program that needs to be analyzed. Using this approach, we are able to analyse the case study in under 2 minutes on a standard desktop machine, with no false negatives and a low level of false positives.  </paper_abstract>
<query_num> 401 </query_num>
<text>   . However, as discussed slicing is not prescriptive, and we plan to investigate other forms of dependency analysis in future work. Some related work has been made in analysing transparent persistence =-=[29]-=- using program analysis, although the focus here is on providing optimisation of queries. This work bears similarity to work on extracting queries from legacy applications [9], however these technique   </text>
<query_num> 402 </query_num>
<text>   dataflow analyses =-=[25]-=-. k-CFA analyses are where all or some of the propagated data in the dataflow analysis include, in their definition, a call string that represents the last k calling call-sites =-=[14]-=-. The approach of Christensen et al. is context-insensitive which, whilst being computationally less expensive than the analysis we propose, causes a loss of precision which makes impact analysis diff  is that as k increases, k-CFA analysis becomes computationally expensive, especially for large programs. In fact, k-CFA analyses where k &amp;gt; 0 have exponential complexity with respects to program size =-=[14]-=-. This presents us with the problem of how to increase the precision of the analysis sufficiently, whilst keeping the computational cost feasible. The way we address this problem, is to reduce the cos e directly used to execute or represent the results of a query. 3.5.1 Increasing Context-Sensitivity In order to increase Choi et al’s algorithm from 1-CFA to k-CFA we use the techniques described in =-=[14]-=-. This involves modifying the property space of the dataflow analysis, so that abstract variables 2 A point of interest is referred to as a slicing criterion in the program slicing literature. and abs  </text>
<query_num> 403 </query_num>
<text>   iable results for large programs. There may be substantial opportunities to improve the speed of the slicing algorithms we are using, as shown by the timing statistics reported by Binkley and Harmann =-=[5]-=-. However, we predict that such improvements may not be possible for the dataflow analysis. This is because the dataflow analysis we describe is not representable as an efficient bit-vector problem [1 les, as well as some object oriented features such as inheritance. We are not aware of any work we could have used to corroborate our results, other than the work on slice sizes in C and C++ programs =-=[5]-=-. We present this work despite these drawbacks, as manual inspection and testing indicate that our results are valid, as problems caused by slicing inaccuracies occur rarely in our current case study.   </text>
<query_num> 404 </query_num>
<text>   istence technologies could be incorporated into our approach in the future, and insights into formalisation of such techniques. There has also been some related work produced by the testing community =-=[15]-=-. However, the program analysis used in this research, suffers from the same precision problems as the other program analysis techniques mentioned. We also note that database oriented testing in gener   </text>
<query_num> 405 </query_num>
<text>   orithms in detail in this paper. We refer interested readers Frank Tip’s survey of slicing techniques [28]. We base our prototype implementation on the slicing algorithm proposed by Liang and Harrold =-=[19]-=-. The slicing output will be a subset of the original program. This gives us enough information to simply assemble a subset of the source program, on which we can perform the dataflow analysis. 3.5 Da   </text>
<query_num> 406 </query_num>
<text>   particular form of program analysis where the possible runtime values of string variables are predicted for selected locations in the program. An example of this is the approach taken by Gould et al =-=[12]-=-, in which they use string analysis to predict the values of strings passed to the Java JDBC library methods, in order to check that the queries are type safe with respect to the database schema. The  he trade-offs involved in the selection of program analyses. Increased precision comes at a cost, therefore we are not claiming any improvement over the techniques of previous work on string analysis =-=[8, 12]-=-, we are simply identifying that cost-benefit ratios in the case of impact analysis are different, and therefore call for different techniques. 3.3 Approach Overview We have motivated the need for a p applications. Amongst these were two important papers which initially inspired our work. Firstly the string analysis of Christensen et al. and secondly, the dynamic query type checker of Gould et al. =-=[12]-=-, both of which we discussed in Section 3. We also discussed the string analysis of Choi et al. [7], the basis of our own query analy-sis. Although we have only discussed k-CFA as a way of implementi   </text>
<query_num> 407 </query_num>
<text>   program slicing and dataflow analysis is built using Microsoft’s Phoenix framework [21]. We then use the results of the dataflow analysis to reason about dependency relationships, and we use CrocoPat =-=[4]-=- for the final impact calculation. We evaluate our approach using a commercial content management system as a case study. We have considered a version history of two years, which had 62 different vers ffectedExecLines(x) :=EX(y, AffectedExecutions(y) &amp; ExecutedAtLine(y, x)); PRINT [&amp;quot; &amp;quot;] AffectedExecLines(x); } Figure 5: Example RML program To execute our impact calculation programs we use CrocoPat =-=[4]-=-. This tool allows efficient execution of relational programs against arbitrary relational data. The tool uses the RSF, and RML file formats for specifying input data and relational programs respectiv ible interactions between them. As described above, we use the RSF file format [30] to store the results of the dataflow analysis, we write impact calculation programs in RML, and we use the CrocoPat =-=[4]-=- tool to execute these programs. The eventual output of this process is a text-based impact report. We are currently targetting only C# applications that use SQL Server databases. We have implemented   </text>
<query_num> 408 </query_num>
<text>   s asdefined in Section 2. We describe these stages next. 3.4 Program Slicing A program slice contains ‘the parts of a program that (potentially) affect the values computed at some point of interest’ =-=[28]-=-. By taking a series of program slices where the point of interest 2 is a database call, we can extract a subset of the source application that can affect, or be affected by, these database calls. Thi ram, and we identify this as an important area for future research. We do not describe slicing algorithms in detail in this paper. We refer interested readers Frank Tip’s survey of slicing techniques =-=[28]-=-. We base our prototype implementation on the slicing algorithm proposed by Liang and Harrold [19]. The slicing output will be a subset of the original program. This gives us enough information to sim   </text>
<query_num> 409 </query_num>
<text>   schema change upon applications is currently often estimated manually by application experts [2]. Assessing these effects manually is both fragile and difficult [17], and can be frequently incorrect =-=[20]-=-. Moreover impact analysis from code inspections can be prohibitively expensive [22]. Therefore, in this paper we address the problem of assessing the effects of database schema change in a more relia   </text>
<query_num> 410 </query_num>
<text>   shall conduct further evaluation of the usefulness, accuracy and scalability of this approach in future work. 6. RELATED WORK There is a great deal of work related to software change impact analysis =-=[3, 23, 17]-=-. However, we are only aware of one similar project that focussed on impact analysis of database schemas [16]. This earlier work focuses on object-oriented databases whereas we consider relational dat   </text>
<query_num> 411 </query_num>
<text>   string analysis of Choi et al. =-=[7]-=-, the basis of our own query analy-sis. Although we have only discussed k-CFA as a way of implementing a context-sensitive dataflow analysis, there are alternatives =-=[18]-=-. Such alternatives need to make sure that they also provide the ability to associate query definitions, executions and the use of query results across multiple method calls. We suspect that such alte   </text>
<query_num> 412 </query_num>
<text>   therefore we have used program slicing to reduce the amount of dataflow analysis. There are other options to reduce the cost of k-CFA analysis such as demand driven interprocedural dataflow analysis =-=[13]-=-. However, such techniques are comparable to slicing, producing very similar results. We chose slicing as it is a well-understood concept that can be applied to many different programming languages. H   </text>
<query_num> 413 </query_num>
<text>   to the Java JDBC library methods, in order to check that the queries are type safe with respect to the database schema. The string analysis used was the JSA application created by Christensen et al. =-=[8]-=-. Whilst this string analysis is suitable for many such applications, we have found that it was not precise enough for the purpose of schema change impact analysis. We now describe why impact analysis he trade-offs involved in the selection of program analyses. Increased precision comes at a cost, therefore we are not claiming any improvement over the techniques of previous work on string analysis =-=[8, 12]-=-, we are simply identifying that cost-benefit ratios in the case of impact analysis are different, and therefore call for different techniques. 3.3 Approach Overview We have motivated the need for a p   </text>
<query_num> 414 </query_num>
<text>   uires changes. The effects of database schema change upon applications is currently often estimated manually by application experts [2]. Assessing these effects manually is both fragile and difficult =-=[17]-=-, and can be frequently incorrect [20]. Moreover impact analysis from code inspections can be prohibitively expensive [22]. Therefore, in this paper we address the problem of assessing the effects of   shall conduct further evaluation of the usefulness, accuracy and scalability of this approach in future work. 6. RELATED WORK There is a great deal of work related to software change impact analysis =-=[3, 23, 17]-=-. However, we are only aware of one similar project that focussed on impact analysis of database schemas [16]. This earlier work focuses on object-oriented databases whereas we consider relational dat   </text>
</top>
<top>
<paper_num> 5 </paper_num>
<paper_title>   Discovering Novel Attack Strategies from INFOSEC Alerts.  </paper_title>
<paper_abstract>   Abstract. Correlating security alerts and discovering attack strategies are important and challenging tasks for security analysts. Recently, there have been several proposed techniques to analyze attack scenarios from security alerts. However, most of these approaches depend on a priori and hard-coded domain knowledge that lead to their limited capabilities of detecting new attack strategies. In this paper, we propose an approach to discover novel attack strategies. Our approach includes two complementary correlation mechanisms based on two hypotheses of attack step relationship. The first hypothesis is that attack steps are directly related because an earlier attack enables or positively affects the later one. For this type of attack relationship, we develop a Bayesian-based correlation engine to correlate attack steps based on security states of systems and networks. The second hypothesis is that for some related attack steps, even though they do not have obvious and direct relationship in terms of security and performance measures, they still have temporal and statistical patterns. For this category of relationship, we apply time series and statistical analysis to correlate attack steps. The security analysts are presented with aggregated information on attack strategies from these two correlation engines. We evaluate our approach using DARPA’s Grand Challenge Problem (GCP) data sets. The results show that our approach can discover novel attack strategies and provide a quantitative analysis of attack scenarios. 1  </paper_abstract>
<query_num> 501 </query_num>
<text>   ased on our experience and domain knowledge. CPT values associated with each node adapt to new evidence and therefore can be updated accordingly. We apply an adaptive algorithm originally proposed by =-=[1]-=- and further developed by [6]. The motivation of using adaptive Bayesian network is that we want to fine-tune the parameters of the model and adapt the model to the evidence to fix the initial CPTs th   </text>
<query_num> 502 </query_num>
<text>   cate and consequence relationship. They use clustering algorithms to detect attack scenarios and situations. This approach pre-defines consequences of attacks in a configuration file. Morin and Debar =-=[21]-=- apply chronicle formalism to aggregate and correlate alerts. The approach performs attack scenario pattern recognition based on known malicious event sequences. Therefore, this approach is similar to   </text>
<query_num> 503 </query_num>
<text>   erefore a core component in a security management system. Recently, there have been several alert correlation proposals. With respect to correlation techniques, most of the proposed approaches (e.g., =-=[5, 9, 12, 22]-=-) rely on various forms of prior knowledge of individual attacks such as attack pre-conditions and consequences. It is difficult for these approaches to recognize new attack strategies where the attac ion based on known malicious event sequences. Therefore, this approach is similar to misuse detection and cannot detect new attack sequences. Ning et al. =-=[22]-=-, Cuppens and Miège =-=[7]-=- and Cheung et al. =-=[5]-=- build alert correlation systems based on matching the pre-/post-conditions of individual alerts. The idea of this approach is that prior attack steps prepare for later ones. Therefore, the consequenc   </text>
<query_num> 504 </query_num>
<text>   erefore a core component in a security management system. Recently, there have been several alert correlation proposals. With respect to correlation techniques, most of the proposed approaches (e.g., =-=[5, 9, 12, 22]-=-) rely on various forms of prior knowledge of individual attacks such as attack pre-conditions and consequences. It is difficult for these approaches to recognize new attack strategies where the attac oach performs attack scenario pattern recognition based on known malicious event sequences. Therefore, this approach is similar to misuse detection and cannot detect new attack sequences. Ning et al. =-=[22]-=-, Cuppens and Miège [7] and Cheung et al. [5] build alert correlation systems based on matching the pre-/post-conditions of individual alerts. The idea of this approach is that prior attack steps prep ond to the prerequisites of later attacks. The correlation engine searches alert pairs that have a consequence and prerequisite matching. Further correlation graphs can be built with such alert pairs =-=[22]-=-. One challenge to this approach is that a new attack cannot be paired with any other attacks because its prerequisites and consequences are not defined. Recently, Ning et al. [24] have extended the p   </text>
<query_num> 505 </query_num>
<text>   per and point out some ongoing and future work in Section 5. 2 Related Work Recently, there have been several proposed techniques of alert correlation and attack scenario analysis. Valdes and Skinner =-=[30]-=- use probabilistic-based reasoning to correlate alerts by measuring and evaluating the similarities of alert attributes. Alert aggregation and scenario construction are conducted by enhancing or relax  an exploit attack follows a probe than the other way around. We use domain-specific knowledge based on prior experience and empirical studies to estimate appropriate probability values. Related work =-=[30]-=- also helps us on the probability estimation. In alert correlation, the pair of alerts being evaluated in the correlation engine (as shown in Figure 1(b)) is only known at run-time. Therefore, we cann   </text>
<query_num> 506 </query_num>
<text>   pplied in econometrics, it has been widely applied in other areas, such as weather analysis (e.g., [18]), automatic control system (e.g., [4, 11]) and neurobiology (e.g., [17, 16]). In our prior work =-=[3, 2]-=-, we have applied GCT-based analysis for pro-active detection of Distributed-Denial-of-Service (DDoS) attacks using MIB II [29] variables. The results have demonstrated the correlation strength of GCT e believe that there are a large number of attacks, e.g., worms, with such attack steps. Thus, we believe that causality analysis is as10 X. Qin and W. Lee very useful technique. As also discussed in =-=[3, 2]-=-, when there is sufficient training data available, we can use GCT off-line to compute and validate very accurate “causal” relationships from alert data. We can then update the knowledge base with the   </text>
<query_num> 507 </query_num>
<text>   s and corresponding security events for further inference and correlation. Porras et al. design a “mission-impact-based” correlation system with a focus on the attack impacts on the protected domains =-=[26]-=-. The system uses clustering algorithms to aggregate and correlate alerts. Security incidents are ranked based on the security interests and the relevance of attack to the protected networks and syste   </text>
<query_num> 508 </query_num>
<text>   ter raw alerts, then prioritize the aggregated alerts before conducting further alert correlation. The corresponding algorithms for alert aggregation and prioritization can be found in our prior work =-=[27]-=-. Briefly, alert aggregation and clustering reduces the redundancy of raw alerts while retaining important alert attributes, such as time stamp, source IP, destination IP, port(s), attack class. In th relation engine based on statistical analysis, in particular, the Granger Causality Test (GCT) [13]. In this section, we briefly introduce our GCT-based correlation mechanism. Details can be found in =-=[27]-=-. Granger Causality Test (GCT) is a time series-based statistical analysis method that aims to test if a time series variable X correlates with another time series variable Y by performing a statistic he following steps: (1) First, we apply Bayesian-based correlation engine on target hyper alerts. Target alerts are hyper alerts with high priorities computed by the alert priority computation module =-=[27]-=-. Thus, they should be the main interests in the correlation analysis to correlate with all the other hyper alerts. The result of this step can be a set of isolated correlation graphs. (2) Second, for by IETF [14]. ��£ ¢sDiscovering Novel Attack Strategies from INFOSEC Alerts 13 In order to compare the performance between our current integrated correlation system and the GCT-alone approach used in =-=[27]-=-, we used the same data sets and preprocessed the raw alerts the same way as in [27]. According to the GCP documents that include detailed configurations of protected networks and systems, we establis goals enables us to identify the servers of interest and assign interest score to corresponding alerts targeting at the important hosts. The alert priority is computed based on our model described in =-=[27]-=-. For performance evaluation, we define two measures: true positive correlation rate, (i.e., ¡¡s� ¤ ����§�§�©�� ������§�§�©���¥���©���¥���©�§�� ���£¢ ¡�������¥¨�¤s� ¤ ����§�§�©���¥���©���§�©���¥¨�����   </text>
<query_num> 509 </query_num>
<text>   th such alert pairs [22]. One challenge to this approach is that a new attack cannot be paired with any other attacks because its prerequisites and consequences are not defined. Recently, Ning et al. =-=[24]-=- have extended the pre/post-condition-based correlation technique to correlate some isolated attack scenarios by hypothesizing missed attack steps. Our approach aims to address the challenge of how to onship. We also note that two different approaches have been proposed to integrate isolated correlation graphs. Ning [23] et al. apply graph theory to measure and merge similar correlation graphs. In =-=[24]-=-, Ning et al. link isolated correlation graphs based on attack pre-/post-conditions. Our approach is different from their work in that our integration method is based on the correlation probability ev   </text>
<query_num> 510 </query_num>
<text>   � £¦¥ � identified � © by Bayesian engine or equals when GCT discovers its rela� tionship. We also note that two different approaches have been proposed to integrate isolated correlation graphs. Ning =-=[23]-=- et al. apply graph theory to measure and merge similar correlation graphs. In =-=[24]-=-, Ning et al. link isolated correlation graphs based on attack pre-/post-conditions. Our approach is different from t   </text>
</top>
<top>
<paper_num> 6 </paper_num>
<paper_title>   The Recurrent Temporal Restricted Boltzmann Machine.  </paper_title>
<paper_abstract>   The Temporal Restricted Boltzmann Machine (TRBM) is a probabilistic model for sequences that is able to successfully model (i.e., generate nice-looking samples of) several very high dimensional sequences, such as motion capture data and the pixels of low resolution videos of balls bouncing in a box. The major disadvantage of the TRBM is that exact inference is extremely hard, since even computing a Gibbs update for a single variable of the posterior is exponentially expensive. This difficulty has necessitated the use of a heuristic inference procedure, that nonetheless was accurate enough for successful learning. In this paper we introduce the Recurrent TRBM, which is a very slight modification of the TRBM for which exact inference is very easy and exact gradient learning is almost tractable. We demonstrate that the RTRBM is better than an analogous TRBM at generating motion capture and videos of bouncing balls. 1  </paper_abstract>
<query_num> 601 </query_num>
<text>   Inference in RTRBMs given v T 1 is very easy, which might be surprising in light of its similarity to the TRBM. The reason inference is easy is similar to the reason inference in square ICAs is easy =-=[1]-=-: There is a unique and an easily computable value of the hidden variables that has a nonzero posterior probability. Suppose, for example, that the value of V1 is v1, which means that v1 was produced   </text>
<query_num> 602 </query_num>
<text>   P(H|v) 4. Decrease ∆W by v · h ⊤ Models learned by CD1 are often reasonable generative models of the data [3], but if learning is continued with CD25, the resulting generative models are much better =-=[11]-=-. The RBM also plays a critical role in deep belief networks [4], [5], but we do not use this connection in this paper. 3 The TRBM It is easy to construct the TRBM with RBMs. The TRBM, as described in   </text>
<query_num> 603 </query_num>
<text>   V ) is computationally intractable, and much work has been done on methods for computing approximate values for the expectations that are good enough for practical learning and inference tasks (e.g., =-=[16, 12, 19]-=-, including [15], which works well for the RBM). We will approximate the gradients with respect to the RBM’s parameters using the Contrastive Divergence [3] learning procedure, CDn, whose updates are   </text>
<query_num> 604 </query_num>
<text>   cannot affect H1. probability P ′ (H (j) = 1), and 0 otherwise. In contrast, the statement h ← P ′ (H) means that each h (j) is set to the real value P ′ (H (j) = 1), so this is a “mean-field” update =-=[8, 17]-=-. The symbol P stands for the distribution of some TRBM, while the symbol Q stands for the distribution defined by an RTRBM. Note that the outcome of the operation · ← P(Ht|vt,ht−1) is s(Wvt +W ′ ht−1   </text>
<query_num> 605 </query_num>
<text>   dicting the future of a sequence from its past, be used as a prior for denoising tasks, and be used for other applications such as tracking objects in video. The Temporal Restricted Boltzmann Machine =-=[14, 13]-=- is a recently introduced probabilistic model that has the ability to accurately model complex probability distributions over high-dimensional sequences. It was shown to be able to generate realistic  in TRBMs, on the other hand, is highly non-trivial, since computing even a single Gibbs update requires computing the ratio of two RBM partition functions. The approximate inference procedure used in =-=[13]-=- was heuristic and was not even derived from a variational principle. In this paper we introduce the Recurrent TRBM (RTRBM), which is a model that is very similar to the TRBM, and just as expressive.  H (j) t = 1| everything else) involves evaluating the exact ratio of two RBM partition functions, which can be seen from Eq. 5. This difficulty necessitated the use of a heuristic inference procedure =-=[13]-=-, which is based on the observation that the distribution P(Ht|h t−1 1 ,v t 1) = P(Ht|ht−1,vt) is factorial by definition. This inference procedure does not do any kind of smoothing from the future an H). The statement h ∼ P ′ (H) means that h is sampled from the factorial distribution P ′ (H), so each h (j) is set to 1 with 2 This is a slightly simplified description of the inference procedure in =-=[13]-=-.Figure 2: The graphical structure of the RTRBM, Q. The variables Ht are real valued while the variables H ′ t are binary. The conditional distribution Q(Vt,H ′ t|ht−1) is given by the equation Q(vt, radient would be computed exactly if CD were to return the exact gradient of the RBM’s log probability. 5 Experiments We report the results of experiments comparing an RTRBM to a TRBM. The results in =-=[14, 13]-=- were obtained using TRBMs that had several delay-taps, which means that each hidden unit could directly observe several previous timesteps. To demonstrate that the RTRBM learns to use the hidden unit   </text>
<query_num> 606 </query_num>
<text>   ge visible to hidden connections. 5.2 Motion capture data We used a dataset that represents human motion capture data by sequences of joint angle, translations, and rotations of the base of the spine =-=[14]-=-. The total number of frames in the dataset set was 3000, from which the model learned on subsequences of length 50. Each frame has 49 dimensions, and both models have 200 hidden units. The data is re y the TRBM; samples from these models are provided as videos 6,7 (RTRBM) and videos 8,9 (TRBM); video 10 is a sample training sequence. Part of the Gaussian noise was removed in a manner described in =-=[14]-=- in both models. 5.3 Details of the learning procedures Each problem was trained for 100,000 weight updates, with a momentum of 0.9, where the gradient was normalized by the length of the sequence for   </text>
<query_num> 607 </query_num>
<text>   h and videos, that is inherently sequential. A good model for these data sources could be useful for finding an abstract representation that is helpful for solving “natural” discrimination tasks (see =-=[4]-=- for an example of this approach for the non-sequential case). In addition, it could be also used for predicting the future of a sequence from its past, be used as a prior for denoising tasks, and be   reasonable generative models of the data [3], but if learning is continued with CD25, the resulting generative models are much better [11]. The RBM also plays a critical role in deep belief networks =-=[4]-=-, [5], but we do not use this connection in this paper. 3 The TRBM It is easy to construct the TRBM with RBMs. The TRBM, as described in the introduction, is a sequence of RBMs arranged in such a way   </text>
<query_num> 608 </query_num>
<text>   it. It follows from these equations that the TRBM is a directed graphical model that has an (undirected) RBM at each timestep (a related directed sequence of Boltzmann Machines has been considered in =-=[7]-=-). As in most probabilistic models, the weight update is computed by solving the inference problem and computing the weight update as if the inferred variables were observed. fully-visible case. If th   </text>
<query_num> 609 </query_num>
<text>   ntractable, and much work has been done on methods for computing approximate values for the expectations that are good enough for practical learning and inference tasks (e.g., [16, 12, 19], including =-=[15]-=-, which works well for the RBM). We will approximate the gradients with respect to the RBM’s parameters using the Contrastive Divergence [3] learning procedure, CDn, whose updates are computed by the   </text>
<query_num> 610 </query_num>
<text>   uncing in a box [13], as well as complete and denoise such sequences. As a probabilistic model, the TRBM is a directed graphical model consisting of a sequence of Restricted Boltzmann Machines (RBMs) =-=[3]-=-, where the state of one or more previous RBMs determines the biases of the RBM in next timestep. This probabilistic formulation straightforwardly implies a learning procedure where approximate infere because it learns to convey more information through its hidden-to-hidden connections. 2 Restricted Boltzmann Machines The building block of the TRBM and the RTRBM is the Restricted Boltzmann Machine =-=[3]-=-. An RBM defines a probability distribution over pairs of vectors, V ∈ {0,1} NV and H ∈ {0,1} NH (a shorthand for visible and hidden) by the equation P(v,h) = P(V = v,H = h) = exp(v ⊤ bV + h ⊤ bH + v  earning and inference tasks (e.g., [16, 12, 19], including [15], which works well for the RBM). We will approximate the gradients with respect to the RBM’s parameters using the Contrastive Divergence =-=[3]-=- learning procedure, CDn, whose updates are computed by the following algorithm. Algorithm 1 (CDn) 1. Sample (v,h) ∼ P(H|V ) ˜ P(V ) 2. Set ∆W to v · h ⊤ 3. repeat n times: sample v ∼ P(V |h), then sa   </text>
<query_num> 611 </query_num>
<text>   units to their full potential [2]. However, this disadvantage is common to many other probabilistic models, and it can be partially alleviated using techniques such as the long short term memory RNN =-=[6]-=-. Acknowledgments This research was partially supported by the Ontario Graduate Scholarship and by the Natural Council of Research and Engineering of Canada. The mocap data used in this project was ob   </text>
</top>
<top>
<paper_num> 7 </paper_num>
<paper_title>   A probabilistic framework for relational clustering.  </paper_title>
<paper_abstract>   Relational clustering has attracted more and more attention due to its phenomenal impact in various important applications which involve multi-type interrelated data objects, such as Web mining, search marketing, bioinformatics, citation analysis, and epidemiology. In this paper, we propose a probabilistic model for relational clustering, which also provides a principal framework to unify various important clustering tasks including traditional attributes-based clustering, semi-supervised clustering, co-clustering and graph clustering. The proposed model seeks to identify cluster structures for each type of data objects and interaction patterns between different types of objects. Under this model, we propose parametric hard and soft relational clustering algorithms under a large number of exponential family distributions. The algorithms are applicable to relational data of various structures and at the same time unifies a number of stat-of-the-art clustering algorithms: co-clustering algorithms, the k-partite graph clustering, and semi-supervised clustering based on hidden Markov random fields.  </paper_abstract>
<query_num> 701 </query_num>
<text>   . Graph partitioning has been studied for decades and a number of different approaches, such as spectral approaches [7, 42, 13] and multilevel approaches [6, 20, 28], have been proposed. Some efforts =-=[17, 43, 21, 21, 1]-=- based on stochastic block modeling also focus on homogeneous relations. Compared with co-clustering and homogeneous-relationbased clustering, clustering on general relational data, which may consist   </text>
<query_num> 702 </query_num>
<text>   . Graph partitioning has been studied for decades and a number of different approaches, such as spectral approaches [7, 42, 13] and multilevel approaches [6, 20, 28], have been proposed. Some efforts =-=[17, 43, 21, 21, 1]-=- based on stochastic block modeling also focus on homogeneous relations. Compared with co-clustering and homogeneous-relationbased clustering, clustering on general relational data, which may consist  r MMRC throughout the rest of the paper). Assume that each type of objects X (j) has kj latent classes. We represent the membership vectors for all the objects in X (j) as a membership matrix Λ (j) ∈ =-=[0, 1]-=- k j ×n j such that the sum of elements of each column Λ (j) ·p is 1 and Λ (j) ·p denotes the membership vector for object x (j) p , i.e., Λ (j) gp denotes the probability that object x (j) p associat   </text>
<query_num> 703 </query_num>
<text>   In this section, we present experiments on the MMRC algorithm under normal distribution in comparison with two representative graph partitioning algorithms, the spectral graph partitioning (SGP) from =-=[36]-=- that is generalized to work with both normalized cut and ratio association, and the classic multilevel algorithm, METIS [28]. The graphs based on the text data have been widely used to test graph par   </text>
<query_num> 704 </query_num>
<text>   al relational data, which may consist of more than two types of data objects with various structures, has not been well studied in the literature. Several noticeable efforts are discussed as follows. =-=[45, 19]-=- extend the the probabilistic relational model to the clustering scenario by introducing latent variables into the model; these models focus on using attribute information for clustering. [18] formula   </text>
<query_num> 705 </query_num>
<text>   data, is called co-clustering or biclustering. Several previous efforts related to co-clustering are model based [22, 23]. Spectral graph partitioning has also been applied to bi-type relational data =-=[11, 25]-=-. These algorithms formulate the data matrix as a bipartite graph and seek to find the optimal normalized cut for the graph. Due to the nature of a bipartite graph, these algorithms have the restricti to work with both normalized cut and ratio association, and the classic multilevel algorithm, METIS [28]. The graphs based on the text data have been widely used to test graph partitioning algorithms =-=[13, 11, 25]-=-. In this study, we use various data sets from the 20-newsgroups [32], WebACE and TREC [27], which cover data sets of different sizes, different balances and different levels of difficulties. The data Poisson distribution to clustering bi-type relational data, word-document data, and tri-type relational data, worddocument-category data. Two algorithms, Bi-partite Spectral Graph partitioning (BSGP) =-=[11]-=- and Relation Summary Network under Generalized I-divergence (RSN-GI) [34], are used as comparison in bi-clustering. For tri-clustering, Consistent Bipartite Graph Co-partitioning (CBGC) [18] and RSN-   </text>
<query_num> 706 </query_num>
<text>   duling. Existing graph partitioning approaches are mainly based on edge cut objectives, such as KernighanLin objective [30], normalized cut [42], ratio cut [7], ratio association[42], and min-max cut =-=[13]-=-. Graph clustering is equivalent to clustering on single-type relational data of one homogeneous relation matrix S. The log-likelihood function of the hard clustering under MMRC model is −Dφ(S, (C) T  to work with both normalized cut and ratio association, and the classic multilevel algorithm, METIS [28]. The graphs based on the text data have been widely used to test graph partitioning algorithms =-=[13, 11, 25]-=-. In this study, we use various data sets from the 20-newsgroups [32], WebACE and TREC [27], which cover data sets of different sizes, different balances and different levels of difficulties. The data   </text>
<query_num> 707 </query_num>
<text>   e reinforcement clustering process. There are no sound objective function and theoretical proof on the effectiveness and correctness (convergence) of the mutual reinforcement clustering. Some efforts =-=[26, 50, 49, 5]-=- in the literature focus on how to measure the similarities or choosing cross-relational attributes. To summarize, the research on relational data clustering has attracted substantial attention, espec   </text>
<query_num> 708 </query_num>
<text>   ering (partitioning) [7, 42, 13, 6, 20, 28] can be viewed as clustering on singly-type relational data consisting of only homogeneous relations (represented as a graph affinity matrix); co-clustering =-=[12, 2]-=- which arises in important applications such as document clustering and micro-array data clustering, can be formulated as clustering on bi-type relational data consisting of only heterogeneous relatio  maximize the mutual information between the clustered random variables subject to the constraints on the number of row and column clusters. A more generalized co-clustering framework is presented by =-=[2]-=- wherein any Bregman divergence can be used in the objective function. Recently, coclustering has been addressed based on matrix factorization. [35] proposes an EM-like algorithm based on multiplicati simplicity, scalability, and broad applicability, k-means algorithm has become one of the most popular clustering algorithms. Hence, it is desirable to extend k-means to relational data. Some efforts =-=[47, 2, 12, 33]-=- in the literature work in this direction. However, these approaches apply to only some special and simple cases of relational data, such as bi-type heterogeneous relational data. As traditional k-mea upervised clustering algorithms. 5.2 Co-clustering Co-clustering or bi-clustering arise in many important applications, such as document clustering, micro-array data clustering.A number of approaches =-=[12, 8, 33, 2]-=- have been proposed for co-clustering. These efforts can be generalized as solving the following matrix approximation problem [34], arg min C,Υ D(R, (C(1) ) T ΥC (2) ) (33) where R ∈ R n1×n2 is the da bership parameters, maximizing log-likelihood function of hard clustering on a heterogeneous relation matrix under the MMRC model is equivalent to the minimization in (33). The algorithms proposed in =-=[12, 8, 33, 2]-=- can be viewed as special cases of hard EF-MMRC. At the same time, soft EF-MMRC provides another family of new algorithms for co-clustering. Our previous work [34] proposes the relation summary networ   </text>
<query_num> 709 </query_num>
<text>   ering (partitioning) [7, 42, 13, 6, 20, 28] can be viewed as clustering on singly-type relational data consisting of only homogeneous relations (represented as a graph affinity matrix); co-clustering =-=[12, 2]-=- which arises in important applications such as document clustering and micro-array data clustering, can be formulated as clustering on bi-type relational data consisting of only heterogeneous relatio orithms have the restriction that the clusters from different types of objects must have one-to-one associations. Informationtheory based co-clustering has also attracted attention in the literature. =-=[12]-=- proposes a co-clustering algorithm to maximize the mutual information between the clustered random variables subject to the constraints on the number of row and column clusters. A more generalized co simplicity, scalability, and broad applicability, k-means algorithm has become one of the most popular clustering algorithms. Hence, it is desirable to extend k-means to relational data. Some efforts =-=[47, 2, 12, 33]-=- in the literature work in this direction. However, these approaches apply to only some special and simple cases of relational data, such as bi-type heterogeneous relational data. As traditional k-mea upervised clustering algorithms. 5.2 Co-clustering Co-clustering or bi-clustering arise in many important applications, such as document clustering, micro-array data clustering.A number of approaches =-=[12, 8, 33, 2]-=- have been proposed for co-clustering. These efforts can be generalized as solving the following matrix approximation problem [34], arg min C,Υ D(R, (C(1) ) T ΥC (2) ) (33) where R ∈ R n1×n2 is the da bership parameters, maximizing log-likelihood function of hard clustering on a heterogeneous relation matrix under the MMRC model is equivalent to the minimization in (33). The algorithms proposed in =-=[12, 8, 33, 2]-=- can be viewed as special cases of hard EF-MMRC. At the same time, soft EF-MMRC provides another family of new algorithms for co-clustering. Our previous work [34] proposes the relation summary networ   </text>
<query_num> 710 </query_num>
<text>   ibution assumptions for the data. How to decide the optimal distribution assumption is beyond the scope of this paper. For performance measure, we elect to use the Normalized Mutual Information (NMI) =-=[44]-=- between the resulting cluster labels and the true cluster labels, which is a standard way to measure the cluster quality. The final performance score is the average of ten runs. 6.1 Graph Clustering   </text>
<query_num> 711 </query_num>
<text>   ing. Mixed membership models, which assume that each object has mixed membership denoting its association with 4 F (5)sclasses, have been widely used in the applications involving soft classification =-=[16]-=-, such as matching words and pictures [39], race genetic structures [39, 48], and classifying scientific publications [15]. In this paper, we propose a relational mixed membership model to cluster rel   </text>
<query_num> 712 </query_num>
<text>   istribution functions of exponential families can be formulated as a general form. This nice property facilitates us to derive a general EM algorithm for the MMRC model. It is shown in the literature =-=[3, 9]-=- that there exists bijection between exponential families and Bregman divergences [40]. For example, the normal distribution, Bernoulli distribution, multinomial distribution and exponential distribut   </text>
<query_num> 713 </query_num>
<text>   mbership denoting its association with 4 F (5)sclasses, have been widely used in the applications involving soft classification [16], such as matching words and pictures [39], race genetic structures =-=[39, 48]-=-, and classifying scientific publications [15]. In this paper, we propose a relational mixed membership model to cluster relational data (we refer to the model as mixed membership relational clusterin   </text>
<query_num> 714 </query_num>
<text>   number of important clustering problems, which have been of intensive interest in the literature, can be viewed as special cases of relational clustering. For example, graph clustering (partitioning) =-=[7, 42, 13, 6, 20, 28]-=- can be viewed as clustering on singly-type relational data consisting of only homogeneous relations (represented as a graph affinity matrix); co-clustering [12, 2] which arises in important applicati objects based on pairwise similarities, which can be viewed as homogeneous relations. Graph partitioning has been studied for decades and a number of different approaches, such as spectral approaches =-=[7, 42, 13]-=- and multilevel approaches [6, 20, 28], have been proposed. Some efforts [17, 43, 21, 21, 1] based on stochastic block modeling also focus on homogeneous relations. Compared with co-clustering and hom omains, such as circuit partitioning, VLSI design, task scheduling. Existing graph partitioning approaches are mainly based on edge cut objectives, such as KernighanLin objective [30], normalized cut =-=[42]-=-, ratio cut [7], ratio association[42], and min-max cut [13]. Graph clustering is equivalent to clustering on single-type relational data of one homogeneous relation matrix S. The log-likelihood funct   </text>
<query_num> 715 </query_num>
<text>   o various clustering tasks. 6. EXPERIMENTS This section provides empirical evidence to show the effectiveness of the MMRC model and algorithms. Since a number of stat-of-the-art clustering algorithms =-=[12, 8, 33, 2, 3, 4]-=- can be viewed as special cases of EF-MMRC model and algorithms, the experimental results in these efforts also illustrate the effectiveness of the MMRC model and algorithms. In this paper, we apply M   </text>
<query_num> 716 </query_num>
<text>   only some special and simple cases of relational data, such as bi-type heterogeneous relational data. As traditional k-means can be formulated as a hard version of Gaussian mixture model EM algorithm =-=[29]-=-, we propose the hard version of MMRC algorithm as a general relational k-means algorithm (from now on, we call Algorithm 1 as soft EF-MMRC), which applies to various relational data. To derive the ha   </text>
<query_num> 717 </query_num>
<text>   raph under a broad range of distortion measures. The above graph-based algorithms do not consider attribute information. Some efforts on relational clustering are based on inductive logic programming =-=[37, 24, 31]-=-. Base on the idea of mutual reinforcement clustering, [51] proposes a framework 1 2 F (1) 1 2 2 3 1 5 (a) (b) (c) Figure 1: Examples of the structures of relational data. for clustering heterogeneous   </text>
<query_num> 718 </query_num>
<text>   relational data consisting of only heterogeneous relations, such as the word-document data, is called co-clustering or biclustering. Several previous efforts related to co-clustering are model based =-=[22, 23]-=-. Spectral graph partitioning has also been applied to bi-type relational data [11, 25]. These algorithms formulate the data matrix as a bipartite graph and seek to find the optimal normalized cut for   </text>
<query_num> 719 </query_num>
<text>   sed algorithms do not consider attribute information. Some efforts on relational clustering are based on inductive logic programming [37, 24, 31]. Base on the idea of mutual reinforcement clustering, =-=[51]-=- proposes a framework 1 2 F (1) 1 2 2 3 1 5 (a) (b) (c) Figure 1: Examples of the structures of relational data. for clustering heterogeneous Web objects and [47] presents an approach to improve the c   </text>
<query_num> 720 </query_num>
<text>   uch as document clustering and micro-array data clustering, can be formulated as clustering on bi-type relational data consisting of only heterogeneous relations. Recently, semi-supervised clustering =-=[46, 4]-=- has attracted significant attention, which is a special type of clustering using both labeled and unlabeled data. In section 5, we show that semi-supervised clustering can be formulated as clustering w to the existing clustering approaches from various important data mining applications. 5.1 Semi-supervised Clustering Recently, semi-supervised clustering has become a topic of significant interest =-=[4, 46]-=-, which seeks to cluster a set of data points with a set of pairwise constraints. Semi-supervised clustering can be formulated as a special case of relational clustering, clustering on the single-type  and the qth object. =-=[4]-=- provides a general model for semi-supervised clustering based on Hidden Markov Random Fields (HMRFs). We show that it can be formulated as a special case of MMRC model. As in [4], we define the homogeneous relation matrix S as follows, ���fM (xp, xq) if (xp, xq) ∈ M Spq = fC(xp, xq) if (xp, xq) ∈ C 0 otherwise where M denotes a set of must-link constraints; C denotes a set of on of must-link constraint; fC(xp, xq) is a penalty function for cannot-links. If we assume Gibbs distribution [41] for S, P r(S) = 1 exp(−�Spq). (30) z1 where z1 is the normalization constant. Since =-=[4]-=- focuses on p,q only hard clustering, we omit the soft member parameters in the MMRC model to consider hard clustering. Based on Eq.(30) and Eq.(4), the likelihood function of hard semisupervised clus formulated as L(Θ|F) = 1 z exp(−� k� � Spq) exp(− Dφ(F·p, Λ·g)) p,q g=1 p:Cgp=1 (32) The above likelihood function is equivalent to the objective function of semi-supervised clustering based on HMRFs =-=[4]-=-. Furthermore, when applied to optimizing the objective function in Eq.(32), hard MMRC provides a family of semisupervised clustering algorithms similar to HMRF-KMeans in [4]; on the other hand, soft   </text>
<query_num> 721 </query_num>
<text>   upervised clustering algorithms. 5.2 Co-clustering Co-clustering or bi-clustering arise in many important applications, such as document clustering, micro-array data clustering.A number of approaches =-=[12, 8, 33, 2]-=- have been proposed for co-clustering. These efforts can be generalized as solving the following matrix approximation problem [34], arg min C,Υ D(R, (C(1) ) T ΥC (2) ) (33) where R ∈ R n1×n2 is the da k1×n1 and C (2) ∈ {0, 1} k2×n2 are indicator matrices, Υ ∈ R k1×k2 is the relation representative matrix, and D is a distance function. For example, [12] uses KL-divergences as the distance function; =-=[8, 33]-=- use Euclidean distances. Co-clustering is equivalent to clustering on relational data of one heterogeneous relation matrix R. Based on Eq.(9), by omitting the soft membership parameters, maximizing l o various clustering tasks. 6. EXPERIMENTS This section provides empirical evidence to show the effectiveness of the MMRC model and algorithms. Since a number of stat-of-the-art clustering algorithms =-=[12, 8, 33, 2, 3, 4]-=- can be viewed as special cases of EF-MMRC model and algorithms, the experimental results in these efforts also illustrate the effectiveness of the MMRC model and algorithms. In this paper, we apply M   </text>
<query_num> 722 </query_num>
<text>   ws. [45, 19] extend the the probabilistic relational model to the clustering scenario by introducing latent variables into the model; these models focus on using attribute information for clustering. =-=[18]-=- formulates star-structured relational data as a star-structured m-partite graph and develops an algorithm based on semi-definite programming to partition the graph. [34] formulates multi-type relatio g (BSGP) [11] and Relation Summary Network under Generalized I-divergence (RSN-GI) [34], are used as comparison in bi-clustering. For tri-clustering, Consistent Bipartite Graph Co-partitioning (CBGC) =-=[18]-=- and RSN-GI are used as comparison. The bi-type relational data, word-document data, are constructed based on various subsets of the 20-Newsgroup data. We pre-process the data by selecting the top 200 The links between documents and categories are constructed such that if a document belongs to k categories, the weights of links between this document and these k category nodes are 1/k (please refer =-=[18]-=- for details). The true taxonomy structures for two data sets, TP-TM1 and TP-TM2, are documented in Table 3. Figure 3 and Figure 4 show the NMI comparison of the three algorithms on bi-type and tri-ty   </text>
</top>
<top>
<paper_num> 8 </paper_num>
<paper_title>   PROXIMUS: a framework for analyzing very high dimensional discrete-attributed datasets.  </paper_title>
<paper_abstract>   This paper presents an efficient framework for error-bounded compression of high-dimensional discrete attributed datasets. Such datasets, which frequently arise in a wide variety of applications, pose some of the most significant challenges in data analysis. Subsampling and compression are two key technologies for analyzing these datasets. PROXIMUS provides a technique for reducing large datasets into a much smaller set of representative patterns, on which traditional (expensive) analysis algorithms can be applied with minimal loss of accuracy. We show desirable properties of PROXIMUS in terms of runtime, scalability to large datasets, and performance in terms of capability to represent data in a compact form. We also demonstrate applications of PROXIMUS in association rule mining. In doing so, we establish PROXIMUS as a tool for preprocessing data before applying computationally expensive algorithms or as a tool for directly extracting correlated patterns. Our experimental results show that use of the compressed data for association rule mining provides excellent precision and recall values (near 100%) across a range of support thresholds while reducing the time required for association rule mining drastically.  </paper_abstract>
<query_num> 801 </query_num>
<text>   . SDD alsosnds application in image compression and pattern matching and has been shown to provide fast and accurate pattern matching though performing slightly worse than DCT-based image compression =-=[2,6]-=-. McConnell and Skillicorn show that SDD diers from SVD in that it is extremely eective insnding outlier clusters in datasets and works well in information retrieval for datasets containing a large   </text>
<query_num> 802 </query_num>
<text>   large scale data analysis. Variants of orthogonal and non-orthogonal matrix transformations such as truncated SVD, SDD, Centroid Decomposition and PDDP have been widely used in information retrieval =-=[3, 4, 6, 15, 16]-=-. In the rest of this section, we summarize commonly used orthogonal and non-orthogonal matrix transformations and their applications in data analysis and explore alternate approaches for binary datas equires 1.5 bits, thus enabling a higher rank representation for a given amount of memory. SDD applied to LSI has been shown to do as well as truncated SVD, while using less than onetenth the storage =-=[15]-=-. SDD alsosnds application in image compression and pattern matching and has been shown to provide fast and accurate pattern matching though performing slightly worse than DCT-based image compression   </text>
<query_num> 803 </query_num>
<text>   large scale data analysis. Variants of orthogonal and non-orthogonal matrix transformations such as truncated SVD, SDD, Centroid Decomposition and PDDP have been widely used in information retrieval =-=[3, 4, 6, 15, 16]-=-. In the rest of this section, we summarize commonly used orthogonal and non-orthogonal matrix transformations and their applications in data analysis and explore alternate approaches for binary datas ing to small singular values aresltered. 2.2 Semi-Discrete Decomposition(SDD) SDD is a variant of SVD in which the values of the entries in matrices U and V are constrained to be in the set f1; 0; 1g =-=[16]-=-. The main advantage of SDD is its lower storage requirement, since each element only requires 1.5 bits, thus enabling a higher rank representation for a given amount of memory. SDD applied to LSI has mall clusters [20]. Since the entries of the singular vectors are constrained to be in the set f-1,0,1g, computation of SDD becomes an integer programming problem, which is NP-hard. Kolda and O&amp;apos;Leary =-=[16]-=- propose an iterative alternating heuristic to solve the problem ofsnding rank-one approximations to a matrix in polynomial time. Each iteration of this heuristic has linear time complexity. 2.3 Centr can be evaluated in O(m) time. Similarly, we can compute vector y that maximizes Cd (x; y) for asxed x in linear time. This leads to an alternating iterative algorithm based on the computation of SDD =-=[16]-=-, namely initialize y , then solve for x . Now, solve for y based on updated value of x . Repeat this process until there is no improvement in the objective function. Indeed, this technique is distant   </text>
<query_num> 804 </query_num>
<text>   ply partitioning heuristics on this representation. Graph-based approaches represent similarity between pairs of data items using weights assigned to edges and cost functions on this similarity graph =-=[8, 10]-=-. Hypergraph-based approaches observe that discreteattribute datasets are naturally described by hypergraphs and directly dene cost functions on the corresponding hypergraph [11, 22]. Our approach di   </text>
<query_num> 805 </query_num>
<text>   rgely focused on clustering very large categorical datasets. A class of approaches is based on well-known techniques such as vector-quantization [9] and k-means clustering [19]. The k-modes algorithm =-=[12]-=- extends k-means to the discrete domain by dening new dissimilarity measures. Another class of algorithms is based on similarity graphs and hypergraphs. These methods represent the data as a graph or   </text>
<query_num> 806 </query_num>
<text>   roaches to analysis of large scale data focus on probabilistic subsampling and data compression. Data reduction techniques based on probabilistic subsampling have been explored by several researchers =-=[13, 23, 24, 25]-=-. Data compression techniques are generally based on the idea ofsnding compact representations for data through discovery of dominant patterns or signals. A natural way of compressing data relies on m   </text>
<query_num> 807 </query_num>
<text>   ture to improve both the quality and eciency of the analysis. Our approach is based on recursively computing discrete rank-one approximations to the matrix to extract dominant patterns hierarchically =-=[17]-=-. The problem of error-bounded approximation can also be thought of assnding dense patterns in sparse matrices. A binary rank-one approximation for a matrix is dened as an outer product of two binary   </text>
<query_num> 808 </query_num>
<text>   work on summarizing discrete-attributed datasets is largely focused on clustering very large categorical datasets. A class of approaches is based on well-known techniques such as vector-quantization =-=[9]-=- and k-means clustering [19]. The k-modes algorithm [12] extends k-means to the discrete domain by dening new dissimilarity measures. Another class of algorithms is based on similarity graphs and hyp   </text>
</top>
<top>
<paper_num> 9 </paper_num>
<paper_title>   On Distributed Object Checkpointing and Recovery.  </paper_title>
<paper_abstract>   Recovery by checkpointing on distributed shared memory systems is investigated in this paper. The notion of consistent global states on a sequentially consistent shared memory system is defined. We investigate how consistent checkpoints can be obtained in these systems. In addition, a novel lazy checkpointing approach is proposed. It allows a controlled degree of concurrency and, at the same time, limits the amount of rollback propagation during recovery. Correctness requirements for efficient checkpointing are explored first and algorithms satisfying the requirements are developed subsequently. Several interesting properties of checkpointing on distributed shared memory systems are discovered. In particular, we show that for low levels of laziness, one can achieve better concurrency with more stable storage.  1 Introduction  Among various programming paradigms, shared memory programming is relatively easier because of its good abstraction of communication and synchronization. Low cost...  </paper_abstract>
<query_num> 901 </query_num>
<text>   ading from the owner, and establishing a write-into relation. Whilst this is in general true in a majority of DSM protocols, it is not true in remote procedure call based DSM protocols, as in Split-C =-=[7]-=-. One way to incorporate these protocols into our model is to let writer processes checkpoint the variables they updated even though they are no longer the owners. Another way is to model a remote wri   </text>
<query_num> 902 </query_num>
<text>   and a proper shared memory access interface. These systems are commonly called distributed shared memory (DSM) systems or shared virtual memory (SVM) systems. A number of such systems have been built =-=[16, 21, 28]-=- since the concept was grounded by Li and Hudak [19]. We consider the important issues of maintaining a robust shared memory system in this paper. As the size of a parallel system increases, it become   </text>
<query_num> 903 </query_num>
<text>   applied on a shared memory system, a number of authors have pointed out that more efficient checkpointing is possible if the content of the messages implementing the shared memory system can be used =-=[13]-=-. However, as far as we know, models for recoverable shared memory based computation have not yet been clearly and formally defined. Without a carefully defined model, it becomes difficult to justify  ory is sequential consistent [17], Janssens and Fuchs [12] consider efficient checkpointing on DSM with relaxed consistency. All the above mentioned approaches are considered as communication-induced =-=[13]-=-. They assume shared memory systems are implemented on a message passing platform. Consequently, a consistent global state of a shared memory system can be inferred from a consistent global state of t  and write shared memory operations to reduce system state dependency.Janssens and Fuchs argue that several messages are irrelevant to the consistency of global state in certain shared memory systems =-=[13]-=-. Checkpoint recovery can be made more efficient by ignoring the effects of these messages. So far, there is no well-defined model purely based on the consistency requirements and dependency of shared   </text>
<query_num> 904 </query_num>
<text>   eedup, it is therefore necessary to design systems that can tolerate various kinds of errors and failures. No system can tolerate any arbitrary kind of failures without imposing stringent assumptions =-=[8]-=-. In practice, it is usually assumed that some stable storage exists that always survives failures. On these systems, checkpointing is perhaps the most popular technique for fault-tolerant computing.   </text>
<query_num> 905 </query_num>
<text>   ere is no well-defined model purely based on the consistency requirements and dependency of shared memory systems. 3 System Model 3.1 State Transition Systems Our distributed system model is based on =-=[4]-=-. A system consists of a set of processes (or processors), a set of events, and a set of read/write memory objects. Every object is assumed to have a serial specification [9]. We do not distinguish be ency defined by Lamport [17] as our correctness requirement, since this is the most widely accepted correctness criterion. The definition of sequential consistency formally stated by Attiya and Welch =-=[4]-=- is as follows: Definition 2 (Sequential Consistency) An executionsoe is sequentially consistent if there exists a legal sequencesof operations such that, for each process p, the restriction of events   </text>
<query_num> 906 </query_num>
<text>   lerant technique in parallel and distributed computing. In the context of message passing based computations, a wide variety of algorithms have been proposed for maintaining recoverable global states =-=[5, 6, 15, 22]-=-. Chandy and Lamport define the notion of a consistent global state, on which a distributed computation should be based [6]. They propose taking checkpoints on processes in a coordinated manner. Recov s y in program order, and 8 x 2 P1 : 9 y 2 P2 : x = ysx precedes y in program order. The notion of consistent global states for message passing systems has been formally defined by Chandy and Lamport =-=[6]-=-. Surprisingly, the notion of consistent global states for shared memory systems has never been formally defined in the literature. In general, consistent global states for shared memory systems can o cut is consistent if every read preceding the cut is written into by a write preceding the cut. 2 Our definition of cuts should not be confused with the definition of cuts of distributed computations =-=[6]-=-. Proof: Let oe be an execution and oe 0 ae oe be an execution that leads the system to the global state corresponding to a cut. Suppose every read preceding the cut is written into by a write precedi ying the sequences0 on the initial state is the same as that of the cut. The proof follows. 2 The correctness criterion derived from Lemma 1 is weaker than traditional models based on message passing =-=[6]-=-. As an example, the checkpoints in Figure 1 form a consistent global state when the semantics of read/write operations is considered. The reason is that the read request is sent at a and the response heckpoint c 1 . Furthermore, the write operation occurs before checkpoint c 2 . By virtue of Lemma 1, checkpoints c 1 and c 2 form a consistent global state. However, under Chandy and Lamport&amp;apos;s model =-=[6]-=-, the response message is an orphan message that has been received but not yet been sent in the cut formed by the checkpoints. Hence c 1 and c 2 do not form a consistent global state under their model   </text>
<query_num> 907 </query_num>
<text>   ns of a process i are totally ordered by the individual program order of the process, denoted by ) i . The union of these individual program orders is denoted by ). The notion of writeintosdefined in =-=[3]-=- is adopted in this paper. The writeintosrelation relates operations according to the semantics of read and write operations. If a write operation o 1 = w(x; v)() writes into a read operation o 2 = r( te when it is clear from context to which state we are referring. A cut is a set 1 For simplicity, we have adopted the assumption that writes to the same variable are associated with different values =-=[3]-=-. An alternative definition of write-into without making such an assumption is also possible. of local checkpoints, which contains exactly one checkpoint from each process 2 . A minimal cut of a set o   </text>
<query_num> 908 </query_num>
<text>   re commonly called distributed shared memory (DSM) systems or shared virtual memory (SVM) systems. A number of such systems have been built [16, 21, 28] since the concept was grounded by Li and Hudak =-=[19]-=-. We consider the important issues of maintaining a robust shared memory system in this paper. As the size of a parallel system increases, it becomes more important for the system to be reliable. Part set of correctness conditions that a family of lazy checkpointing algorithms should observe are defined and proved. From the conditions, a lazy checkpointing algorithm based on the ownership protocol =-=[19]-=- is derived. The algorithm is a generalization of several existing algorithms. For example, when the writepenetration bound G is set to 0, the algorithm behaves as if it were a synchronous checkpointi   </text>
<query_num> 909 </query_num>
<text>   system model is based on [4]. A system consists of a set of processes (or processors), a set of events, and a set of read/write memory objects. Every object is assumed to have a serial specification =-=[9]-=-. We do not distinguish between shared memory and local memory. Instead, memory appears to be local if one and only one process accesses it in any execution. A process is an automaton with states and   sequential consistency. It is an open question whether Theorem 1 will hold or not for Gs3 if sequential consistency is strengthened, e.g. to realizable sequential consistency [10] or linearizability =-=[9]-=-. 5 Realizing a Lazy Checkpointing Algorithm In this section, we present a lazy checkpointing algorithm based on Theorem 1. To ensure correctness, our algorithm needs to maintain the safety requiremen   </text>
<query_num> 910 </query_num>
<text>   tificially delaying the delivery of a message to control the amount of rollback propagation [26]. A number of authors discuss about providing faulttolerance in distributed shared memory (DSM) systems =-=[11, 12, 23, 27]-=-. Wu and Fuchs propose the twin page approach in which a page is used to store the valid data and its twin acts as a checkpoint or represents an obsolete version of the data [27]. Stumm and Zhou [23]   </text>
<query_num> 911 </query_num>
<text>   tificially delaying the delivery of a message to control the amount of rollback propagation [26]. A number of authors discuss about providing faulttolerance in distributed shared memory (DSM) systems =-=[11, 12, 23, 27]-=-. Wu and Fuchs propose the twin page approach in which a page is used to store the valid data and its twin acts as a checkpoint or represents an obsolete version of the data [27]. Stumm and Zhou [23]  mory. Pending shared memory operations are allowed to complete and a consistent checkpoint is taken. While most work on recovery on DSM assume memory is sequential consistent [17], Janssens and Fuchs =-=[12]-=- consider efficient checkpointing on DSM with relaxed consistency. All the above mentioned approaches are considered as communication-induced [13]. They assume shared memory systems are implemented on   </text>
</top>
<top>
<paper_num> 10 </paper_num>
<paper_title>   Efficient and secure threshold-based event validation for VANETs.  </paper_title>
<paper_abstract>   Determining whether the number of vehicles reporting an event is above a threshold is an important mechanism for VANETs, because many applications rely on a threshold number of notifications to reach agreement among vehicles, to determine the validity of an event, or to prevent the abuse of emergency alarms. We present the first efficient and secure threshold-based event validation protocol for VANETs. Quite counter-intuitively, we found that the z-smallest approach offers the best tradeoff between security and efficiency since other approaches perform better for probabilistic counting. Analysis and simulation shows that our protocol provides&amp;gt; 99 % accuracy despite the presence of attackers, collection and distribution of alerts in less than 1  </paper_abstract>
<query_num> 1001 </query_num>
<text>   a specific probabilistic counting scheme, and a discussion of probabilistic counting schemes’ trade-offs and limitations. Probabilistic counting selects several representative elements, or a synopsis =-=[22]-=-, as an estimator for the total number of distinct elements =-=[1, 3, 9]-=-. The synopsis summarizes the entire element set and thus permits estimation of the total size. Probabilistic counting provides a t mallest observed element, and e0 and e1 the minimal and maximal value, respectively. Generally a probabilistic counting scheme provides three functions on synopses: Generation, Fusion, and Evaluation =-=[22]-=-. A Generation function selects the representative items from the input set I to use as a synopsis S. In this paper, we consider a class of probabilistic counting schemes whose Fusion function prevent   </text>
<query_num> 1002 </query_num>
<text>   algorithm outputs “threshold detected” while in fact n&amp;lt;τ (a large FP rate of P [D =1|n &amp;lt;τ] &amp;gt;δ1). We assume jamming and denial-of-service attacks can be mitigated by techniques such as spread spectrum =-=[6]-=-, channel switching [28] or adaptive authentication [30]; providing reliable wireless communication is outside the scope of this paper. We do not consider deflation attacks, where an attacker covers u   </text>
<query_num> 1003 </query_num>
<text>   an be provided by the Global Positioning Sys1 VANETs can leverage multiple keys per vehicle to provide privacy [28, 33]. However, only one key pair is valid at any given time to prevent Sybil attacks =-=[8]-=- where one vehicle poses as many vehicles.tem (GPS), which is available in many vehicles nowadays and necessary for VANET safety applications. We do not require secure positioning, and thus we tolera   </text>
<query_num> 1004 </query_num>
<text>   articipate in alert collection and distribution are honest. A honest participant complies with all VANET protocols and reports correct information. A temporary, localized dishonest majority may exist =-=[21]-=- (e.g., 7 out of 10 vehicles in one block are dishonest). However, such a small-scale dishonest majority has a limited impact on MH-relevant applications because the number of malicious entities is to   </text>
<query_num> 1005 </query_num>
<text>   ces. 36] to defend against attacks, but their assumption of known network topology conflicts with vehicle mobility. Probabilistic counting has been proposed for efficient data dissemination in VANETs =-=[20]-=-. However, to secure probabilistic counting, most schemes need to store hundreds of signatures =-=[12, 19]-=-; such overhead is impractical for VANET. Aggregate signatures. Aggregate signatures have been wi   </text>
<query_num> 1006 </query_num>
<text>   d several kilometers so drivers can find another route (e.g., take an exit to avoid a congested part of a highway). Despite the great potential of VANET applications, security has long been a concern =-=[10, 18, 24, 25, 28, 31]-=-, and thus it is imperative to provide functionality to validate an event reported by vehicles in both types of applications. Although the IEEE 1609.2 [14] standard is proposed to secure VANETs using   </text>
<query_num> 1007 </query_num>
<text>   d several kilometers so drivers can find another route (e.g., take an exit to avoid a congested part of a highway). Despite the great potential of VANET applications, security has long been a concern =-=[10, 18, 24, 25, 28, 31]-=-, and thus it is imperative to provide functionality to validate an event reported by vehicles in both types of applications. Although the IEEE 1609.2 [14] standard is proposed to secure VANETs using   Time and location information is required in each event description E. The information can be provided by the Global Positioning Sys1 VANETs can leverage multiple keys per vehicle to provide privacy =-=[28, 33]-=-. However, only one key pair is valid at any given time to prevent Sybil attacks [8] where one vehicle poses as many vehicles.tem (GPS), which is available in many vehicles nowadays and necessary for shold detected” while in fact n&amp;lt;τ (a large FP rate of P [D =1|n &amp;lt;τ] &amp;gt;δ1). We assume jamming and denial-of-service attacks can be mitigated by techniques such as spread spectrum [6], channel switching =-=[28]-=- or adaptive authentication [30]; providing reliable wireless communication is outside the scope of this paper. We do not consider deflation attacks, where an attacker covers up the occurrence of an e ts as an unique identifier of an alert, and allows our scheme to detect a threshold number of vehicles by estimating the number of distinct alerts. In VANETs, authorities assign key pairs to vehicles =-=[28]-=-. This prevents an attacker from selecting a specific public key as part of a decision changing attack; vehicles are limited to the public keys assigned to them. The advantage of hashing the above rat ly (e.g., every 5 minutes) to prevent long-term location tracing. When vehicles are unable to connect to keying authorities on a frequent basis, the vehicles are allowed to preload multiple key pairs =-=[28]-=-. Let TPK be the average time length between a public key is known by its owner and the key is being used. For example, TPK = 6 months when vehicles download a year worth key pairs for the next year d   </text>
<query_num> 1008 </query_num>
<text>   e-hop alerts are counted or assumes all alerts are available for analysis regardless how the alert distribution works. Dietzel et al. adopts the notion of data-centric trust [29] for event validation =-=[7]-=-. However, their scheme results in high dissemination delay. In contrast, our protocol enables a bandwidth-efficient solution to promptly distribute alerts and provides the event validity indicator fo   </text>
<query_num> 1009 </query_num>
<text>   ll avoid this road, providing the selfish driver with an improved driving experience. Counting the number of vehicles that report an event allows a recipient to evaluate the validity of a VANET event =-=[13, 16, 29]-=-. For example, a traffic jam reported by 2 vehicles is likely to be fake (or just started), but alerts from 50 vehicles is a strong indicator of road congestion. Particularly, we focus on threshold-ba  counting (detailed discussion on probabilistic counting is provided in Section 2). VANET event validation. The number of alerts from nearby vehicles is a strong indicator of the validity of an event =-=[13, 16, 29]-=-. However, prior work either focuses on one-hop-relevant applications where only one-hop alerts are counted or assumes all alerts are available for analysis regardless how the alert distribution works   </text>
<query_num> 1010 </query_num>
<text>   nopsis set may also include alerts that are already known to others. To avoid transmitting such redundant alerts and thus further optimize the message exchange protocol, we instead use a Bloom filter =-=[4]-=- as the digest. A Bloom filter allows constant time membership queries. Hence, the vehicle can reduce bandwidth usage by identifying absent alerts in the sender’s synopsis set, and only broadcast thos   </text>
<query_num> 1011 </query_num>
<text>   s proposed schemes for probabilistic counting to estimate the total number of items (e.g., unique elements in a database) based on a single pass over the data while requiring significantly less space =-=[1, 3, 9]-=-. In this paper, we leverage such counting schemes to perform probabilistic threshold-based event validation where a vehicle that receives a small subset of alerts can distinguish between a small numb obabilistic counting schemes’ trade-offs and limitations. Probabilistic counting selects several representative elements, or a synopsis [22], as an estimator for the total number of distinct elements =-=[1, 3, 9]-=-. The synopsis summarizes the entire element set and thus permits estimation of the total size. Probabilistic counting provides a trade-off between synopsis size and accuracy: the more elements in the  the space of random items, not over the entire distribution of n, i.e., n is taken as given. In this paper, we consider four error-bounded probabilistic counting schemes (KeepAll, AMS [1], FM sketch =-=[9]-=-,Table 1: Error bounded probabilistic counting schemes. ɛ&amp;lt; 1 for z-smallest and FM sketch. w&amp;gt;4 for AMS. The right most column shows the approximate size of a synopsis when n = 10000, ɛ = 0.1, δ =0.05  the approach where every unique item is part of the synopsis. Due to space limitations, we only provide a summary of z-smallest below, and refer readers to the original publications for more details =-=[1, 3, 9]-=-. After the example and a discussion of the accuracy and efficiency trade-off for probabilistic counting schemes, we discuss how maliciously crafted inputs can cause probabilistic counting schemes to   </text>
<query_num> 1012 </query_num>
<text>   s proposed schemes for probabilistic counting to estimate the total number of items (e.g., unique elements in a database) based on a single pass over the data while requiring significantly less space =-=[1, 3, 9]-=-. In this paper, we leverage such counting schemes to perform probabilistic threshold-based event validation where a vehicle that receives a small subset of alerts can distinguish between a small numb obabilistic counting schemes’ trade-offs and limitations. Probabilistic counting selects several representative elements, or a synopsis [22], as an estimator for the total number of distinct elements =-=[1, 3, 9]-=-. The synopsis summarizes the entire element set and thus permits estimation of the total size. Probabilistic counting provides a trade-off between synopsis size and accuracy: the more elements in the BL(n) BU (n) synopsis size KeepAll n n n 10000 z-smallest n(1 − ɛ) n(1 + ɛ) O( ln(1/δ) ɛ2 ) 128 ln(1/δ) AMS n/w wn 2(1/2−2/w) 2 150 ln 1/δ ln n FM sketch n(1 − ɛ) n(1 + ɛ) O( ɛ2 ) 1700 and z-smallest =-=[3]-=-) which satisfy such requirements as examples for theoretical analysis and simulation. KeepAll is the approach where every unique item is part of the synopsis. Due to space limitations, we only provid stic counting schemes, we discuss how maliciously crafted inputs can cause probabilistic counting schemes to produce unrealistically large estimates. Probabilistic Counting Example. Bar-Yossef et al. =-=[3]-=- proposed using the z th -smallest hash value (vz) as an estimator of the number of distinct elements (n). The intuition is that if the hashes of the elements are uniformly distributed in [0, 1], the   </text>
<query_num> 1013 </query_num>
<text>   s proposed schemes for probabilistic counting to estimate the total number of items (e.g., unique elements in a database) based on a single pass over the data while requiring significantly less space =-=[1, 3, 9]-=-. In this paper, we leverage such counting schemes to perform probabilistic threshold-based event validation where a vehicle that receives a small subset of alerts can distinguish between a small numb obabilistic counting schemes’ trade-offs and limitations. Probabilistic counting selects several representative elements, or a synopsis [22], as an estimator for the total number of distinct elements =-=[1, 3, 9]-=-. The synopsis summarizes the entire element set and thus permits estimation of the total size. Probabilistic counting provides a trade-off between synopsis size and accuracy: the more elements in the ns: Generation: SG(.) S = SG(I), where S ⊆ I. Fusion: SF(.,.) SF(S1, S2) = SG(I1 ∪ I2) when S1 = SG(I1) and S2 = SG(I2). Evaluation: SE(.) ñ = SE(S). Pr[BL(n) ≤ ñ ≤ BU (n)] &amp;gt; 1 − δ, (1) where δ is in =-=[0, 1]-=-, and BL(·) and BU (·) are monotonically increasing functions that indicate the lower bound and the upper bound of ñ, respectively. This probability is taken over the space of random items, not over t  the approach where every unique item is part of the synopsis. Due to space limitations, we only provide a summary of z-smallest below, and refer readers to the original publications for more details =-=[1, 3, 9]-=-. After the example and a discussion of the accuracy and efficiency trade-off for probabilistic counting schemes, we discuss how maliciously crafted inputs can cause probabilistic counting schemes to  f et al. [3] proposed using the z th -smallest hash value (vz) as an estimator of the number of distinct elements (n). The intuition is that if the hashes of the elements are uniformly distributed in =-=[0, 1]-=-, the expected number of hashes falling into [0,vz] is vzn. Hence, the estimator is ñ = z/vz. For example, if the resulted hash set is {0.05, 0.1, 0.15, 0.2,...}, with elements perfectly uniformly dis   </text>
<query_num> 1014 </query_num>
<text>   st storm [23] — severe link-layer contention and collision due to an excessive number of replicated messages. Various techniques have been proposed to alleviate the broadcast storm problem in general =-=[17, 23, 32, 35]-=-. Built upon existing broadcast storm solutions, we describe a customized message exchange protocol that can further reduce the bandwidth overhead by suppressing redundant broadcasts of synopses. For   </text>
<query_num> 1015 </query_num>
<text>   τ (a large FP rate of P [D =1|n &amp;lt;τ] &amp;gt;δ1). We assume jamming and denial-of-service attacks can be mitigated by techniques such as spread spectrum [6], channel switching [28] or adaptive authentication =-=[30]-=-; providing reliable wireless communication is outside the scope of this paper. We do not consider deflation attacks, where an attacker covers up the occurrence of an event by dropping alerts or jammi   </text>
</top>
<top>
<paper_num> 11 </paper_num>
<paper_title>   Key Agreement Using Statically Keyed Authenticators.  </paper_title>
<paper_abstract>   A family of authenticators based on static shared keys is  identified and proven secure. The authenticators can be used in a variety  of settings, including identity-based ones. Application of the authenticators  to Di#e-Hellman variants in appropriate groups leads to  authenticated key agreement protocols which have attractive properties  in comparison with other proven-secure protocols. We explore two key  agreement protocols that result.  </paper_abstract>
<query_num> 1101 </query_num>
<text>   an be used to generate efficient protocols with similar properties to some existing ones, but with the benefit of a formal security proof. In a later refinement of the technique, Canetti and Krawczyk =-=[9]-=- designed a MAC-based authenticator which uses a pre-existing shared secret as the MAC key. In our earlier work [7], we focussed on deniability properties of protocols resulting from taking an identit ators on the basic Diffie-Hellman protocol. 2 Authenticators and the Canetti–Krawczyk Model In this section we describe the modular approach to protocol proofs [2] and the Canetti–Krawczyk (CK) model =-=[9]-=-. We aim to give an informal understanding of how the approach works, sufficient to follow the rest of the paper. However, we necessarily omit the formal details and refer the interested reader to the versary can distinguish between the correct key in a test session and a random string of the same length is no more than 1/2 plus a negligible function in the security parameter. Canetti and Krawczyk =-=[9]-=- show that a protocol that is SK-secure in the AM is transformed into an SK-secure protocol in the UM if an authenticator is used. In order to explain what an authenticator is, we must first define th euristically not to disturb the protocol security. 3 Two Authenticators Our MT-authenticators can be viewed as variants of the MT-authenticator based on MACs that was proposed by Canetti and Krawczyk =-=[9]-=-. The format of our authenticators is shown in Fig. 1. On successful completion of the protocol, B will output ‘B received m from A’. In Fig. 1 k is a security parameter, m is the message to be transm se of this section is to prove that our two methods of generating FAB still produce authenticators (in the sense of Defn. 3). A second difference between our approach and that of Canetti and Krawczyk =-=[9]-=- is that we replace the MAC by a hash function where the static shared key is included as an input to the hash. The reason for this is that it makes both the authenticator and the proofsa little simpl   </text>
<query_num> 1102 </query_num>
<text>   ct answer at that point. Therefore we can improve the success probability to the following: Pr(J ) ≥ ɛ1(k). 3.2 Authenticator using Identity-Based Static Keys Using the notation of Boneh and Franklin =-=[6]-=-, we let G1 be an additive group of prime order q and G2 be a multiplicative group of the same order q. We assume the existence of an efficiently computable, non-degenerate, bilinear map ê from G1 × G dded to itself a times, also called scalar multiplication of Q by a. As a consequence of bilinearity, we have that, for any Q, W ∈ G1 and a, b ∈ Zq: ê(aQ, bW ) = ê(Q, W ) ab = ê(abQ, W ). We refer to =-=[1, 6, 11]-=- for a more comprehensive description of how these groups, pairings and other parameters should be selected in practice for efficiency and security. In this setting, the initialisation function I on i   </text>
<query_num> 1103 </query_num>
<text>   dded to itself a times, also called scalar multiplication of Q by a. As a consequence of bilinearity, we have that, for any Q, W ∈ G1 and a, b ∈ Zq: ê(aQ, bW ) = ê(Q, W ) ab = ê(abQ, W ). We refer to =-=[1, 6, 11]-=- for a more comprehensive description of how these groups, pairings and other parameters should be selected in practice for efficiency and security. In this setting, the initialisation function I on i   </text>
<query_num> 1104 </query_num>
<text>   formal security proof. In a later refinement of the technique, Canetti and Krawczyk [9] designed a MAC-based authenticator which uses a pre-existing shared secret as the MAC key. In our earlier work =-=[7]-=-, we focussed on deniability properties of protocols resulting from taking an identity-based approach to obtaining keys for the MACbased authenticator of Canetti and Krawczyk [9]. In this paper, we pr our identity-based MT-authenticator to the basic Diffie-Hellman protocol and then optimising is the identity-based key agreement protocol shown as Protocol 2. Its properties are explored in detail in =-=[7]-=-, in particular its strong deniability feature. Shared Information: Fixed key FAB derived from static Diffie-Hellman key: FAB = ê(QA, QB) s ; A generator P of G1. A B rA ∈R Zq TA = rAP Verify hash A,   </text>
<query_num> 1105 </query_num>
<text>   n both parties cooperate with each other, if one party defects by revealing its random input, then the other cannot deny taking part. Following the definition of deniable encryption by Canetti et al. =-=[8]-=- we may say that a two-party protocol is deniable for party A, if a legitimate party B could have simulated the protocol without the presence of A. Since all protocols using our authenticators can be   </text>
<query_num> 1106 </query_num>
<text>   ngs. In this section we give more detail of how to apply our identity-based authenticator to derive an identity-based key agreement protocol. We compare this protocol with a protocol proven secure in =-=[10]-=-. We assume the same algebraic setting as in Section 3.2. In principle we can use Diffie-Hellman in any group for the AM protocol. However, it is practical to choose a group that is already implemente  Protocol 4], the parties exchange ephemeral values rAQA and rBQB and calculate the shared secret ZAB = ê(QA, QB) s(rA+rB) . MACs are used in a standard way to provide key confirmation. Protocol 4 of =-=[10]-=- requires each party to compute one elliptic curve pairing and two elliptic curve multiplications (equivalent to exponentiations in a multiplicative group). Therefore the computational effort required rotocol instances take place between the same parties, Protocol 2 can cache and re-use the same FAB value and therefore requires only one pairing for all of these protocols. In contrast Protocol 4 of =-=[10]-=- requires one pairing for every protocol run, even between the same parties. Although Protocol 4 of [10] does not provide forward secrecy, Chen and Kudla do provide alternatives with this property. Wi entity of the other party in order to derive the session key. Therefore it seems that, like our protocols, identity protection against active adversaries cannot be efficiently achieved. Protocol 4 of =-=[10]-=- does provide confirmation of knowledge of the peer entity. 6 Conclusion We have shown that the CK model can be profitably used to design novel, provably secure key exchange protocols. We obtain proto   </text>
<query_num> 1107 </query_num>
<text>   of any new protocol. In recent years the number of key establishment protocols that carry a security proof has increased enormously. Most popular has been the model introduced by Bellare and Rogaway =-=[3, 4]-=- and later refined by themselves and others. In the modular approach to protocol design and proof [2], Bellare, Canetti and Krawczyk introduced the notion of an authenticator as a protocol translator.   </text>
<query_num> 1108 </query_num>
<text>   ol 1 has better computational and bandwidth efficiency than SIGMA. An exact comparison relies on the details of the signature scheme used in SIGMAsand the size of various parameters. The MQV protocol =-=[14]-=- has slightly smaller computational requirements in total but currently has no published security proof. In terms of bandwidth, Protocol 1 and the UMP also seem to be optimal. The only components incl   </text>
<query_num> 1109 </query_num>
<text>   oof has increased enormously. Most popular has been the model introduced by Bellare and Rogaway [3, 4] and later refined by themselves and others. In the modular approach to protocol design and proof =-=[2]-=-, Bellare, Canetti and Krawczyk introduced the notion of an authenticator as a protocol translator. Protocols may be proven secure in an ideal model (the so-called authenticated links model, or simply and an authenticator. We remark that, despite the extensive theoretical framework that has been built up, there have been few new protocols proven secure as a result of this technique. Bellare et al. =-=[2]-=- designed two general-purpose authenticators, one based on signatures and the other based on public key encryption. They showed how these authenticators can be used to generate efficient protocols wit at result from using these two authenticators on the basic Diffie-Hellman protocol. 2 Authenticators and the Canetti–Krawczyk Model In this section we describe the modular approach to protocol proofs =-=[2]-=- and the Canetti–Krawczyk (CK) model [9]. We aim to give an informal understanding of how the approach works, sufficient to follow the rest of the paper. However, we necessarily omit the formal detail able from π with A.sDefinition 3. An authenticator is a mapping of protocols that transforms a protocol π in the AM to a protocol π ′ in the UM such that π ′ emulates π. In common with Bellare et al. =-=[2]-=-, we require the indistinguishability in Defn. 2 to be computational. That is, there should be no efficient algorithm that can distinguish the output of the two protocols. All the authenticators we ta nding and receiving of a message m entails it being first stored in, and then removed from, the message store M. A UM protocol is an MT-authenticator if it emulates the AM protocol MT. Bellare et al. =-=[2]-=- showed that the mapping of protocols obtained by replacing each message M in an AM protocol by an MT-authenticator corresponding to M is an authenticator. Therefore, given an SK-secure protocol in th   </text>
<query_num> 1110 </query_num>
<text>   protocol that is secure in the AM, and using various optimisations, we obtain two concrete protocols that are provably secure in the UM. We compare our first protocol with the Unified Model protocol =-=[5]-=- and with the SIGMA protocol of Krawczyk [13]. We compare our second protocol to recent protocols of Chen and Kudla [10]. Analysis shows that our protocols are competitive with these existing protocol col based on static Diffie-Hellman 4.2 Comparison with Related Protocols The most similar proven secure protocol to Protocol 1 is the Unified Model Protocol (UMP) analysed by Blake-Wilson and Menezes =-=[5]-=- in the Bellare–Rogaway model. In the variant of the UMP that we consider, the shared secret ZAB is equal to g rArB ||FAB, the concatenation of ephemeral and static Diffie-Hellman keys. A MAC key (for   </text>
<query_num> 1111 </query_num>
<text>   ssumption, that the gap-DH problem is hard. With this assumption we may allow V access to an oracle that will distinguish between Diffie-Hellman triples and random triples in G. The gap-DH assumption =-=[15]-=- is that CDHP is still hard even given access to this oracle. In this case V can test if U hassasked a critical query (one involving g xy ) of the random oracle and can abort the protocol run with cer   </text>
</top>
<top>
<paper_num> 12 </paper_num>
<paper_title>   Compiling Language Definitions: The ASF+SDF Compiler  </paper_title>
<paper_abstract>   Machine Interfaces|The Bottleneck Eect  High-level transformations have to be applied with extreme care, especially if their purpose is to simplify the compiler by reducing the number of dierent constructs Compiling Language Denitions: The ASF+SDF Compiler  11 that have to be handled later on. For instance, by rst transforming conditional rewrite rules to unconditional ones or associative list matching to term matching, the compiler can be simplied considerably, but at the expense of a serious degradation in the performance of the generated code. Similarly, transformation of default rules (which can be applied only when all other rules fail) to sets of ordinary rewrite rules that catch the same cases would lead to very inecient code. These transformations would perhaps be appropriate in a formal semantics of ASF+SDF, but in a compiler they cause a bottleneck whose eect is hard to undo at a later stage.  Since it would require a high-level transformation phase of the above kind, the compiler does not generate code for the Abstract Rewrite Machine (ARM) [Fokkink et al. 1998]. In fact, any xed abstract machine interface is a potential bottleneck in the compilation process. The modularization advantage gained by introducing it may be oset by a serious loss in opportunities for generating ecient code. This happens when, in the words of Franz [1994, Sec. 2], \the code generator eectively needs to reconstruct at considerable expense, information that was more easily accessible in the front-end, but lost in the transition to the intermediate representation.&amp;quot;  The factors involved in the use of an abstract machine have a qualitatively dierent character. The abstract machine interface facilitates construction and verication  of the compiler, but possibly at the expens...  </paper_abstract>
<query_num> 1201 </query_num>
<text>   995] (railroad safety) |Action Semantics [=-=van Deursen 1994-=-] (programming language semantics) |Manifold [=-=Rutten and Thiebaux 1992-=-], ToolBus [=-=Bergstra and Klint 1998-=-] (coordination languages) |ALMA-0 [=-=Apt et al. 1998-=-] (backtracking and search) |Languages of the ASF+SDF Meta-Environment itself [=-=van den Brand et al. 2001-=-]: |SDF (syntax denition) |Box (prettyprinting specication) |ASF+SDF (language denition|this   </text>
<query_num> 1202 </query_num>
<text>   c specication. Rewriting is also important in functional programming, program transformation and optimization, and equational theorem proving. Useful theoretical surveys of rewriting are [=-=Klop 1992; Dershowitz and Jouannaud 1990-=-], but we assume only a basic understanding of rewrite systems on the part of the reader. In addition to regular rewrite rules, ASF+SDF features conditional rewrite rules, associative (  </text>
<query_num> 1203 </query_num>
<text>   m Transformation |Interactive program transformation for Clean [=-=van den Brand et al. 1995-=-] and Prolog [=-=Brunekreef 1996-=-] |PIM [=-=Field 1992-=-] (compiler toolkit) |Automatic program transformation for C++ [=-=Dinesh et al. 2001-=-] Software Renovation |Description of the multiplicity of languages and dialects encountered in software renovation applications such as Cobol (including embedded languages like SQL and CICS) [=-=van Deu -=-  </text>
<query_num> 1204 </query_num>
<text>   mple is constant caching. 4. MAJOR DESIGN CONSIDERATIONS The design of the compiler was in  </text>
<query_num> 1205 </query_num>
<text>   nd algebraic specication. Rewriting is also important in functional programming, program transformation and optimization, and equational theorem proving. Useful theoretical surveys of rewriting are [=-=Klop 1992; Dershowitz and Jouannaud 1990-=-], but we assume only a basic understanding of rewrite systems on the part of the reader. In addition to regular rewrite rules, ASF+SDF features conditional rewrite rule   </text>
<query_num> 1206 </query_num>
<text>   o run, namely, Clean [=-=Plasmeijer and van Eekelen 1994; Smetsers et al. 1991-=-], Elan [=-=Kirchner and Moreau 2001-=-], Haskell [=-=Peyton Jones et al. 1993; Peyton Jones 1996-=-], Maude [=-=Clavel et al. 1999-=-], Opal [=-=Didrich et al. 1994-=-], and SML [=-=Appel 1992-=-]. This article is organized as follows: brief survey of the ASF+SDF language (Sec. 2); general compilation scheme (Sec. 3); major design considerations for the ASF+SDF compiler   Strictness annotations  Polymorphic typing Maude Rewriting logic language Interpreted [=-=Clavel et al. 1999-=-]  First-order Core Maude  Re  </text>
<query_num> 1207 </query_num>
<text>   piler and compare its performance with that of other rewrite system and functional language compilers we were able to run, namely, Clean [=-=Plasmeijer and van Eekelen 1994; Smetsers et al. 1991-=-], Elan [=-=Kirchner and Moreau 2001-=-], Haskell [=-=Peyton Jones et al. 1993; Peyton Jones 1996-=-], Maude [=-=Clavel et al. 1999-=-], Opal [=-=Didrich et al. 1994-=-], and SML [=-=Appel 1992-=-]. This article is organized as follows: brief survey of the ASF+SD  via [=-=Plasmeijer and van Eekelen 1994-=-]  Higher-order ABC abstract [=-=Smetsers et al. 1991-=-]  Lazy graph rewriting  Strictness annotations machine  Polymorphic typing Elan Rewriting logic language C [=-=Kirchner and Moreau 2001-=-]  First-order  Strategy specication  AC-rewriting Haskell Functional language C [=-=Peyton Jones et al. 1993-=-]  Higher-order [=-=Peyton Jones 1996-=-]  Lazy  Strictness annotations  Polymorphic typing   </text>
</top>
<top>
<paper_num> 13 </paper_num>
<paper_title>   Sense of direction in distributed computing.  </paper_title>
<paper_abstract>   Sense of Direction is a property of labeled graphs which has been shown to have a definite  impact on computability and complexity in systems of communicating entities, and whose  applicability ranges from the analysis of graph classes to distributed object systems. The full  consequences of this property are still not known; in fact, the ongoing investigations continue to  bring new (often surprising) results, to establish unsuspected links with other research and/or application  areas, and to pose more questions than they answer. The aim of this paper is to provide  a view of the current status of research, describing some of the relevant results, and providing  pointers to future research directions.  1 Introduction  In its more general formulation, a distributed system is a collection of computational entities communicating by exchanging finite amounts of information, which we shall call messages. The exact nature of the entities (i.e., processors, processes, network nodes, agents,...  </paper_abstract>
<query_num> 1301 </query_num>
<text>   8 Depth First Traversal Election Fault-Tolerant Election local orientation with topological awareness O(n 2 ) \Theta(n log n) [41] \Omega\Gamma n log k + kf) [38] chordal SD \Theta(n) [25] \Theta(n) =-=[45, 49, 62]-=- \Theta(n + kf) [50] arbitrary SD \Theta(n) [26] \Theta(n) [18] ? Table 2: Complete Networks. Wake-Up Time \Theta Bits local orientation \Theta(n 2 ) with topological awareness chordal SD O(n log 2 n) ion in complexity is achievable in presence of any SD [18]. Subsequent research has focused on improvements of the constant factor of the message complexity and on improvements of the time complexity =-=[49, 62]-=-, and adding fault-tolerance to the requirements [48, 50, 52, 61]. In the results for fault-tolerant election, f is the number of faulty processors, k is the number of initiators, and the faults are &amp;quot;   </text>
<query_num> 1302 </query_num>
<text>   For example, in absence of any type-T knowledge (and assuming distinct identities), the problems LE , SPT and MF are equivalent [59]; the communication complexity of B is not greater than that of LE =-=[26]-=-; and, clearly, any solution to DFT is also a solution to B. 6 consistency Broadcast Election Depth First Traversal Spanning Tree Minimum Finding local orientation \Omega\Gamma e) [26] \Omega\Gamma e   [58] neighboring SD \Theta(min(e; n 1+\Theta(1) )) [3] O(n log n) [44] \Theta(n) [60] chordal SD \Theta(n) [25] O(n log n) [48] c-group SD \Theta(n) [65] arbitrary SD 2n \Gamma 2 [26] 3n log n +O(n) =-=[26]-=- Table 1: Universal Protocols The studies have concentrated both on the inherent limits (lower bounds) and on the achievable results (upper bounds) for solving these problems in presence of sense of d entation alone), and in absence of additional type-T knowledge, both the broadcast and the depth first traversal problems require\Omega\Gamma e) messages even if the entities have distinct identities =-=[26]-=-; such a bound can be easily achieved (e.g., see [14]). An improvement in the message complexity had been shown to exist if each entity has a distinct identity and knows the identities of all its neig  to the larger class of labelings with commutative group sense of direction [65]. Finally, it has been shown that any sense of direction allows these two problems to be solved with only O(n) messages =-=[26]-=-. On a related note, it has been shown that, without sense of direction, a O(n) broadcast protocol exists only if the size of messages is unbounded [3]. 3.2.2 Election, Spanning Tree and Minimum Findi ful chordal sense of direction [25]. These two labeling-dependent results have been recently fully generalized; in fact, any sense of direction allows for O(n log n) solutions to these three problems =-=[26]-=-. 3.3 Network-specific Protocols Most of the original research on sense of direction has focused on specific network topologies. Clearly, the solution algorithms, to be executable, require at each ent   </text>
<query_num> 1303 </query_num>
<text>   O(n) messages in all those graphs, since they are sparse; thus, the research has focused on the multiplicative constant. In tori, the 3n + 1 trivial upperbound has been reduced first to 2n + O( p n) =-=[16]-=-, and subsequently to 10 7 n + O( p (n) [20]; the current lower-bound is 8 7 n +O(1)[20]. It is interesting to note that there exists graphs in which broadcasting without sense of direction can be don   </text>
<query_num> 1304 </query_num>
<text>   [39] jSj = O(log n) O(n log n) O(n log n) [54] S = f1; p ng O(n) [47] O(n) Fabri&amp;Tel95 in [65] arbitrary S; jSj = k O(nk + n log n) ? Table 4: Chordal Ring. Broadcast Election local orientation O(n) =-=[15, 22]-=- O(n log log n) [19] neighboring SD O(n) [3] O(n log log n) dimensional SD O(n) \Theta(n) [24, 57, 64, 66] chordal SD O(n) O(1) [24] arbitrary SD O(n) O(n) [18] Table 5: Hypercube. sense of direction.  to the broadcast problem. There is widespread suspicion that it might be insensitive also in the case of the election problem. It is interesting to note that all algorithms for hypercubes without SD =-=[15, 19, 22]-=- exploit the well-known fact that the hypercube of size n has a dominating set of size O( n log n ); active nodes build such a dominating set to efficiently solve problems such as broadcast and electi   </text>
<query_num> 1305 </query_num>
<text>   ]. Subsequent research has focused on improvements of the constant factor of the message complexity and on improvements of the time complexity [49, 62], and adding fault-tolerance to the requirements =-=[48, 50, 52, 61]-=-. In the results for fault-tolerant election, f is the number of faulty processors, k is the number of initiators, and the faults are &amp;quot;fail-stop&amp;quot;. In the case of synchronous complete networks, a good   </text>
<query_num> 1306 </query_num>
<text>   able to the system entities (the nodes). The computational power of sense of direction in anonymous systems has been studied in [28] and shown to be linked to the notion of surrounding, introduced in =-=[29]-=-. 11 Before describing the notion of surrounding, we need the following definition of labeled graph isomorphism. Given two labeled graphs (G = (V; E); ) and (G 0 = (V 0 ; E 0 );s0 ), a labeled graph i fourth label is necessary. 7.2 Surrounding Symmetry, Cayley Graphs, and Minimality. Necessary and sufficient conditions for a labeled graph (G; ) to have minimal sense of direction have been given in =-=[29]-=- for the case of symmetric labelings and in [33] for the case of non-symmetric labelings. We remind that a labelingsis symmetric if there exists a bijection / : \Sigma ! \Sigma such that for each hx;  ma \Gamma \Gamma \Gamma Figure 7: Petersen&amp;apos;s graph. Section 4): a regular graph with symmetric labeling has a minimal SD iff it is surrounding symmetric (i.e., if every node has the same surrounding) =-=[29]-=-. On the other hand, a labeled graph (G; ) is surrounding symmetric iff it is a Cayley graph with a Cayley labeling. A Cayley graph is a graph where nodes correspond to the elements of a group and edg e group. As a consequence, we have a characterization of minimal SD in terms of Cayley graphs: a regular graph with symmetric labeling has a minimal SD iff it is a Cayley graph with a Cayley labeling =-=[29]-=-; this result holds also for directed graphs [10]. b b b a a a b a t t t t Figure 8: A minimal SD with a non-symmetric labeling. If the graph is not symmetric, this necessary and sufficient condition   </text>
<query_num> 1307 </query_num>
<text>   bitrary S; jSj = k O(nk + n log n) ? Table 4: Chordal Ring. Broadcast Election local orientation O(n) [15, 22] O(n log log n) [19] neighboring SD O(n) [3] O(n log log n) dimensional SD O(n) \Theta(n) =-=[24, 57, 64, 66]-=- chordal SD O(n) O(1) [24] arbitrary SD O(n) O(n) [18] Table 5: Hypercube. sense of direction. However, the same bound has been obtained assuming only local orientation and topological awareness [15,  nt on the O(n log n) bound implied by the universal protocol. In presence of the dimensional sense of direction (traditional for hypercubes), several \Theta(n) election algorithms have been presented =-=[24, 57, 64, 66]-=-. Most of these solutions exploit the implicit region partitioning of the topology and an efficient and implicit scheme to compute and represent shortest paths. These results can be obtained also by u ing the more recent modular technique of [21]. An interesting additional result is that the chordal sense of direction would break the symmetry of the hypercube so to allow a surprising O(1) solution =-=[24]-=-. Recently, the question of whether a O(n) solution exists with any sense of direction has been positively answered [18]. 3.4 Insensitivity to Sense of Direction Some graphs have particular structural   </text>
<query_num> 1308 </query_num>
<text>   dication of the impact of SD. This result 8 Depth First Traversal Election Fault-Tolerant Election local orientation with topological awareness O(n 2 ) \Theta(n log n) [41] \Omega\Gamma n log k + kf) =-=[38]-=- chordal SD \Theta(n) [25] \Theta(n) [45, 49, 62] \Theta(n + kf) [50] arbitrary SD \Theta(n) [26] \Theta(n) [18] ? Table 2: Complete Networks. Wake-Up Time \Theta Bits local orientation \Theta(n 2 ) w   </text>
<query_num> 1309 </query_num>
<text>   impler problems can be used as building blocks for solving the complex ones; for instance, a solution to SPT can be constructed starting from any network traversal protocol (e.g., a solution to DFT ) =-=[40]-=-. Furthermore, large-scale problems and applications are likely to be (and are) modularly decomposed in terms of these basic problems; hence, the intrinsic cost of solving such basic problems has a di   </text>
<query_num> 1310 </query_num>
<text>   irection is that, for every string ff 2 \Sigma + of labels, there exists a walk with that sequence of labels; actually, there is one starting from every node. 16 7.1 Cycle Symmetry and Minimality. In =-=[23]-=- it has been shown that the existence of a minimal sense of direction is related to the notion of cycle symmetry of a graph: a graph is cycle symmetric if each node belongs to the same number of cycle ae(x); ae(y)) 2 E iff (x; y) 2 E. Since a vertex transitive graph is also cycle symmetric, a weaker necessary condition for the existence of minimal sense of direction is given by vertex transitivity =-=[23]-=-. Notice that the two notions are not equivalent; in fact it has been recently shown that a cycle symmetric graph is not necessarily vertex transitive [34]. x y b) a) J J J J J J &amp;apos; &amp;apos; &amp;apos; &amp;apos; &amp;apos; &amp;apos; S S S S S ion for the existence of minimal sense of direction. As an example consider the Petersen&amp;apos;s graph G (see Figure 7); this graph is vertex transitive and, thus, cycle symmetric, but it has been shown in =-=[23]-=- that for any labelingswhich uses three labels, (G; ) does not have sense of direction. In order to have a sense of direction a fourth label is necessary. 7.2 Surrounding Symmetry, Cayley Graphs, and   </text>
<query_num> 1311 </query_num>
<text>   ists graphs in which broadcasting without sense of direction can be done with the absolutely minimum cost: n-1 messages. A characterization of regular networks 10 having this property can be found in =-=[17]-=-; the general problem to decide whether a graph has this property is NP-hard [17]. For some classes of regular networks with non-constant degree, it is clearly possible that their structural propertie   </text>
<query_num> 1312 </query_num>
<text>   larly interesting open problem is the characterization of the computability relationship between sense of direction and other forms of consistency; some results have already been recently obtained in =-=[30]-=-. 5 Impact of Sense of Direction on Systems: Object Naming Sense of direction can be used to construct an efficient naming scheme in systems of distributed objects [4]. A distributed object system con l networks, wireless communication media, etc. Preliminary investigations have identified the notion of backward sense of direction as the relevant consistency property and results are very promising =-=[30]-=-. The variety in the nature and scope of the problems still open, several indicated throughout the paper, offers a unique opportunity for researchers with very distinct background. A portal on Structu   </text>
<query_num> 1313 </query_num>
<text>   log n) O(n log n) [54] S = f1; p ng O(n) [47] O(n) Fabri&amp;Tel95 in [65] arbitrary S; jSj = k O(nk + n log n) ? Table 4: Chordal Ring. Broadcast Election local orientation O(n) [15, 22] O(n log log n) =-=[19]-=- neighboring SD O(n) [3] O(n log log n) dimensional SD O(n) \Theta(n) [24, 57, 64, 66] chordal SD O(n) O(1) [24] arbitrary SD O(n) O(n) [18] Table 5: Hypercube. sense of direction. However, the same b tained assuming only local orientation and topological awareness [15, 22]. For the election problem, in absence of sense of direction, a solution has been developed which uses O(n log log n) messages =-=[19]-=-; this is an improvement on the O(n log n) bound implied by the universal protocol. In presence of the dimensional sense of direction (traditional for hypercubes), several \Theta(n) election algorithm  to the broadcast problem. There is widespread suspicion that it might be insensitive also in the case of the election problem. It is interesting to note that all algorithms for hypercubes without SD =-=[15, 19, 22]-=- exploit the well-known fact that the hypercube of size n has a dominating set of size O( n log n ); active nodes build such a dominating set to efficiently solve problems such as broadcast and electi   </text>
<query_num> 1314 </query_num>
<text>   o identified. Subsequent investigations have reduced the size of the structure necessary to achieve linear election algorithms from O(log n) to O(log log n) [39], to O(log log log n) [54], to O(1) in =-=[47, 65]-=-. 3.3.3 Hypercube Network The results for broadcasting in the hypercube show a general insensitivity of this topology to sense of direction. In fact, broadcasting can be performed with O(n) messages i ection Election local orient. chordal SD S = f1; 2; : : : ; kg O(nk + n log n) O( n k log n k ) [1] jSj = O(log log n) O(n log n) O(n) [39] jSj = O(log n) O(n log n) O(n log n) [54] S = f1; p ng O(n) =-=[47]-=- O(n) Fabri&amp;Tel95 in [65] arbitrary S; jSj = k O(nk + n log n) ? Table 4: Chordal Ring. Broadcast Election local orientation O(n) [15, 22] O(n log log n) [19] neighboring SD O(n) [3] O(n log log n) di ch area. The knowledge to date is limited to some vertex-symmetric regular graphs of constant degree. In particular, the election problem can be solved with the exchange of \Theta(n) messages in tori =-=[55, 47]-=-, in the class of chordal rings h1; p ni n [47], as well as in butterflies and cube-connected-cycles [21]. Broadcast can be easily performed with O(n) messages in all those graphs, since they are spar raditional sense of direction for cliques, hypercubes, 15 Cost Type of SD Complete Network \Theta(n 2 ) [63] chordal SD Hypercube \Theta(n log n) [63] dimensional SD Torus ( p n \Theta p n) \Theta(n) =-=[47]-=- compass SD Double Loop h1; p ni n \Theta(n) [47] chordal SD Table 6: Constructing SD. and tori, exchanges at least\Omega\Gamma e \Gamma 1 2 n) messages in a network with n nodes and e edges. The same   </text>
<query_num> 1315 </query_num>
<text>   of the positive impact is already quite strong, as this paper hopes to have shown. For example, the notion of sense of direction has just recently found applications in the area of self-stabilization =-=[12, 13]-=-. The scope of the investigations on sense of direction and on the consistency of the labelings continues to expand. For example, research has recently focused to consider cases where the communicatio   </text>
<query_num> 1316 </query_num>
<text>   osts of many distributed algorithms, the problem of its construction (e.g., in preprocessing phase) is of obvious relevance. The problem of constructing SD in an unlabeled network has been studied in =-=[63, 65]-=-. It has been shown that any algorithm constructing the traditional sense of direction for cliques, hypercubes, 15 Cost Type of SD Complete Network \Theta(n 2 ) [63] chordal SD Hypercube \Theta(n log  e \Gamma 1 2 n) messages in a network with n nodes and e edges. The same result hold for constructing a chordal calSD in arbitrary graphs. Algorithms matching the lower bounds have also been proposed =-=[63]-=-; most of the solutions cannot avoid a complete flooding of the communication links in the network. These results are not attractive for dense topologies, like cliques and hypercubes, while they are m   </text>
<query_num> 1317 </query_num>
<text>   ready been recently obtained in [30]. 5 Impact of Sense of Direction on Systems: Object Naming Sense of direction can be used to construct an efficient naming scheme in systems of distributed objects =-=[4]-=-. A distributed object system consists of a collection of objects and relations between objects; each object has a state (for example, expressed by local variables) and a behavior (set of actions it m d on the location of the objects (e.g., internet, email addresses). On the other hand, locality of names and independence on the location are two very desirable characteristics of a naming scheme. In =-=[4]-=-, a naming scheme in which a name of an object depends neither on its state nor on its characteristics, not even on its location, has been proposed based on sense of direction as described in the foll   </text>
<query_num> 1318 </query_num>
<text>   t of research has been devoted to the study of computability in anonymous systems; i.e., the study of what problems can be solved when there are no distinct identifiers associated to the nodes (e.g., =-=[2, 9, 5, 42, 43, 51, 68]-=-). Clearly, which problems can be solved depends on many factors including the structural properties of the system as well as the amount and type of structural knowledge available to the system entiti   </text>
<query_num> 1319 </query_num>
<text>   the motivation for studying the interplay between the topology of the system and the properties that a labeling must satisfy for having sense of direction. Such an interplay has been investigated in =-=[31, 32]-=-, where the following questions have been considered and answered: ffl In which graphs, every labeling guarantees the existence of sense of direction? ffl For which graphs, the fact that the labeling   </text>
<query_num> 1320 </query_num>
<text>   the system as well as the amount and type of structural knowledge available to the system entities (the nodes). The computational power of sense of direction in anonymous systems has been studied in =-=[28]-=- and shown to be linked to the notion of surrounding, introduced in [29]. 11 Before describing the notion of surrounding, we need the following definition of labeled graph isomorphism. Given two label sense of direction. In particular, what is computable in these systems depends on the number of distinct surroundings, as well as on their multiplicity (i.e., how many nodes have a given surrounding) =-=[28]-=-. An interesting property is that the classes of nodes having the same surrounding have the same cardinality ae (G;) , called surrounding symmetricity of (G; ). Notice that the availability at node u  rty of sense of direction is the following: Given a graph G and a type-T knowledge K, if a problem P is K-solvable in G=G then P is solvable in W=G (where knowledge of the coding function is assumed) =-=[28]-=-. In other words, with weak sense of direction, no other knowledge is needed. This result is based on the fact that, in a labeled graph with weak sense of direction (and without any other information  irection and sense of direction are equivalent; this is not so from a complexity viewpoint. 4.3 Results on Specific Problems. The computational power of anonymous systems with WSD has been studied in =-=[28]-=- considering some specific, typical problems: leader election LE , edge election EE , spanning tree construction SPT , topology recognition T R, complete topology recognition CT . 12 Solvability of th   </text>
<query_num> 1321 </query_num>
<text>   tion of minimal SD in terms of Cayley graphs: a regular graph with symmetric labeling has a minimal SD iff it is a Cayley graph with a Cayley labeling [29]; this result holds also for directed graphs =-=[10]-=-. b b b a a a b a t t t t Figure 8: A minimal SD with a non-symmetric labeling. If the graph is not symmetric, this necessary and sufficient condition does not hold. Consider, for example, the labeled   </text>
<query_num> 1322 </query_num>
<text>   uld be stable and reliable, that is, names must remain valid, and keep denoting the same objects when the system changes its state. In order to achieve these goals, the existing naming schemes (e.g., =-=[36, 53, 67]-=-) either use an approach based on global names (e.g., each object uses as its name an intrinsic property which distinguishes it from the other objects) or they use hierarchical names based on the loca   </text>
</top>
<top>
<paper_num> 14 </paper_num>
<paper_title>   Robust incentive techniques for peer-to-peer networks.  </paper_title>
<paper_abstract>   Lack of cooperation (free riding) is one of the key problems that confronts today&amp;apos;s P2P systems. What makes this problem particularly difficult is the unique set of challenges that P2P systems pose: large populations, high turnover, asymmetry of interest, collusion, zero-cost identities, and traitors. To tackle these challenges we model the P2P system using the Generalized Prisoner&amp;apos;s Dilemma (GPD), and propose the Reciprocative decision function as the basis of a family of incentives techniques. These techniques are fully distributed and include: discriminating server selection, maxflowbased subjective reputation, and adaptive stranger policies. Through simulation, we show that these techniques can drive a system of strategic users to nearly optimal levels of cooperation.  </paper_abstract>
<query_num> 1401 </query_num>
<text>   behavior among selfish players. The unique challenges imposed by peer-to-peer systems inspired additional body of work [5] [37], mainly in the context of packet forwarding in wireless ad-hoc routing =-=[8]-=- [27] [30] [35], and file sharing [15] [31]. Friedman and Resnick [13] consider the problem of zero-cost identities in online environments and find that in such systems punishing all newcomers is inev   </text>
<query_num> 1402 </query_num>
<text>   case of our work) is consistent with the objective of incentivizing the desired behavior among selfish players. The unique challenges imposed by peer-to-peer systems inspired additional body of work =-=[5]-=- [37], mainly in the context of packet forwarding in wireless ad-hoc routing [8] [27] [30] [35], and file sharing [15] [31]. Friedman and Resnick [13] consider the problem of zero-cost identities in o   </text>
<query_num> 1403 </query_num>
<text>   e of our work) is consistent with the objective of incentivizing the desired behavior among selfish players. The unique challenges imposed by peer-to-peer systems inspired additional body of work [5] =-=[37]-=-, mainly in the context of packet forwarding in wireless ad-hoc routing [8] [27] [30] [35], and file sharing [15] [31]. Friedman and Resnick [13] consider the problem of zero-cost identities in online   </text>
<query_num> 1404 </query_num>
<text>   ess for complexity of computation. We show that the maxflow-based algorithm scales better than private history in the presence of colluders without the centralized trust required in previous work [9] =-=[20]-=-. • Adaptive Stranger Policy: Zero-cost identities allows noncooperating peers to escape the consequences of not cooperating and eventually destroy cooperation in the system if not stopped. We show th cular player for the entire population to observe it, thus scales better to large populations and high turnovers, and also tolerates asymmetry of interest. Some examples of shared history schemes are =-=[20]-=- [23] [28]. Figure 7 shows the effectiveness of shared history under high turnover rates. In this figure, we fix the population size and vary the turnover rate. While selective players with private hi   </text>
<query_num> 1405 </query_num>
<text>   failure rate increase when users do not share their resources [3]. In a wireless ad-hoc network, overall packet latency and loss rate increase when nodes refuse to forward packets on behalf of others =-=[26]-=-. Further examples are file preservation [25], discussion boards [17], online auctions [16], and overlay routing [6]. In many of these systems, users have natural disincentives to cooperate because co   </text>
<query_num> 1406 </query_num>
<text>   gic players results in the socially desired outcome. Distributed Algorithmic Mechanism Design seeks solutions within this framework that are both fully distributed and computationally tractable [12]. =-=[10]-=- and [11] are examples of applying DAMD to BGP routing and multicast cost sharing. More recently, DAMD has been also studied in dynamic environments [38]. In this context, demonstrating the superiorit   </text>
<query_num> 1407 </query_num>
<text>   iveness for complexity of computation. We show that the maxflow-based algorithm scales better than private history in the presence of colluders without the centralized trust required in previous work =-=[9]-=- [20]. • Adaptive Stranger Policy: Zero-cost identities allows noncooperating peers to escape the consequences of not cooperating and eventually destroy cooperation in the system if not stopped. We sh whitewashing can be nearly eliminated from the system. The adaptive stranger policy does this without requiring centralized allocation of identities, an entry fee for newcomers, or rate-limiting [13] =-=[9]-=- [25]. • Short-term History: History also creates the possibility that a previously well-behaved peer with a good reputation will turn traitor and use his good reputation to exploit other peers. The p ng all newcomers is inevitable. Using a theoretical model, they demonstrate that such a system can converge to cooperation only for sufficiently low turnover rates, which our results confirm. [6] and =-=[9]-=- show that whitewashing and collusion can have dire consequences for peer-to-peer systems and are difficult to prevent in a fully decentralized system. Some commercial file sharing clients [1] [2] pro   </text>
<query_num> 1408 </query_num>
<text>   l with collusion, entities can compute reputation subjectively, where player A weighs player B’s opinions based on how much player A trusts player B. Our subjective algorithm is based on maxflow [24] =-=[32]-=-. Maxflow is a graph theoretic problem, which given a directed graph with weighted edges asks what is the greatest rate at which “material” can be shipped from the source to the target without violati   </text>
<query_num> 1409 </query_num>
<text>   player for the entire population to observe it, thus scales better to large populations and high turnovers, and also tolerates asymmetry of interest. Some examples of shared history schemes are [20] =-=[23]-=- [28]. Figure 7 shows the effectiveness of shared history under high turnover rates. In this figure, we fix the population size and vary the turnover rate. While selective players with private history   </text>
<query_num> 1410 </query_num>
<text>   rs results in the socially desired outcome. Distributed Algorithmic Mechanism Design seeks solutions within this framework that are both fully distributed and computationally tractable [12]. [10] and =-=[11]-=- are examples of applying DAMD to BGP routing and multicast cost sharing. More recently, DAMD has been also studied in dynamic environments [38]. In this context, demonstrating the superiority of a co   </text>
<query_num> 1411 </query_num>
<text>   strategic players results in the socially desired outcome. Distributed Algorithmic Mechanism Design seeks solutions within this framework that are both fully distributed and computationally tractable =-=[12]-=-. [10] and [11] are examples of applying DAMD to BGP routing and multicast cost sharing. More recently, DAMD has been also studied in dynamic environments [38]. In this context, demonstrating the supe   </text>
<query_num> 1412 </query_num>
<text>   their resources [3]. In a wireless ad-hoc network, overall packet latency and loss rate increase when nodes refuse to forward packets on behalf of others [26]. Further examples are file preservation =-=[25]-=-, discussion boards [17], online auctions [16], and overlay routing [6]. In many of these systems, users have natural disincentives to cooperate because cooperation consumes their own resources and ma ewashing can be nearly eliminated from the system. The adaptive stranger policy does this without requiring centralized allocation of identities, an entry fee for newcomers, or rate-limiting [13] [9] =-=[25]-=-. • Short-term History: History also creates the possibility that a previously well-behaved peer with a good reputation will turn traitor and use his good reputation to exploit other peers. The peer c   </text>
<query_num> 1413 </query_num>
<text>   to-peer (P2P) systems rely on cooperation among selfinterested users. For example, in a file-sharing system, overall download latency and failure rate increase when users do not share their resources =-=[3]-=-. In a wireless ad-hoc network, overall packet latency and loss rate increase when nodes refuse to forward packets on behalf of others [26]. Further examples are file preservation [25], discussion boa n this context is the “tragedy of the commons” [18] where resources are under-provisioned due to selfish users who free-ride on the system’s resources, and is especially common in large networks [29] =-=[3]-=-. The problem has been extensively studied adopting a game theoretic approach. The prisoners’ dilemma model provides a natural framework to study the effectiveness of different strategies in establish   </text>
<query_num> 1414 </query_num>
<text>   unique challenges imposed by peer-to-peer systems inspired additional body of work [5] [37], mainly in the context of packet forwarding in wireless ad-hoc routing [8] [27] [30] [35], and file sharing =-=[15]-=- [31]. Friedman and Resnick [13] consider the problem of zero-cost identities in online environments and find that in such systems punishing all newcomers is inevitable. Using a theoretical model, the   </text>
</top>
<top>
<paper_num> 15 </paper_num>
<paper_title>   Universal Approximation Capability of Cascade Correlation for Structures.  </paper_title>
<paper_abstract>   Cascade correlation (CC) constitutes a training method for neural networks which determines the weights as well as the neural architec-ture during training. Various extensions of CC to structured data have been proposed: recurrent cascade correlation (RCC) for sequences, re-cursive cascade correlation (RecCC) for tree structures with limited ¤ This work has been partially supported by MIUR grant 2002093941 004. We would like to thank two anonymous referees for profound and valuable comments on an earlier version of the manuscript.  </paper_abstract>
<query_num> 1501 </query_num>
<text>   , 2002; Lodhi et al., 2000; Watkins, 1999). Alternatively, recurrent neural models can be extended to complex dependencies which occur in tree structures, graphs, or spatial data (=-=Baldi et al., 1999; Frasconi, Gori, and Sperduti, 1998; Goller and Küchler, 1996; Hammer, 2000; Micheli, Sona, and Sperduti, 2000; Pollastri et al., 2002; Sperduti, Majidi, and Starita, 1996; Sperduti and Starita, 1997; Wakuya and Zurada, 2001-=-). These st for tree structured inputs (=-=Bianucci et al., 2000; Sperduti, Majidi, and Starita, 1996; Micheli, 2003-=-). Recursive networks constitute the analogous extension of recurrent networks to tree structures (=-=Frasconi, Gori, and Sperduti, 1998; Goller and Küchler, 1996; Sperduti and Starita, 1997-=-). For the latter, the universal approximation capability has been established in (=-=Hammer, 2000-=-). However, RecCC networks share the particularly e ecursive structure, recursive processing of each vertex in the tree starting from the leaves up to the root is possible. This leads to the notion of recursive networks (RecNNs) as introduced e.g. in (=-=Frasconi, Gori, and Sperduti, 1998; Goller and Küchler, 1996; Sperduti and Starita, 1997-=-) and recursive cascade correlation (RecCC) as proposed in (=-=Bianucci et al., 2000; Sperduti, Majidi, and Starita, 1996-=-). As beforehand, for RecNN  e structures as input, respectively. They have successfully been applied for various type of classification and regression problems on structured data ranging from picture processing up to chemistry (=-=Frasconi, Gori, and Sperduti, 1998-=-). However, it is well known that training recurrent networks faces severe problems (=-=Bengio, Simard, and Frasconi, 1994-=-) and the generalization ability might be considerably worse compared to standard   </text>
<query_num> 1502 </query_num>
<text>   =-=ie, Eskin, and Noble, 2002; Lodhi et al., 2000; Watkins, 1999-=-). Alternatively, recurrent neural models can be extended to complex dependencies which occur in tree structures, graphs, or spatial data (=-=Baldi et al., 1999; Frasconi, Gori, and Sperduti, 1998; Goller and Küchler, 1996; Hammer, 2000; Micheli, Sona, and Sperduti, 2000; Pollastri et al., 2002; Sperduti, Majidi, and Starita, 1996; Sperduti and Starita, 1997-=- models have been reported in different areas such as image and document processing, logic, natural language processing, chemistry, DNA processing, homology detection, or protein structure prediction (=-=Baldi et al., 1999; Diligenti, Frasconi, and Gori, 2003; Goller, 1997; 3sJaakkola, Diekhans, and Haussler, 2000; Leslie, Eskin, and Noble, 2002; Lodhi et al., 2000; deMauro et al., 2003; Pollastri et al., 2002;-=- Sturt e ogic or chemistry. Therefore extensions of recursive models for sequences and tree structures have been proposed which also take additional contextual information into account. The model proposed in (=-=Baldi et al., 1999-=-) trains two RNNs simultaneously to predict the secondary structure of proteins. The networks are thereby transforming the given sequence in reverse directions such that the full information is availa nlike the above mentioned approaches, contextual information is here directly integrated into the recursive part of the network such that the information can be used in an earlier stage than e.g. in (=-=Baldi et al., 1999; Pollastri et al., 2002-=-). Thanks to these characteristics, in the cascade approach, the processing of contextual information is efficiently extended, with respect to the previous approches, from sequ red data such as sequences or grids have been proposed in the literature and successfully been applied e.g. in bioinformatics where often spatial data, instead of temporal data, has to be dealt with (=-=Baldi et al., 1999; Micheli, Sona, and Sperduti, 2000; Pollastri et al., 2002; Wakuya and Zurada, 2001-=-). However, unlike contextual cascade models, structure processing is restricted to special subclasses and integrati   </text>
<query_num> 1503 </query_num>
<text>   Frasconi, and Gori, 2003; Goller, 1997; 3sJaakkola, Diekhans, and Haussler, 2000; Leslie, Eskin, and Noble, 2002; Lodhi et al., 2000; deMauro et al., 2003; Pollastri et al., 2002; Sturt et al., 2003; =-=Vullo and Frasconi, 2003-=-). Here, we are interested in a variant of so-called recursive networks which constitute a straightforward generalization of well known recurrent networks to tree structures and for which excellent re   </text>
<query_num> 1504 </query_num>
<text>   and successfully applied as prediction models for chemical structures: recursive cascade correlation and contextual recursive cascade correlation, respectively (=-=Bianucci et al., 2000; Micheli, 2003;Micheli, Sona, and Sperduti, 2003 b; Micheli, Sona, and Sperduti, 2002; Micheli, Sona, and Sperduti, 2000; Sperduti, Majidi, and Starita, 1996-=-). The efficient training algorithm of CC can be transferred to these latter models and very y depends on the recurrence being restricted and it cannot be transferred to fully recursive networks. This generalization of RecCC networks towards context integration has recently been proposed in (=-=Micheli, Sona, and Sperduti, 2003 b; Micheli, Sona, and Sperduti, 2002-=-). These models are applied to prediction tasks for chemical structures, and they compare favorably to models without context integration. Plots of the principal co racy in experiments can be accompanied by more general theoretical results on the universal approximation capibility of the model. A first study of the enlarged capacity of the model can be found in (=-=Micheli, Sona, and Sperduti, 2003 b; Micheli, 2003-=-), where it is shown that, in a directed acyclic graph, contextual RecCC (CRecCC) takes into account the context of each vertex, including both successors and predecessors of the verte xtual RecCC can implement classes of contextual functions where a desired output for a given vertex may depend on the predecessors of the vertex, up to the whole structure (contextual transductions) (=-=Micheli, Sona, and Sperduti, 2003 b; Micheli, 2003-=-). Of course, contextual RecCC networks can still implement all the functions which are computable by RCC models (=-=Micheli, Sona, and Sperduti, 2003b; Micheli, 2003-=-). We will show in th . I.e. the root has height�. Note that a tree is just a specific form of DPAGs where each vertex (except for the root) possesses exactly one parent. Contextual recursive cascade correlation (CRecCC) (=-=Micheli, Sona, and Sperduti, 2003b; Micheli, Sona, and Sperduti, 2002; Micheli, 2003-=-) can also take into account information about parents. Assume DPAGs with limited fan-in at most and fan-out at most ¤ ¡ are given. In the following,   </text>
<query_num> 1505 </query_num>
<text>   as shown above. Because FNNs with activation function¦ and linear outputs possess the universal approximation property, we can find a FNN which maps these encodings exactly to the outputs given by� (=-=Sontag, 1992-=-). The composition of these two networks can be interpreted as a RCC network because of Lemma 3.1 with the desired properties. ¨ Note that we have restricted ourselves to outputs in�, for convenience.   </text>
<query_num> 1506 </query_num>
<text>   e design of efficient training algorithms which yield sparse models with good generalization ability is a key issue for recursive models for structures. Cascade correlation (CC) has been proposed by (=-=Fahlmann and Lebiere, 1990-=-) as a particularly efficient training algorithm for feedforward networks which simultaneously determines the architecture of the network and the parameters. Hidden neurons are created and trained con  solutions with good generalization ability. CC proved to be particularly appropriate 4sfor training problems where standard feedforward network training is difficult such as the two spirals problem (=-=Fahlmann and Lebiere, 1990-=-). The main difference of CC networks with respect to feedforward networks lies in the particularly efficient constructive training method. Since any feedforward network structure can be embedded into rained at each stage, this algorithm is very fast. Moreover, it usually provides excellent classification results with a small number of hidden neurons which, because of the specific training method (=-=Fahlmann and Lebiere, 1990-=-), serve as error correcting units. This training scheme is also used for all further cascade correlation models which are introduced in this article. Since we are here not interested in the specific   </text>
<query_num> 1507 </query_num>
<text>   l data (=-=Baldi et al., 1999; Frasconi, Gori, and Sperduti, 1998; Goller and Küchler, 1996; Hammer, 2000; Micheli, Sona, and Sperduti, 2000; Pollastri et al., 2002; Sperduti, Majidi, and Starita, 1996; Sperduti and Starita, 1997; Wakuya and Zurada, 2001-=-). These structure processing models have the advantage that complex data such as spatial data, tree structures, or graphs can be directly tackled and thus a possibly time con ajidi, and Starita, 1996; Micheli, 2003). Recursive networks constitute the analogous extension of recurrent networks to tree structures (=-=Frasconi, Gori, and Sperduti, 1998; Goller and Küchler, 1996; Sperduti and Starita, 1997-=-). For the latter, the universal approximation capability has been established in (=-=Hammer, 2000-=-). However, RecCC networks share the particularly efficient iterative training scheme of RCC and CC netwo  tree starting from the leaves up to the root is possible. This leads to the notion of recursive networks (RecNNs) as introduced e.g. in (=-=Frasconi, Gori, and Sperduti, 1998; Goller and Küchler, 1996; Sperduti and Starita, 1997-=-) and recursive cascade correlation (RecCC) as proposed in (=-=Bianucci et al., 2000; Sperduti, Majidi, and Starita, 1996-=-). As beforehand, for RecNN we assume direct dependence of neuron from neurons�sin   </text>
<query_num> 1508 </query_num>
<text>   models can be extended to complex dependencies which occur in tree structures, graphs, or spatial data (=-=Baldi et al., 1999; Frasconi, Gori, and Sperduti, 1998; Goller and Küchler, 1996; Hammer, 2000; Micheli, Sona, and Sperduti, 2000; Pollastri et al., 2002; Sperduti, Majidi, and Starita, 1996; Sperduti and Starita, 1997; Wakuya and Zurada, 2001-=-). These structure processing models have the advantage that complex data such as spat  recursive cascade correlation and contextual recursive cascade correlation, respectively (=-=Bianucci et al., 2000; Micheli, 2003; Micheli, Sona, and Sperduti, 2003b; Micheli, Sona, and Sperduti, 2002; Micheli, Sona, and Sperduti, 2000; Sperduti, Majidi, and Starita, 1996-=-). The efficient training algorithm of CC can be transferred to these latter models and very good results have been reported. Here, we are interested in the in-pri r to integrate all available information for each vertex of a regular two-dimensional lattice. Similar problems for sequences have been tackled by a cascade correlation approach including context in (=-=Micheli, Sona, and Sperduti, 2000-=-). Another model with similar ideas has been proposed in (=-=Wakuya and Zurada, 2001-=-). Note that several of these models simply construct disjoint recursive networks according to different causality assu uences or grids have been proposed in the literature and successfully been applied e.g. in bioinformatics where often spatial data, instead of temporal data, has to be dealt with (=-=Baldi et al., 1999; Micheli, Sona, and Sperduti, 2000; Pollastri et al., 2002; Wakuya and Zurada, 2001-=-). However, unlike contextual cascade models, structure processing is restricted to special subclasses and integration of contextual information is her   </text>
<query_num> 1509 </query_num>
<text>   nted by a specific locally recurrent network architecture with the Heaviside activation function can be limited regardless of the number of neurons and thus restrictions apply. The work presented in (=-=Giles et al., 1995-=-) explicitely investigates RCC architectures and shows that these networks equipped with the Heaviside function or a monotonically increasing activation function cannot implement all finite automata.  res constitutes a proper subclass of the class of recurrent network architectures. This might restrict the capability of these type of networks compared to the general model, as demonstrated e.g. in (=-=Giles et al., 1995-=-). Hence the question was whether the restriction of the recurrence in cascade models poses severe restrictions on the applicability of the models because possibly only very simple functions can be re bability in spite of their restricted recurrence. So if fundamental differences concerning the capacity of these systems in comparison to their fully recurrent counterparts exist (as demonstrated in (=-=Giles et al., 1995-=-)) then these fundamental differences will not manifest for any given finite set of training data or for any given finite time horizon or height of the trees, respectively, provided enough neurons are   </text>
<query_num> 1510 </query_num>
<text>   orted in different areas such as image and document processing, logic, natural language processing, chemistry, DNA processing, homology detection, or protein structure prediction (=-=Baldi et al., 1999;Diligenti, Frasconi, and Gori, 2003; Goller, 1997; 3sJaakkola, Diekhans, and Haussler, 2000; Leslie, Eskin, and Noble, 2002; Lodhi et al., 2000; deMauro et al., 2003; Pollastri et al., 2002; Sturt et al., 2003; Vullo and Frasconi, 2003 -=-  </text>
<query_num> 1511 </query_num>
<text>   ralization ability might be fundamentally worse compared to simple pattern recognition tools because the capacity of the models (measured e.g. in terms of the VC dimension) depends on the structures (=-=Hammer, 2001-=-). For recursive models for structures, the long-term dependencies problem is often reduced because structural representations via trees or graphs yield more compact representation of data. However, t known that training recurrent networks faces severe problems (=-=Bengio, Simard, and Frasconi, 1994-=-) and the generalization ability might be considerably worse compared to standard feedforward networks (=-=Hammer, 2001-=-). Moreover, classification tasks for sequences and tree structures are likely more difficult than standard learning problems for feedforward networks due to the increased complexity of data. Hence co   </text>
<query_num> 1512 </query_num>
<text>   urrent neural models can be extended to complex dependencies which occur in tree structures, graphs, or spatial data (=-=Baldi et al., 1999; Frasconi, Gori, and Sperduti, 1998; Goller and Küchler, 1996; Hammer, 2000; Micheli, Sona, and Sperduti, 2000; Pollastri et al., 2002; Sperduti, Majidi, and Starita, 1996; Sperduti and Starita, 1997; Wakuya and Zurada, 2001-=-). These structure processing models have the advan networks to tree structures (=-=Frasconi, Gori, and Sperduti, 1998; Goller and Küchler, 1996; Sperduti and Starita, 1997-=-). For the latter, the universal approximation capability has been established in (=-=Hammer, 2000-=-). However, RecCC networks share the particularly efficient iterative training scheme of RCC and CC networks and thus recurrence is restricted to recursive connections of hidden neurons to neurons int unction from sequences into a real vector space can be approximated arbitrarily well, i.e. the distance of the desired output from the network output is at most¦for inputs of probability at least�� ¥(=-=Hammer, 2000-=-). We will later show that an analogous result holds also for RCC networks with restricted recurrence. 2.2 Recursive models Both, RNNs and RCC networks have been generalized to tree structures in the  s ��¡��¢¡����������spact notation shows how the dependencies can be transferred to larger fan-out ¤ : the operator is then substituted byscom�¡£¢�ch��£¡����������¡£¢�ch��£¡����. It has been shown in (=-=Hammer, 2000-=-) that fully recursive networks constitute ��¡£¢�¢¡���¤ universal approximators on tree structures. We will show in this article an analogous result for RecCC networks with multiplicative neurons. 2.3 C We start with the simplest case, recurrent networks for sequence processing. The key ideas of the proof borrow steps from the universal approximation proof of recursive neural networks as given in (=-=Hammer, 2000-=-). For convenience, we first outline the general steps of the proof some of which can be reused for recursive cascade correlation and contextual recursive cascade correlation. The key points are: (I)   </text>
</top>
<top>
<paper_num> 16 </paper_num>
<paper_title>   Pruning and Summarizing the Discovered Associations.  </paper_title>
<paper_abstract>   Association rules are a fundamental class of patterns that exist in data. The key strength of association rule mining is its completeness. It finds all associations in the data that satisfy the user specified minimum support and minimum confidence constraints. This strength, however, comes with a major drawback. It often produces a huge number of associations. This is particularly true for data sets whose attributes are highly correlated. The huge number of associations makes it very difficult, if not impossible, for a human user to analyze in order to identify those interesting/useful ones. In this paper, we propose a novel technique to overcome this problem. The technique first prunes the discovered associations to remove those insignificant associations, and then finds a special subset of the unpruned associations to form a summary of the discovered associations. We call this subset of associations the direction setting (DS) rules  as they set the directions that are followed by the...  </paper_abstract>
<query_num> 1601 </query_num>
<text>   24] do not perform summarization. There are also a number of other methods in classification research for rule pruning such as pessimistic error rate [21] and minimum description length based pruning =-=[15]-=-. In our work, we choose the widely used chisquare test statistics for rule pruning. [1] introduces a technique to remove two types of redundant rules, i.e., simple and strict redundancy. Essentially,   </text>
<query_num> 1602 </query_num>
<text>   ain. The system then finds those unexpected rules by comparing the user&amp;apos;s knowledge with the discovered rules. Again, these methods do not prune and do not attempt to summarize the association rules. =-=[25]-=- and [18] investigated using item constraints specified by the user in rule mining to generate only the relevant rules. Essentially, the constraints restrict the items or combination of items allowed   </text>
<query_num> 1603 </query_num>
<text>   ated. The user can now obtain a complete picture of the domain without being overwhelmed by a huge number of rules. 2. Related Work The problem of too many rules has been studied by many researchers. =-=[8]-=- proposed an approach to allow the user to specify what he/she wants to see using templates. The system then retrieves those match rules from the set of discovered rules. This method, however, does no In subjective interestingness research in data mining, [22, 11, 12, 19] proposed a number of methods for finding unexpected rules. Instead of asking the user to specify what he/she wants to see as in =-=[8]-=-, these approaches ask the user Pruning Significant rules Summarization DS rules Discovered large rules Non-DS rules to specify his/her existing knowledge about the domain. The system then finds those   </text>
<query_num> 1604 </query_num>
<text>   been applied successfully to a number of real-life applications. 1. Introduction Association rules are a class of important regularities in data. Association rule mining is commonly stated as follows =-=[2]-=-: Let I = {i 1 , ..., i n } be a set of items, and D be a set of data cases. Each data case consists of a subset of items in I. An association rule is an implication of the form X �� Y, where X �� I,   to know how other attributes are related to this target attribute (which can have many values) [13, 4]. With a target attribute, we can express an itemset as follows (instead of a set of items as in =-=[2]-=-): X �� y where y is an item (or a value) of the target attribute, and X is a set of items from the rest of the attributes. For To appear in ACM SIGKDD International Conference on Knowledge Discovery   </text>
<query_num> 1605 </query_num>
<text>   blic domain data sets are mainly used for classification research. However, classification can only give the user a partial picture of the data, i.e., many interesting/useful rules are not discovered =-=[20, 13]-=-. With our proposed technique, we show that association rule mining can now be effectively and practically applied to these data sets to give a complete picture of the underlying relationships in the   </text>
<query_num> 1606 </query_num>
<text>   h is the difference between the confidence of a rule and the confidence of any proper subrule with the same consequent. Those rules that do not meet this minimum improvement in confidence are pruned. =-=[24]-=- also proposed a related technique. Our pruning method is similar. However, we use chi-square test as the basis for pruning (minimum improvement can be easily incorporated in our framework). We will s   </text>
<query_num> 1607 </query_num>
<text>   n rule mining in such data is typically targeted at a specific attribute because the user normally wants to know how other attributes are related to this target attribute (which can have many values) =-=[13, 4]-=-. With a target attribute, we can express an itemset as follows (instead of a set of items as in [2]): X �� y where y is an item (or a value) of the target attribute, and X is a set of items from the  blic domain data sets are mainly used for classification research. However, classification can only give the user a partial picture of the data, i.e., many interesting/useful rules are not discovered =-=[20, 13]-=-. With our proposed technique, we show that association rule mining can now be effectively and practically applied to these data sets to give a complete picture of the underlying relationships in the  by our users. For all these data sets, even with a target attribute, the numbers of associations discovered are huge (note that our modified association rule miner is able to use the target attribute =-=[13]-=-). Many data sets cause combinatorial explosion. Due to this reason, we set a hard limit of 80,000 on the total number of large rules processed in memory. Even with such a large limit, mining cannot b 95% significance level for the c 2 test is commonly used [17]. A lower level of 90% is used to show the effect of this on the results. We used the minimum supports of 2% and 1% because it is shown in =-=[13]-=- that for these data sets, rules with these support thresholds are sufficiently predictive. Table 1 gave the results obtained using the significance level of 95% for the c 2 test and minsup = 1%. The   </text>
<query_num> 1608 </query_num>
<text>   n rule mining in such data is typically targeted at a specific attribute because the user normally wants to know how other attributes are related to this target attribute (which can have many values) =-=[13, 4]-=-. With a target attribute, we can express an itemset as follows (instead of a set of items as in [2]): X �� y where y is an item (or a value) of the target attribute, and X is a set of items from the  t we do not use minimum confidence in our framework, although we can also use it (see Section 6). Minimum confidence does not reflect the underlying relationship of the domain represented by the data =-=[4]-=-. Instead we use statistical correlation as the basis for finding rules that represent the fundamental relations of the domain. Figure 1 shows the conceptual flow of the proposed technique 1 , which c he advantage of association rules, its completeness, is lost. Clearly, a better approach is to summarize the discovered rules. From this summary, the user can obtain an overall picture of the domain. =-=[4]-=- proposed a rule pruning technique using minimum improvement, which is the difference between the confidence of a rule and the confidence of any proper subrule with the same consequent. Those rules th vement can be easily incorporated in our framework). We will see in Section 8 that even after pruning the number of rules left can still be very large. Summarization is thus important. The methods in =-=[4, 24]-=- do not perform summarization. There are also a number of other methods in classification research for rule pruning such as pessimistic error rate [21] and minimum description length based pruning [15 ssociation rules that use only one fixed attribute on the right hand side. Also our proposed technique does not use minimum confidence for rule generation (see the problems with minimum confidence in =-=[4]-=-), but statistical correlation (or significance). Other related work includes correlation rule mining in [5]. It uses chi-square test to measure the correlation. A correlation rule is a set of correla   </text>
<query_num> 1609 </query_num>
<text>   ng two conditions are met: 1. The rule&amp;apos;s support exceeds s. 2. The c 2 value for the rule with respect to the whole data does not exceed the c 2 value at the significance level c. Similar to those in =-=[5]-=-, we define three types of correlation of a rule. For this work, we also call them the directions of a rule. Definition 3 (type of correlation or direction): Positive correlation: if X and y of a rule   </text>
<query_num> 1610 </query_num>
<text>   set of discovered rules. This method, however, does not prune those insignificant rules and does not provide a summary of the discovered rules. In subjective interestingness research in data mining, =-=[22, 11, 12, 19]-=- proposed a number of methods for finding unexpected rules. Instead of asking the user to specify what he/she wants to see as in [8], these approaches ask the user Pruning Significant rules Summarizat   </text>
</top>
<top>
<paper_num> 17 </paper_num>
<paper_title>   Proof-Carrying Code.  </paper_title>
<paper_abstract>   Dependent session types allow us to describe not only properties of the I/O behavior of processes but also of the exchanged data. In this paper we show how to exploit dependent session types to express proof-carrying communication. We further introduce two modal operators into the type theory to provide detailed control about how much information is communicated: one based on traditional proof irrelevance and one integrating digital signatures.  </paper_abstract>
<query_num> 1701 </query_num>
<text>   be written as {x:unit | σ}, except that the proof object is always erased. Under some restrictions on σ, subset types can be seen as predicate-based type refinement as available, for example, in Fine =-=[16]-=- used for secure communication in distributed computation.10 Luis Caires, Frank Pfenning, and Bernardo Toninho 4 Affirmation In many distributed communicating systems there are tradeoffs between trus  constructs for concurrency. Their system is based on natural deduction and is substantially different from ours, and they consider neither dependent types nor unrestricted sessions. The work on Fine =-=[16]-=-, F7 [2], and more recently F* [15] has explored the integration of dependent and refinement types in a suite of fully fledged functional programming languages, with the aim of statically checking ass   </text>
<query_num> 1702 </query_num>
<text>   cated: one based on traditional proof irrelevance and one integrating digital signatures. Keywords: Process calculus, session types, proof irrelevance, proofcarrying code 1 Introduction Session types =-=[9]-=- provide high-level specifications for the communication behavior of interacting processes along bidirectional channels. Recently, logical foundations for session types have been established via Curry  we will briefly review dependent session types and the facilities they provide in terms of proof-carrying communication. Dependent session types [3, 17] are a conservative extension of session types =-=[9, 4, 5, 8]-=- that allow us to not only describe the behavior of processes in terms of their input and output behavior but also enable us to describe rich properties of the communicated data themselves. In [17], t   </text>
<query_num> 1703 </query_num>
<text>   communication behavior of interacting processes along bidirectional channels. Recently, logical foundations for session types have been established via Curry-Howard correspondences with linear logic =-=[4, 10]-=-. Besides clarifying and unifying concepts in session types, such logical underpinnings provide simple means for generalization. One such extension to dependent session types [3, 17] allows us to expr  check such signatures as authentic. Ours is one amongst several Curry-Howard interpretations connecting linear logic to concurrency. Perhaps closest to session types is work by Mazurak and Zdancewic =-=[10]-=- who develop a Curry-Howard interpretation of classical linear logic as a functional programming language with explicit constructs for concurrency. Their system is based on natural deduction and is su   </text>
<query_num> 1704 </query_num>
<text>   communication behavior of interacting processes along bidirectional channels. Recently, logical foundations for session types have been established via Curry-Howard correspondences with linear logic =-=[4, 10]-=-. Besides clarifying and unifying concepts in session types, such logical underpinnings provide simple means for generalization. One such extension to dependent session types [3, 17] allows us to expr  we will briefly review dependent session types and the facilities they provide in terms of proof-carrying communication. Dependent session types [3, 17] are a conservative extension of session types =-=[9, 4, 5, 8]-=- that allow us to not only describe the behavior of processes in terms of their input and output behavior but also enable us to describe rich properties of the communicated data themselves. In [17], t   </text>
<query_num> 1705 </query_num>
<text>   hat the verifier can sign a certificate containing the proposition [Πx:nat. f(x) ≤ x] and the server can reliably and efficiently check the signature. Our experience with a proof-carrying file system =-=[7]-=- shows that digitally signed certificates are much more compact and can be checked much more quickly than proofs themselves and are one of the cornerstones to make the architecture practical. In this   and pass the signed certificate (without the proof term) on to fpt. fpt〈λx. x〉. v〈nat → nat〉. v〈λx. x〉. v(c). fpt〈c〉. fpt(y). fpt(q). 0. In fact, the implementation of the proof-carrying file system =-=[7]-=- (PCFS) provides such a generic trusted service. In PCFS, the access control policy is presented as a logical theory in the access control logic. Access to a file is granted if a proof of a correspond   </text>
<query_num> 1706 </query_num>
<text>   rem can be constructed with the theory in access control logic and is presented to the file system. Such proofs are generally small when compared to proof-carrying code in the sense of Necula and Lee =-=[12, 11]-=- in which the type safety and memory safety of binary code is certified, but they are still too big to be transmitted and checked every time a file is accessed. Instead, we call upon the trusted verif   </text>
<query_num> 1707 </query_num>
<text>   ts for concurrency. Their system is based on natural deduction and is substantially different from ours, and they consider neither dependent types nor unrestricted sessions. The work on Fine [16], F7 =-=[2]-=-, and more recently F* [15] has explored the integration of dependent and refinement types in a suite of fully fledged functional programming languages, with the aim of statically checking assertions   </text>
<query_num> 1708 </query_num>
<text>   tting a proof q of y = f(y). But we do not want to erase this requirement entirely, of course, just avoid sending a proof term. We can do this by using the type-theoretic concept of proof irrelevance =-=[14, 13, 1]-=-. Generally, a type [A] (pronounced “bracket A”) is the type inhabited by proofs of A, all of which are identified. This is only meaningful if such proofs play no computational role, so there is some  esuppose or “bake in” any particular analysis or strategy, but formulate the type theory so that we can seemlessly move between different specifications. This is what a modality for proof irrelevance =-=[14, 13, 1]-=- in the type theory allows us to do. Proof irrelevance is a technique that allows us to selectively hide portions of a proof (and by the proofs-as-programs principle, portions of a program). The idea   </text>
<query_num> 1709 </query_num>
<text>   tting a proof q of y = f(y). But we do not want to erase this requirement entirely, of course, just avoid sending a proof term. We can do this by using the type-theoretic concept of proof irrelevance =-=[14, 13, 1]-=-. Generally, a type [A] (pronounced “bracket A”) is the type inhabited by proofs of A, all of which are identified. This is only meaningful if such proofs play no computational role, so there is some  esuppose or “bake in” any particular analysis or strategy, but formulate the type theory so that we can seemlessly move between different specifications. This is what a modality for proof irrelevance =-=[14, 13, 1]-=- in the type theory allows us to do. Proof irrelevance is a technique that allows us to selectively hide portions of a proof (and by the proofs-as-programs principle, portions of a program). The idea  mmunication. T † 2 � ∀p:(Σn:nat. unit). ∃q:(Σy:nat. unit). 1 P † 2 :: x : T2 � x(〈n, 〉). x〈 〈n + 1, 〈〉〉 〉. 0 This solution is popular in type theory, where Σx:τ. [σ] is a formulation of a subset type =-=[14]-=-, {x:τ | σ}. Conversely, bracket types [σ] can be written as {x:unit | σ}, except that the proof object is always erased. Under some restrictions on σ, subset types can be seen as predicate-based type   </text>
<query_num> 1710 </query_num>
<text>   we will briefly review dependent session types and the facilities they provide in terms of proof-carrying communication. Dependent session types [3, 17] are a conservative extension of session types =-=[9, 4, 5, 8]-=- that allow us to not only describe the behavior of processes in terms of their input and output behavior but also enable us to describe rich properties of the communicated data themselves. In [17], t   </text>
<query_num> 1711 </query_num>
<text>   with linear logic [4, 10]. Besides clarifying and unifying concepts in session types, such logical underpinnings provide simple means for generalization. One such extension to dependent session types =-=[3, 17]-=- allows us to express and enforce complex properties of data transmitted during sessions. In this paper we build upon dependent session types to model various aspects of systems employing certified co s Calculus 3 2 Dependent session types In this section we will briefly review dependent session types and the facilities they provide in terms of proof-carrying communication. Dependent session types =-=[3, 17]-=- are a conservative extension of session types [9, 4, 5, 8] that allow us to not only describe the behavior of processes in terms of their input and output behavior but also enable us to describe rich   </text>
<query_num> 1712 </query_num>
<text>   with linear logic [4, 10]. Besides clarifying and unifying concepts in session types, such logical underpinnings provide simple means for generalization. One such extension to dependent session types =-=[3, 17]-=- allows us to express and enforce complex properties of data transmitted during sessions. In this paper we build upon dependent session types to model various aspects of systems employing certified co s Calculus 3 2 Dependent session types In this section we will briefly review dependent session types and the facilities they provide in terms of proof-carrying communication. Dependent session types =-=[3, 17]-=- are a conservative extension of session types [9, 4, 5, 8] that allow us to not only describe the behavior of processes in terms of their input and output behavior but also enable us to describe rich o the client the nok signal, if the client has insufficient balance, or send the charge information to the bank and inform the client that the transaction went through. 5 Progress and Preservation In =-=[17]-=- we established the type safety results of progress and preservation for our dependent session type theory for an unspecified functional language. In fact, we made no mention of when reduction of the  r also being type safe in this sense (which can easily be seen to be the case for the connectives we have presented in this development). The proof of type preservation then follows the same lines of =-=[17]-=-, using a series of reduction lemmas that relate process reductions with parallel composition through an instance of the cut rule and appealing to the type preservation of the functional layer when ne peal to the reduction lemmas mentioned above (and to type preservation of the functional language when the premises of cut are of existential or universal type), which are presented in more detail in =-=[17]-=-. The case for the proof of progress is identical. The result in [17] combined with progress of the functional language establishes progress for the system of this paper. For the purpose of having a s   </text>
</top>
<top>
<paper_num> 18 </paper_num>
<paper_title>   An Entailment Relation for Reasoning on the Web.  </paper_title>
<paper_abstract>   Abstract. Reasoning on the Web is receiving an increasing attention because of emerging fields such as Web adaption and Semantic Web. Indeed, the advanced functionalities striven for in these fields call for reasoning capabilities. Reasoning on the Web, however, is usually done using existing techniques rarely fitting the Web. As a consequence, additional data processing like data conversion from Web formats (e.g. XML or HTML) into some other formats (e.g. classical logic terms and formulas) is often needed and aspects of the Web (e.g. its inherent inconsistency) are neglected. This article first gives requirements for an entailment tuned to reasoning on the Web. Then, it describes how classical logic’s entailment can be modified so as to enforce these requirements. Finally, it discusses how the proposed entailment can be used in applying logic programming to reasoning on the Web. 1  </paper_abstract>
<query_num> 1801 </query_num>
<text>   5.3 in formalising the satisfaction of semistructured expressions and atomic formulas in an interpretation. The following definition is inspired from [11,12] and refines the simulation considered in =-=[13]-=-. Recall that a (directed) rooted graph G = (V, E, r) consists in a set V of vertices, a set E of edges (i.e. ordered pairs of vertices), and a vertex r called the root of G such that there is in G a  roduced in the previous sections has been conceived so as to support such applications that are briefly described below. It provides with the semantics of a prototype Web query language called Xcerpt =-=[13,18]-=-. Dynamic Web pages. Dynamic Web pages are texts or data that are dynamically generated when called. They make it possible for different pages to share data thus ensuring data consistency and to gener age for XML and semistructured data. Xcerpt is based on the logic programming paradigm. The present paper is a contribution to Xcerpt’s semantics. Xcerpt’s operational semantics has been presented in =-=[13]-=-. Further query languages for XML and semistructured data are XSLT [19] and fxt [24]. XSLT is based on the matching of XML elements and on a built-in structural recursion of XML documents and it also   </text>
<query_num> 1802 </query_num>
<text>   a retrieval in terms of root-to-node data item traversals, i.e. a rather procedural approach. Other query languages for XML [5] and semistructured data do not build upon a navigational paradigm. UnQL =-=[21]-=- first proposed to express queries as terms (orspatterns) as in logic and in the present paper. Such query languages can be called “positional” because the relative positions of the data to retrieve,   </text>
<query_num> 1803 </query_num>
<text>   ion of rooted simulation is used below in Section 5.3 in formalising the satisfaction of semistructured expressions and atomic formulas in an interpretation. The following definition is inspired from =-=[11,12]-=- and refines the simulation considered in [13]. Recall that a (directed) rooted graph G = (V, E, r) consists in a set V of vertices, a set E of edges (i.e. ordered pairs of vertices), and a vertex r c   </text>
<query_num> 1804 </query_num>
<text>   ner. E.g. the clause p ← ¬p has a single minimal (inconsistent) paraconsistent model: {¬¬p}. Major advantages of the treatment of negation proposed above is that it extends the Stable Model Semantics =-=[17]-=- and gives it a “minimal model setting”: Proposition 2. Let P be a propositional program. Every consistent model of P (in the sense of Section 6.1) is stable. Every stable model of P (in the sense of  e is {A1, ¬¬A1, . . . , Ak, ¬¬Ak}. Proof. Proposition 2 follows from the characterisation of minimal (paraconsistent) models given below in Proposition 3 which rephrases the program transformation of =-=[17]-=- and extend it to general formulas. Proposition 3. Let M be a (paraconsistent) model of a set of formulas S. Let M = {¬E | E semistructured expression and E �∈ M}∪{¬¬¬E | L semistructured expression a erence procedures local to the formulas considered, thus well-suited to the Web. A third salient aspect of the entailment relation proposed in this paper is that it extends the Stable Model Semantics =-=[17]-=- and gives it a “minimal model setting”. Finally, the article has briefly discussed how the proposed entailment relation can be used in applying logic programming to reasoning on the Web. The work rep   </text>
<query_num> 1805 </query_num>
<text>   p (cf. e.g. [16,28,29]). [30] describes in more detail this approach in the framework of classical logic. Proposition 3 is an adaptation to the nonstandard models of Definition 4 of a result given in =-=[31]-=- for classical logic. 9 Conclusion This article has first given requirement for logics for reasoning on the Web as needed for emerging applications such as Semantic Web and adaptive Web systems. Then,   </text>
<query_num> 1806 </query_num>
<text>   pretations are inconsistent interpretations satisfying ¬¬Cj but not satisfying Cj. Such interpretations “gives room” for interpreting so-called cycles of recursion through negation with an odd length =-=[16]-=- in a quite natural manner. E.g. the clause p ← ¬p has a single minimal (inconsistent) paraconsistent model: {¬¬p}. Major advantages of the treatment of negation proposed above is that it extends the  structured data. The approach to nonmonotonic reasoning described in the present paper is reminiscent of a widespread, often empirical approach consisting in “duplicating” every predicate p (cf. e.g. =-=[16,28,29]-=-). [30] describes in more detail this approach in the framework of classical logic. Proposition 3 is an adaptation to the nonstandard models of Definition 4 of a result given in [31] for classical log   </text>
<query_num> 1807 </query_num>
<text>   roduced in the previous sections has been conceived so as to support such applications that are briefly described below. It provides with the semantics of a prototype Web query language called Xcerpt =-=[13,18]-=-. Dynamic Web pages. Dynamic Web pages are texts or data that are dynamically generated when called. They make it possible for different pages to share data thus ensuring data consistency and to gener   </text>
<query_num> 1808 </query_num>
<text>   structured data. The approach to nonmonotonic reasoning described in the present paper is reminiscent of a widespread, often empirical approach consisting in “duplicating” every predicate p (cf. e.g. =-=[16,28,29]-=-). [30] describes in more detail this approach in the framework of classical logic. Proposition 3 is an adaptation to the nonstandard models of Definition 4 of a result given in [31] for classical log   </text>
<query_num> 1809 </query_num>
<text>   ta. The approach to nonmonotonic reasoning described in the present paper is reminiscent of a widespread, often empirical approach consisting in “duplicating” every predicate p (cf. e.g. [16,28,29]). =-=[30]-=- describes in more detail this approach in the framework of classical logic. Proposition 3 is an adaptation to the nonstandard models of Definition 4 of a result given in [31] for classical logic. 9 C   </text>
</top>
<top>
<paper_num> 19 </paper_num>
<paper_title>   Additive Approximation for Edge-Deletion Problems.  </paper_title>
<paper_abstract>   A graph property is monotone if it is closed under removal of vertices and edges. In this paper we consider the following algorithmic problem, called the edge-deletion problem; given a monotone property P and a graph G, compute the smallest number of edge deletions that are needed in order to turn G into a graph satisfying P. We denote this quantity by E ′ P (G). The first result of this paper states that the edge-deletion problem can be efficiently approximated for any monotone property. • For any fixed ɛ&amp;gt; 0 and any monotone property P, there is a deterministic algorithm, which given a graph G = (V, E) of size n, approximates E ′ P (G) in linear time O(|V | + |E|) to within an additive error of ɛn2. Given the above, a natural question is for which monotone properties one can obtain better additive approximations of E ′ P. Our second main result essentially resolves this problem by giving a precise characterization of the monotone graph properties for which such approximations exist. (1) If there is a bipartite graph that does not satisfy P, then there is a δ&amp;gt; 0 for which it is  </paper_abstract>
<query_num> 1901 </query_num>
<text>   . Many natural properties such as being Perfect, Chordal and Interval are hereditary non-monotone properties. By combining the ideas we used in order to prove Theorem 1.1 along with the main ideas of =-=[6]-=- it can be shown that Theorem 1.1 (as well as Theorem 1.2 and Corollary 1.2) also hold for any hereditary graph property. 3s1.3 On the possibility of better approximations Theorem 1.1 implies that it   </text>
<query_num> 1902 </query_num>
<text>   2 ) edges or the 3-CNF formula has Ω(n 3 ) clauses. Approximations for dense instances of Quadratic Assignment Problems, as well as for additional problems, were obtained by Arora, Frieze and Kaplan =-=[10]-=-. Frieze and Kannan [26] obtained approximations schemes for several dense graph theoretic problems via certain matrix approximations. Alon, Fernandez de la Vega, Kannan and Karpinski [3] obtained res   </text>
<query_num> 1903 </query_num>
<text>   In this section we discuss the basic notions of regularity, some of the basic applications of regular partitions and state the regularity lemmas that we use in the proof of Theorems 1.1 and 1.2. See =-=[35]-=- for a comprehensive survey on the regularity-lemma. We start with some basic definitions. For every two nonempty disjoint vertex sets A and B of a graph G, we define e(A, B) to be the number of edges  be, because we need the graph to be ”closer” to a random graph. This is formulated in the lemma below. Several versions of this lemma were previously proved in papers using the regularity lemma (see =-=[35]-=-). Lemma 2.3 For every real 0 &amp;lt; η &amp;lt; 1, and integers k, f ≥ 1 there exist γ = γ2.3(η, k, f), and N = N2.3(η, k, f) with the following property. Let F be any graph on f vertices, and let U1, . . . , Uk   </text>
<query_num> 1904 </query_num>
<text>   c version of Szemerédi’s Regularity Lemma. They used it to prove that Theorem 1.1 holds for the k-colorability property. The running time of their algorithm was improved by Kohayakawa, Rödl and Thoma =-=[34]-=-. Frieze and Kannan [25] further used the algorithmic version of the regularity lemma, to obtain approximation schemes for additional graph problems. Theorem 1.1 is obtained via the algorithmic versio ] were the first to obtain a polynomial time algorithm for finding the equipartition, whose existence is guaranteed by lemma 2.6. The running time of this algorithm was improved by Kohayakawa et. al. =-=[34]-=- who obtained the following result. Lemma 2.7 ([34]) For every fixed m and γ there is an O(n 2 ) time algorithm that given an equipartition A finds equipartition B as in Lemma 2.6. Our main tool in th   </text>
<query_num> 1905 </query_num>
<text>   eorem 1.3 achieves such a result even for the seemingly easier problem of approximating EP. 1.4.3 Approximation schemes for ”dense” instances Fernandez de la Vega [22] and Arora, Karger and Karpinski =-=[11]-=- showed that many of the classical NP -complete problems such as MAX-CUT and MAX-3-CNF have a PTAS when the instance is dense, namely if the graph has Ω(n 2 ) edges or the 3-CNF formula has Ω(n 3 ) cl   </text>
<query_num> 1906 </query_num>
<text>   formula has Ω(n 3 ) clauses. Approximations for dense instances of Quadratic Assignment Problems, as well as for additional problems, were obtained by Arora, Frieze and Kaplan [10]. Frieze and Kannan =-=[26]-=- obtained approximations schemes for several dense graph theoretic problems via certain matrix approximations. Alon, Fernandez de la Vega, Kannan and Karpinski [3] obtained results analogous to ours f   </text>
<query_num> 1907 </query_num>
<text>   hs G that satisfy EP(G) &amp;lt; δ from those that satisfy EP(G) &amp;gt; ɛ, where 0 ≤ δ &amp;lt; ɛ ≤ 1 are some constants. Recently, there have been several results in this line of work. Specifically, Fischer and Newman =-=[24]-=- have recently shown that if a graph property is testable with number of queries depending on ɛ only, then it is also tolerantly testable for any 0 ≤ δ &amp;lt; ɛ ≤ 1 and with query complexity depending on | In fact, the algorithm implied by Corollary 1.2 is the ”natural” one, where one picks a random subset of vertices S, and approximates EP(G) by computing EP on the graph induced by S. The algorithm of =-=[24]-=- is far more complicated. Furthermore, due to the nature of our algorithm if the input graph satisfies a monotone property P, namely if EP(G) = 0, we will always detect that this is the case. The algo interesting to decide if one can obtain a result analogous to Theorem 1.3 for the family of hereditary properties. • A weaker version of Theorem 1.1 can be derived by combining the results of [7] and =-=[24]-=-. However, this only enables one to approximate EP(G) within an additive error ɛ in time n f(ɛ) , while the running time of our algorithm is of type f(ɛ)n 2 . • Recall that E ′ F (G) denotes the small   </text>
<query_num> 1908 </query_num>
<text>   ie at the core of the proof of Theorem 1.1. This new technique, which may very well have other algorithmic and graph-theoretic applications, applies a result of Alon, Fischer, Krivelevich and Szegedy =-=[4]-=- which is a strengthening of Szemerédi’s Regularity Lemma [44]. We then use an efficient algorithmic version of the regularity lemma, which also implies an efficient algorithmic version of the result  rity lemma, to obtain approximation schemes for additional graph problems. Theorem 1.1 is obtained via the algorithmic version of a strengthening of the standard regularity lemma, which was proved in =-=[4]-=-, and it seems that these results cannot be obtained using the standard regularity lemma. 5s1.4.5 Tolerant Property-Testing In standard Property-Testing (see [23] and [41]) one wants to distinguish be For every fixed m and γ there is an O(n 2 ) time algorithm that given an equipartition A finds equipartition B as in Lemma 2.6. Our main tool in the proof of Theorem 1.1 is Lemma 2.9 below, proved in =-=[4]-=-. This lemma can be considered a strengthening of Lemma 2.6, as it guarantees the existence of an equipartition and a refinement of this equipartition that poses stronger properties compared to those  t in Definition 2.8 we may use an arbitrary function rather than a fixed γ as in Definition 2.5 (such functions will be denoted by E throughout the paper). The following is one of the main results of =-=[4]-=-. Lemma 2.9 ([4]) For any integer m and function E(r) : N ↦→ (0, 1) there is S = S2.9(m, E) such that any graph on at least S vertices has an E-regular equipartition A, B where |A| = k ≥ m and |B| = k ith m = |Ai−1| and γ = E(M)/M 2 and let Ai be the refinement of Ai−1 returned by Lemma 2.6. If Ai−1 and Ai form an E-regular equipartition stop, otherwise set M = |Ai−1| and reiterate. It is shown is =-=[4]-=- that after at most 100/ζ 4 iterations, for some 1 ≤ i ≤ 100/ζ 4 the partitions Ai−1 and Ai form an E-regular equipartition. Moreover, detecting an i for which this holds is very easy, that is, can be   </text>
<query_num> 1909 </query_num>
<text>   lves various combinatorial tools. These include Szemerédi’s Regularity Lemma, and a new result in Extremal Graph Theory, which is stated in Theorem 6.1 (see Section 6) that extends the main result of =-=[14]-=- and [15]. We also use the basic approach of [1], which applies spectral techniques to obtain an NP -hardness result by embedding a blow-up of a sparse instance to a problem, in an appropriate dense p  ′ r(Kn) − o(n2 ) ≤ E ′ H (Kn) ≤ E ′ r(Kn). The main extremal graph-theoretic tool that we use in order to obtain Theorem 1.3 is the following result, which greatly extends one of the main results of =-=[14]-=-. Note, that this result also extends Turán’s 18sTheorem and the Erdős-Stone-Simonovits Theorem as it states that E ′ H (G) and E′ r(G) are very close not only when G is Kn but already when G has a su E) is a graph on n vertices of minimum degree at least (1 − µ)n, then E ′ r(G) − O(n 2−γ ) ≤ E ′ H(G) ≤ E ′ r(G). The assertion of this theorem for the special case of H being a triangle is proved in =-=[14]-=- and in a stronger form in [15]. We note that the n2−γ term in the second item of the theorem cannot be avoided. Note, that the error term we obtain in the second part of the theorem is better than th   </text>
<query_num> 1910 </query_num>
<text>   n be approximated to within any multiplicative constant 1 + ɛ. 1 We assume henceforth that P is not satisfied by all graphs. 4s1.4.2 Hardness of edge-modification problems Natanzon, Shamir and Sharan =-=[39]-=- proved that for various hereditary properties, such as being Perfect and Comparability, computing EP is NP -hard and sometimes even NP -hard to approximate to within some constant. Yannakakis [46] ha   </text>
<query_num> 1911 </query_num>
<text>   ng a false negative, while deleting an edge means correcting a false positive. Computing EP(G) for appropriately defined properties P have important applications in physical mapping of DNA (see [17], =-=[29]-=- and [31]). Computing EP(G) for other properties arises when optimizing the running time of performing Gaussian elimination on a sparse symmetric positive-definite matrix (see [42]). Other modificatio   </text>
<query_num> 1912 </query_num>
<text>   of Theorem 6.1, which is an extension of Turán’s theorem. To this end, we need a result proved for Kr+1-free graphs by Andrásfai, Erdős and Sós [9] and in a more general form by Erdős and Simonovits =-=[21]-=-. Theorem 7.1 ([9],[21]) Let H be a fixed graph with chromatic number r + 1 ≥ 3 which contains an edge e such that χ(H − e) = r. If G is an H-free graph of order n with minimal degree δ(G) &amp;gt; 3r−4 3r−1   </text>
<query_num> 1913 </query_num>
<text>   s testable with number of queries depending on ɛ only, then it is also tolerantly testable for any 0 ≤ δ &amp;lt; ɛ ≤ 1 and with query complexity depending on |ɛ − δ|. Combining this with the main result of =-=[7]-=- implies that any monotone property is tolerantly testable for any 0 ≤ δ &amp;lt; ɛ ≤ 1 and with query complexity depending on |ɛ − δ|. Note, that Corollary 1.2 implicitly states the same. In fact, the algor t seems interesting to decide if one can obtain a result analogous to Theorem 1.3 for the family of hereditary properties. • A weaker version of Theorem 1.1 can be derived by combining the results of =-=[7]-=- and [24]. However, this only enables one to approximate EP(G) within an additive error ɛ in time n f(ɛ) , while the running time of our algorithm is of type f(ɛ)n 2 . • Recall that E ′ F (G) denotes   </text>
<query_num> 1914 </query_num>
<text>   shown that for any hereditary property, which is expressible by a finite number of forbidden induced subgraphs, the problem of computing the edit distance is fixedparameter tractable. Khot and Raman =-=[33]-=- proved that for some hereditary properties P, finding in a given graph G, a subgraph that satisfies P is fixed-parameter tractable, while for other properties finding such a subgraph is hard in an ap   </text>
</top>
<top>
<paper_num> 20 </paper_num>
<paper_title>   On the Validity of IEEE 802.11 MAC Modeling Hypotheses.  </paper_title>
<paper_abstract>   Abstract — We identify common hypotheses on which a large number of distinct mathematical models of WLANs employing IEEE 802.11 are founded. Using data from an experimental test bed and packet-level ns-2 simulations, we investigate the veracity of these hypotheses. We demonstrate that several of these assumptions are inaccurate and/or inappropriate. We consider hypotheses used in the modeling of: saturated and unsaturated 802.11 infrastructure mode networks; saturated 802.11e networks; and saturated and unsaturated 802.11s mesh networks. In infrastructure mode networks we find that, even for small numbers of stations, common hypotheses hold true for saturated stations and also for unsaturated stations with small buffers. However, despite their widespread adoption, common assumptions used to incorporate station buffers are erroneous. This raises questions about the predictive power of all models based on these hypotheses. For saturated 802.11e models that treat differences in AIFS, we find that one of the two fundamental hypotheses is accurate. The other is reasonable for small differences in AIFS, but unreasonable for large differences. For 802.11s mesh networks, we find that assumptions are appropriate only if stations are lightly loaded and are highly inappropriate if they are saturated. By identifying these flawed suppositions, this work identifies areas where mathematical models need to be revisited and revised if they are to be used with confidence by protocol designers and WLAN network planners.  </paper_abstract>
<query_num> 2001 </query_num>
<text>   0 if it results in a success. For a station in an 802.11 network employing DCF, irrespective of whether it is saturated (always having packets to send) or not, many authors (e.g. [2][3][4][5][6][7][8]=-=[9]-=-[10]) assume that: (A1) The sequence of outcomes of attempted transmissions, {Ck}, forms a stochastically independent sequence. (A2) The sequence {Ck} consists of identically distributed random variab redictions, and its predictive accuracy, Bianchi’s basic paradigm has been widely adopted for models that expand on its original range of applicability. A selection of these extensions include: [6][7]=-=[9]-=-[10], which consider the impact of unsaturated stations in the absence of station buffers and enable predictions in the presence of load asymmetries; [11][12][13][14], which treat unsaturated stations y converges to zero indicating little dependence in the the success per attempt sequence, even for N = 2. The (A1) and (A2) assumptions are also adopted in unsaturated models with small buffers [6][7]=-=[9]-=-[10] and big buffers [11][12][13][14]. From experimental data for unsaturated net-4 AutoCovariance Coefficient AutoCovariance Coefficient 0.1 0.08 0.06 0.04 0.02 0 Unsaturated, Small Buffer N=2 λ=400   </text>
<query_num> 2002 </query_num>
<text>   buffers, in addition define Qk := 1 if there is at least one packet awaiting processing after the kth successful transmission and Qk := 0 if the buffer is empty. Then it is commonly assumed (e.g. [11]=-=[12]-=-[13][14]) that: (A3) The sequence {Qk} consists of independent random variables. (A4) The sequence {Qk} consists of identically distributed random variables that, in particular, do not depend on backo lection of these extensions include: [6][7][9][10], which consider the impact of unsaturated stations in the absence of station buffers and enable predictions in the presence of load asymmetries; [11]=-=[12]-=-[13][14], which treat unsaturated stations in the presence of stations with buffers; [15][16][17], which investigate the impact of the variable parameters of 802.11e, including AIFS, on saturated netw ng little dependence in the the success per attempt sequence, even for N = 2. The (A1) and (A2) assumptions are also adopted in unsaturated models with small buffers [6][7][9][10] and big buffers [11]=-=[12]-=-[13][14]. From experimental data for unsaturated net-4 AutoCovariance Coefficient AutoCovariance Coefficient 0.1 0.08 0.06 0.04 0.02 0 Unsaturated, Small Buffer N=2 λ=400 N=5 λ=160 N=10 λ=80 −0.02 0  consider the additional hypotheses introduced to model buffering. V. ASSUMPTIONS (A3) AND (A4) To model stations with buffers serving Poisson traffic, the common idea across various authors, e.g. [11]=-=[12]-=-[13][14], is to treat each station as a queueing system where the service time distribution is identified with the MAC delay distribution based on a Bianchi-like model. The assumptions (A1) and (A2) a   </text>
<query_num> 2003 </query_num>
<text>   f it results in a success. For a station in an 802.11 network employing DCF, irrespective of whether it is saturated (always having packets to send) or not, many authors (e.g. [2][3][4][5][6][7][8][9]=-=[10]-=-) assume that: (A1) The sequence of outcomes of attempted transmissions, {Ck}, forms a stochastically independent sequence. (A2) The sequence {Ck} consists of identically distributed random variables  ictions, and its predictive accuracy, Bianchi’s basic paradigm has been widely adopted for models that expand on its original range of applicability. A selection of these extensions include: [6][7][9]=-=[10]-=-, which consider the impact of unsaturated stations in the absence of station buffers and enable predictions in the presence of load asymmetries; [11][12][13][14], which treat unsaturated stations in  onverges to zero indicating little dependence in the the success per attempt sequence, even for N = 2. The (A1) and (A2) assumptions are also adopted in unsaturated models with small buffers [6][7][9]=-=[10]-=- and big buffers [11][12][13][14]. From experimental data for unsaturated net-4 AutoCovariance Coefficient AutoCovariance Coefficient 0.1 0.08 0.06 0.04 0.02 0 Unsaturated, Small Buffer N=2 λ=400 N=5 er-frame space (AIFS). To model the first three of these, no additional modeling assumptions are necessary beyond (A1)-(A2) for saturated stations or for unsaturated stations with small buffers (e.g. =-=[10]-=-). For unsaturated stations with large buffers, no additional assumptions are necessary beyond (A1)-(A4) (e.g. [14]). However, to capture the full power of 802.11e’s service differentiation, one must   </text>
<query_num> 2004 </query_num>
<text>   ituations where (A0) is false, such as relay topologies that do not have multiple radios and so cannot mitigate interference at non-communicating distances, new approximations are necessary (e.g. [29]=-=[30]-=-[31][32]). Some of these models also use mean-field ideas, usually inspired P(D&amp;gt;t) 10 0 10 −1 10 −2 10 −3 10 −4 10 −5 Unsaturated, Big Buffer, N=5 !=100 Experimental Data Theoretical Data 0 2 4 6 8 10   </text>
<query_num> 2005 </query_num>
<text>   lower AIFS value can decrement their counters while stations with the higher AIFS observe the medium as being continuously busy for longer. The commonly adopted assumptions (e.g. [15][16][17][18][19]=-=[20]-=-) are: (A5) The sequence {Hk} consists of independent random variables. (A6) Each element of the sequence {Hk} is identically distributed, with a distribution that can be identified with one derived i h treat unsaturated stations in the presence of stations with buffers; [15][16][17], which investigate the impact of the variable parameters of 802.11e, including AIFS, on saturated networks; [18][19]=-=[20]-=- which treat unsaturated 802.11e networks; [21], which extends the paradigm from single hop networks to multiple-radio 802.11s mesh networks. Note that the work cited here is a small, selective sub-co y beyond (A1)-(A4) (e.g. [14]). However, to capture the full power of 802.11e’s service differentiation, one must model AIFS and this requires additional innovation and hypotheses [15][16][17][18][19]=-=[20]-=-. Consider N1 + N2 stations, each of which is serving one of two traffic classes with distinct AIFS values: N1 class one stations with AIFS1 and N2 class 2 stations with AIFS2 = AIFS1 + Dσ, where σ is   </text>
<query_num> 2006 </query_num>
<text>   n the time at which the kth successful transmission and the k−1 th successful transmission occurs from a tagged station. If the station’s arrival process is Poisson, then one pair of hypotheses (e.g. =-=[21]-=-) used to enable a tractable mathematical model of 802.11s mesh networks is:1 Assumption Saturated Small Buffer Big Buffer (A1) {Ck} independent ̌ (pairwise) ̌ (pairwise) ̌ (pairwise) (A2) {Ck} ident  stations with buffers; [15][16][17], which investigate the impact of the variable parameters of 802.11e, including AIFS, on saturated networks; [18][19][20] which treat unsaturated 802.11e networks; =-=[21]-=-, which extends the paradigm from single hop networks to multiple-radio 802.11s mesh networks. Note that the work cited here is a small, selective sub-collection within a vast body of literature. To a ent of the network. That is Dk is the difference between the time at which the k th successful transmission and the k − 1 th successful transmission occurs from a tagged station. One hypothesis (e.g. =-=[21]-=-) is that if the arrivals process to the station is Poisson, then the departure process is also Poisson. That is: (A7) {Dk} is a stochastically independent sequence; and (A8) the elements of {Dk} are   </text>
<query_num> 2007 </query_num>
<text>   nd Ck := 0 if it results in a success. For a station in an 802.11 network employing DCF, irrespective of whether it is saturated (always having packets to send) or not, many authors (e.g. [2][3][4][5]=-=[6]-=-[7][8][9][10]) assume that: (A1) The sequence of outcomes of attempted transmissions, {Ck}, forms a stochastically independent sequence. (A2) The sequence {Ck} consists of identically distributed rand make predictions, and its predictive accuracy, Bianchi’s basic paradigm has been widely adopted for models that expand on its original range of applicability. A selection of these extensions include: =-=[6]-=-[7][9][10], which consider the impact of unsaturated stations in the absence of station buffers and enable predictions in the presence of load asymmetries; [11][12][13][14], which treat unsaturated st quickly converges to zero indicating little dependence in the the success per attempt sequence, even for N = 2. The (A1) and (A2) assumptions are also adopted in unsaturated models with small buffers =-=[6]-=-[7][9][10] and big buffers [11][12][13][14]. From experimental data for unsaturated net-4 AutoCovariance Coefficient AutoCovariance Coefficient 0.1 0.08 0.06 0.04 0.02 0 Unsaturated, Small Buffer N=2   </text>
<query_num> 2008 </query_num>
<text>   ons see the medium as being idle. Consider a network of homogeneous saturated class 1 stations and homogeneous class 2 stations. To model the impact of different AIFS values, using the terminology in =-=[15]-=- we have the notion of hold states for class 2 stations. A class 2 station is in a hold state if class 1 stations can decrement their counters while it cannot. As all class 2 stations have the same AI   </text>
<query_num> 2009 </query_num>
<text>   r devices are free to operate in this range, these devices lead to interference. There are extensions to the WLAN modeling paradigms that include failed transmissions due to noise on the medium, e.g. =-=[33]-=-. This approach assumes that packet losses due to noise are i.i.d. and independent of all other stochastic processes in the model. Whether this assumption is appropriate is dependent on the particular   </text>
<query_num> 2010 </query_num>
<text>   s where (A0) is false, such as relay topologies that do not have multiple radios and so cannot mitigate interference at non-communicating distances, new approximations are necessary (e.g. [29][30][31]=-=[32]-=-). Some of these models also use mean-field ideas, usually inspired P(D&amp;gt;t) 10 0 10 −1 10 −2 10 −3 10 −4 10 −5 Unsaturated, Big Buffer, N=5 !=100 Experimental Data Theoretical Data 0 2 4 6 8 10 12 14 x   </text>
<query_num> 2011 </query_num>
<text>   sion and Ck := 0 if it results in a success. For a station in an 802.11 network employing DCF, irrespective of whether it is saturated (always having packets to send) or not, many authors (e.g. [2][3]=-=[4]-=-[5][6][7][8][9][10]) assume that: (A1) The sequence of outcomes of attempted transmissions, {Ck}, forms a stochastically independent sequence. (A2) The sequence {Ck} consists of identically distribute ransmission matches that in the real system, which is an input to the model. If this average is known, this methodology has been demonstrated to make accurate throughput and average delay predictions =-=[4]-=-[5]. In the mean-field approach, the fundamental idea is similar, but the calculation of τ, and thus p, does not require external inputs. One starts by assuming that p is given and each station always accurate in networks with a large number of stations. The p-persistent paradigm has been developed to encompass, for example, saturated 802.11 networks where every station always has a packet to send =-=[4]-=- and saturated 802.11e networks [8]. However, due to its intuitive appeal, its self contained ability to make predictions, and its predictive accuracy, Bianchi’s basic paradigm has been widely adopted te that the work cited here is a small, selective sub-collection within a vast body of literature. To appreciate just how large this literature is, as of November 2009, the ppersistent modeling paper =-=[4]-=- has been cited over 400 times according to ISI Knowledge and over 800 times according to Google Scholar, while the mean-field modeling paper [3] been cited over 1300 times according to ISI Knowledge  nt. IV. ASSUMPTIONS (A1) AND (A2) For a single station, define Ck := 1 if the k th transmission attempt results in a collision and Ck := 0 if it results in a success. The two key assumptions in [2][3]=-=[4]-=-[5] are effectively these: (A1) the sequence {Ck} consists of independent random variables; and (A2) the sequence {Ck} consists of identically distributed random variables. That is, there exists a fix   </text>
<query_num> 2012 </query_num>
<text>   th a lower AIFS value can decrement their counters while stations with the higher AIFS observe the medium as being continuously busy for longer. The commonly adopted assumptions (e.g. [15][16][17][18]=-=[19]-=-[20]) are: (A5) The sequence {Hk} consists of independent random variables. (A6) Each element of the sequence {Hk} is identically distributed, with a distribution that can be identified with one deriv which treat unsaturated stations in the presence of stations with buffers; [15][16][17], which investigate the impact of the variable parameters of 802.11e, including AIFS, on saturated networks; [18]=-=[19]-=-[20] which treat unsaturated 802.11e networks; [21], which extends the paradigm from single hop networks to multiple-radio 802.11s mesh networks. Note that the work cited here is a small, selective su ssary beyond (A1)-(A4) (e.g. [14]). However, to capture the full power of 802.11e’s service differentiation, one must model AIFS and this requires additional innovation and hypotheses [15][16][17][18]=-=[19]-=-[20]. Consider N1 + N2 stations, each of which is serving one of two traffic classes with distinct AIFS values: N1 class one stations with AIFS1 and N2 class 2 stations with AIFS2 = AIFS1 + Dσ, where   </text>
<query_num> 2013 </query_num>
<text>   tigation in the present article due to space constraints. This validation of the standard decoupling assumption, (A1) and (A2), for saturated networks helps to explain why the predictions in [2][3][4]=-=[5]-=- are so precise. Even though intuitively one expects the main model assumptions to be valid for large networks, in fact they are accurate even for small networks. As the assumptions are reasonable, de   </text>
<query_num> 2014 </query_num>
<text>   ty of these assumptions directly, but infer them from the accuracy of model predictions of coarse grained quantities 1 A preliminary report on this work appeared in the Proceedings of IEEE PIMRC 2008 =-=[1]-=-. The hypotheses of 802.11e, 802.11s networks, and ppersistent protocols were not identified and investigated in that article and simulation results were used exclusively rather than experimental meas   </text>
<query_num> 2015 </query_num>
<text>   ural alternative to (A3) and (A4) is to use the approximation that: (A3’) given βk = i, {Qk} is an independent sequence; and (A4’) given βk = i, P (Qk = 1) = qi. As a first step in this direction, in =-=[34]-=- negative consequences of adopting the assumptions (A3) and (A4) are identified. A typical validation scenario employed by modelers is to consider a symmetrically loaded network. While this is unlikel actice, mathematically it leads to homogeneous fixed point equations whose solution can be quickly identified by standard numerical techniques. For stations that are asymmetrically loaded, results in =-=[34]-=- demonstrate that a model based on these assumptions provides inaccurate throughput predictions. That this is a consequence of (A3) and (A4) is established by considering the setting where all station   </text>
</top>
<top>
<paper_num> 21 </paper_num>
<paper_title>   Learning intersections of halfspaces with a margin.  </paper_title>
<paper_abstract>   We give a new algorithm for learning intersections of halfspaces with a margin, i.e. under the assumption that no example lies too close to any separating hyperplane. Our algorithm combines random projection techniques for dimensionality reduction, polynomial threshold function constructions, and kernel methods. The algorithm is fast and simple. It learns a broader class of functions and achieves an exponential runtime improvement compared with previous work on learning intersections of halfspaces with a margin.  </paper_abstract>
<query_num> 2101 </query_num>
<text>   X: The degree of a polynomial threshold function p is simply the degree of the polynomial p. Polynomial threshold functions are well studied in the case where X = f0; 1gn or f\Gamma 1; 1gn (see e.g. =-=[5, 16, 18, 20]-=-) but we will consider other more general subsets X: For S ` fx1; : : : ; xng a multiset of variables, we write xS to denote the monomial Q i2S xi. For p(x) = P S cSxS a polynomial, we write kpk to de   </text>
<query_num> 2102 </query_num>
<text>   esearchers have considered the problem of learning intersections of halfspaces. Efficient algorithms are known for learning intersections of halfspaces under the uniform distribution on the unit ball =-=[7, 22]-=- and on the Boolean cube [15], but less is known about learning under more general probability distributions. Baum [4] gave an algorithm which learns an intersection of two origin-centered halfspaces   </text>
<query_num> 2103 </query_num>
<text>   f the halfspace in R( n+d d ) one must analyze the PTF margin of p rather than its geometric margin. 12s7.2 Lower Bounds on Polynomial Threshold Functions The main result of O&amp;apos;Donnell and Servedio in =-=[19]-=-, if suitably interpreted, proves that there exists a set X ae R2 labelled according to the intersection of two halfspaces with margin ae for which any PTF correctly classifying X must have degree \Om   </text>
<query_num> 2104 </query_num>
<text>   improvement compared with previous work on learning intersections of halfspaces with a margin. \Lambda Supported by an NSF Mathematical Sciences Postdoctoral Research Fellowship. 1sArriaga &amp; Vempala =-=[3]-=- This Paper h1 ^ \Deltas\Deltas\Deltas^ ht n \Deltaspoly i log t ae j + i log t ae j t log t ae ae2 n i t ae jt log t log 1 ae or n i log t ae j q 1 ae log t f (h1; : : : ; ht) -------------- n i t ae to a data set if each of the defining halfspaces has margin ae on the data set; we give precise definitions later.) The margin is a natural parameter to consider; previous work by Arriaga and Vempala =-=[3]-=- on learning intersections of halfspaces has explicitly studied the dependence on this parameter. Since the Perceptron algorithm learns a single halfspace in time O(1=ae2), the ultimate goal in this f n this is the most interesting case) our learning algorithm runs in (1=ae)O(log 1=ae) time, i.e. quasipolynomial in 1=ae: This is an exponential improvement over Arriaga and Vempala&amp;apos;s previous result =-=[3]-=- which was an algorithm that runs in (1=ae)!(1=ae 2) time. (Put another way, our algorithm can learn the intersection of O(1) halfspaces with margin at least 1=2 p log n in poly(n) time, whereas Arria stribution D: The values of m; k and d are given in Section 6. t; this algorithm can learn an intersection of t = n1= log log n many halfspaces in poly(n) time. In contrast, the previous algorithm of =-=[3]-=- has a t!(t) dependence on t and thus runs in poly(n) time only for t = o( log n log log n ) many halfspaces. As described below all our results are achieved using simple iterative algorithms (in fact t a random projection of a set of m points in Rn into Rk (with k ss O( log m ffl2 )) with high probability will not change pairwise distances by more than a (1 \Sigmasffl) factor. Arriaga and Vempala =-=[3]-=- were the first to give learning algorithms based on random projections. Their key insight was that since the geometry of a sample does not change much under random projection, one can run learning al   </text>
<query_num> 2105 </query_num>
<text>   ltasOE(y): The use of kernels in machine learning has received much research attention in recent years (see e.g. [10, 12] and references therein). Given a data set X ae Rn; it is well known (see e.g. =-=[11]-=-) that the Perceptron algorithm can be simulated over OE(X) in the expanded feature space RN using the kernel function K(x; y) to 5syield an implicit representation of a halfspace in RN : If evaluatin   </text>
<query_num> 2106 </query_num>
<text>   lynomials, i.e. Q(x) = a(x)=b(x). The degree of Q is defined as deg(a) + deg(b): Building on earlier results of Newman [17] on rational functions which approximate the absolute value function jxj, in =-=[6]-=- Beigel et al. gave a construction of a low-degree rational function which closely approximates the function sgn(x). We will use the following lemma (Lemma 9 of [6]): Lemma 7 [6] For all integers r; `   </text>
<query_num> 2107 </query_num>
<text>   on of a halfspace in RN : If evaluating K(x; y) takes time T and the Perceptron algorithm is simulated until M mistakes are made on a data set X with jXj = m; the time required is O(mT M 2) (see e.g. =-=[12, 14]-=-). 3 Random Projections We say that an n \Thetask matrix M is a random projection matrix if each entry of M is chosen independently and uniformly from f\Gamma 1; 1g: We will use the following lemma fr   </text>
<query_num> 2108 </query_num>
<text>   perplanes for halfspaces h1; : : : ; ht where each kwik = 1: Let P be the univariate polynomial P (x) = Tr(1 \Gammasx) where r = d p 2=aee: The first part of Lemma 10 implies that jP (x)j ^ 1 for x 2 =-=[0; 2]-=-, and the second part implies that P (x) * 2 for x ^ \Gamma ae 2 : Now consider the polynomial threshold function sign(p(x)) where p(x) = t + 1 2 \GammastX i=1 (P (w i \Deltasx)) dlog 2te : Since P is  Rk via M and then evaluating the k-variable PTF p, it suffices to show that p is a good hypothesis under the distribution M (D) obtained by projecting D down to Rk via M . It is well known (see e.g. =-=[2]-=-) that the VC dimension of the class of degree-d PTFs over k real variables is \Gamma k+d d \Deltas. Thus by the VC theorem [8] in order to learn to accuracy ffl and confidence ffi it suffices to take   </text>
<query_num> 2109 </query_num>
<text>   roblem of learning intersections of halfspaces. Efficient algorithms are known for learning intersections of halfspaces under the uniform distribution on the unit ball [7, 22] and on the Boolean cube =-=[15]-=-, but less is known about learning under more general probability distributions. Baum [4] gave an algorithm which learns an intersection of two origin-centered halfspaces under any symmetric distribut ) 2 [\Gamma 1 \Gammas1 r ; \Gamma 1] for all x 2 [\Gamma 2 `; \Gamma 1]; and (iii) Each coefficient of a(x); b(x) has magnitude at most 2O(` 2 log r) . The following theorem generalizes Theorem 24 in =-=[15]-=-, which addresses the special case of intersections of low-weight halfspaces over the space X = f0; 1gn: Theorem 8 Let X be a subset of Rk with 1 2 ^ kXk ^ 2 and c : R k ! f\Gamma 1; 1g be an intersec ) ). This lower bound implies that our choice of d in the proof of Theorem 15 is essentially optimal with respect to ae. For a discussion of other lower bounds on PTF constructions see Klivans et al. =-=[15]-=-. 7.3 Alternative Algorithms We note that after random projection, in Step 3 of Algorithm A there are several other algorithms that could be used instead of kernel Perceptron. For example, we could ru   </text>
</top>
<top>
<paper_num> 22 </paper_num>
<paper_title>   Anomalous loss performance for mixed real-time and TCP traffic in routers with very small buffers.  </paper_title>
<paper_abstract>   Abstract—In the past few years there has been vigorous debate regarding the size of buffers required at core Internet routers. Recent arguments supported by theory and experimentation show that under certain conditions, core router buffer sizes of a few tens of packets suffice for realizing acceptable end-to-end TCP throughputs. This is a significant step toward the realization of optical packet switched (OPS) networks, which are inherently limited in their ability to buffer optical signals. However, prior studies have largely ignored the presence of real-time traffic, which is increasing in importance as a source of revenue for Internet service providers. In this paper, we study the interaction that happens between realtime (open-loop) and TCP (closed-loop) traffic when they multiplex at buffers of very small size (few tens of packets) and make a significant discovery—namely that in a specific range of buffer size, real-time traffic losses increase as buffer size becomes larger. Our contributions pertaining to this anomalous behavior are threefold. First, we exhibit this anomalous loss performance for realtime traffic via extensive simulations using synthetic traffic and real video traces. Second, we develop quantitative models that reveal the dynamics of buffer sharing between real-time and TCP traffic that lead to this behavior. Third, we show how various factors such as the nature of real-time traffic, mixture of long-lived and short-lived TCP flows, and packet sizes impact the severity of the anomaly. Our study is the first to consider interactions between real-time and TCP traffic in very small (potentially all-optical) buffers and informs router manufacturers and network operators of the factors to consider when dimensioning such small buffer sizes for desired performance balance between real-time and TCP traffic. Index Terms—Anomalous loss performance, mixed TCP and real-time traffic, optical packet switched (OPS) networks, routers with very small buffers. I.  </paper_abstract>
<query_num> 2201 </query_num>
<text>   0 packets instead of the million packets required by the rule of thumb. Since 2004, several new arguments on buffer sizing have been put forth. Stanford University researchers further proposed in [4]–=-=[6]-=- that under certain conditions, which they believe hold in today’s Internet, as few as 20–50 packet buffers suffice for TCP traffic to realize acceptable link utilization, a claim supported by their e needed at that interface. If this ratio is greater than 1, then the loss rate falls exponentially, and only a very small amount of buffering is needed, which corroborates with the results reported in =-=[6]-=-. However, the concern is that if the output/input capacity ratio is lower than 1, then the loss rate follows a power-law reduction and significant buffering is needed. Researchers from the University ter buffer sizing, we refer the reader to our recent survey paper [16]. A. Motivation While most prior studies on buffer sizing have focused on electronic Internet routers, some earlier works such as =-=[6]-=- and [17] have applied buffer sizing principles to optical switches. However, they focus entirely on TCP traffic performance and ignore the performance implications for real-time traffic. From the obs   </text>
<query_num> 2202 </query_num>
<text>   a Gb/s link would require 10 Gb of buffering, which poses a considerable challenge to router design. This buffer sizing rule was first challenged in 2004 by researchers from Stanford University [2], =-=[3]-=- who showed that when a large number of long-lived TCP flows multiplex at a bottleneck link, synchronization does not arise, and near100% utilization of the bottleneck link can be achieved with only b   </text>
<query_num> 2203 </query_num>
<text>   al: large for TCP packets (1000 bytes) and small for UDP packets (say 200 bytes). For brevity, we do not explain the derivation of this model here and instead refer the interested reader to our paper =-=[39]-=- for the detailed analysis. It suffices to state here that the results from the M/D/1/B analysis are qualitatively similar to the M/M/1/B results presented, and both chains predict the inflection poin   </text>
<query_num> 2204 </query_num>
<text>   dent characteristics. It can be noted that the above traffic generation mechanism, which is a combination of several ON–OFF sources with Pareto-distributed ON periods, is in fact long-range-dependent =-=[44]-=-. Fig. 17 plots the empty buffer probability for the four different ratios of long-lived to short-lived flows. Our first observation is that the empty buffer probability falls fairly linearly (on log-   </text>
<query_num> 2205 </query_num>
<text>   ilization of all TCP flows. Buffer size at the bottleneck router is varied in terms of kilobytes. To set the packet sizes, we draw on the fact that several real-time applications, e.g., online gaming =-=[26]-=-, [27], use small UDP packets since they require extremely low latencies. The study showed that almost all packets were under 200 bytes. Our experiments using Skype and Yahoo! Messenger showed that fo   </text>
<query_num> 2206 </query_num>
<text>   ing is needed. Researchers from the University of Illinois at Urbana–Champaign also arrive at a similar conclusion in [11]. Other studies have considered factors such as application layer performance =-=[12]-=-, [13] and fairness [14] influencing buffer sizing. In late 2008, the Stanford group presented experimental results validating the applicability of routers with very small buffers in the core of the I   </text>
<query_num> 2207 </query_num>
<text>   ion of all TCP flows. Buffer size at the bottleneck router is varied in terms of kilobytes. To set the packet sizes, we draw on the fact that several real-time applications, e.g., online gaming [26], =-=[27]-=-, use small UDP packets since they require extremely low latencies. The study showed that almost all packets were under 200 bytes. Our experiments using Skype and Yahoo! Messenger showed that for inte   </text>
<query_num> 2208 </query_num>
<text>   lived to short-lived flows. The total number of TCP flows is kept constant at 2000. In order to incorporate realistic TCP traffic, we consider the closed-loop flow arrival model described in [10] and =-=[43]-=-, operating as follows. A given number of users (up to a maximum of 2000 in our example) perform successive file transfers to their respective destination nodes. The size of the file to be transferred   </text>
<query_num> 2209 </query_num>
<text>   nd fairness [14] influencing buffer sizing. In late 2008, the Stanford group presented experimental results validating the applicability of routers with very small buffers in the core of the Internet =-=[15]-=-. For a comprehensive IEEE Proof Web Version 1063-6692/$26.00 © 2010 IEEE2 IEEE/ACM TRANSACTIONS ON NETWORKING survey on this topic of router buffer sizing, we refer the reader to our recent survey p   </text>
<query_num> 2210 </query_num>
<text>   ofiled and are known to exhibit self-similar and long-range-dependent traffic characteristics. We first illustrate the phenomenon using the video traffic trace from the movie Star Wars, obtained from =-=[29]-=- and references therein. The mean rate is 374.4 kb/s, and the peak rate is 4.446 Mb/s, with the peak-rate-to-mean-rate ratio being nearly 12. The packet size is fixed at 200 bytes. We set the bottlene   </text>
<query_num> 2211 </query_num>
<text>   on, a claim supported by their experimental results at Sprint ATL and Verizon Communications [7]. A measurement study on a Sprint backbone router also found the queue size to seldom exceed 10 packets =-=[8]-=-, while the choice of 50 packet buffers is recommended in [9] to guarantee overall stability. These initial results show the feasibility of building all-optical networks that can be operated at 70%–80   </text>
<query_num> 2212 </query_num>
<text>   phenomenon also occurs under this scenario, we generated fBm traffic at the same average rate of 5 Mb/s. Other parameters are the same as before. The fBm model used is similar to our previous work in =-=[40]-=- and [41]. The traffic model combines a constant mean arrival rate with fractional Gaussian noise (fGn) characterized by zero mean, variance , and Hurst parameter . We use our filtering method in [42]   </text>
<query_num> 2213 </query_num>
<text>   rs has the potential to negatively impact quality of service and lead to worse performance, which could be a significant concern for the operator of the network. Finally, recent work, such as that in =-=[23]-=- and [24], has proposed the use of adaptive strategies wherein routers adapt their buffer size on the fly (depending upon the prevailing network conditions) so as to achieve a desired loss rate and li   </text>
<query_num> 2214 </query_num>
<text>   s from an all-optical router design point of view, where buffering presents a very important but difficult operation, since data is to be retained in the optical domain. Researchers from Georgia Tech =-=[10]-=- revisited the ongoing buffer sizing debate from the perspective of average per-flow TCP throughput rather than focusing purely on link utilization. The authors present evidence to suggest that the ou  of long-lived to short-lived flows. The total number of TCP flows is kept constant at 2000. In order to incorporate realistic TCP traffic, we consider the closed-loop flow arrival model described in =-=[10]-=- and [43], operating as follows. A given number of users (up to a maximum of 2000 in our example) perform successive file transfers to their respective destination nodes. The size of the file to be tr eter 1.5. These chosen values are representative of Internet traffic and comparable to measurement data. After each file transfer, the user transitions into an idle or OFF state, or as the authors of =-=[10]-=- suggest, a “thinking period.” The duration of the thinking period is exponentially distributed with mean 1 s. It is widely believed that Internet traffic exhibits self-similar and long-range-dependen   </text>
<query_num> 2215 </query_num>
<text>   with a Gb/s link would require 10 Gb of buffering, which poses a considerable challenge to router design. This buffer sizing rule was first challenged in 2004 by researchers from Stanford University =-=[2]-=-, [3] who showed that when a large number of long-lived TCP flows multiplex at a bottleneck link, synchronization does not arise, and near100% utilization of the bottleneck link can be achieved with o v model based on some simplifications. 1) Assumption: TCP Packet Arrivals Are Poisson: If a large number (potentially thousands) of long-lived TCP flows multiplex at a bottleneck link, it is believed =-=[2]-=- they do not synchronize their window dynamics behavior and can thus be treated as independent flows. Combined with the fact that each TCP flow’s window will be quite small (since bottleneck buffers a   </text>
</top>
<top>
<paper_num> 23 </paper_num>
<paper_title>   SELP - A System for Studying Strong Equivalence Between Logic Programs.  </paper_title>
<paper_abstract>   Abstract. This paper describes a system called SELP for studying strong equivalence in answer set logic programming. The basic function of the system is to check if two given ground disjunctive logic programs are equivalent, and if not, return a counter-example. This allows us to investigate some interesting properties of strong equivalence, such as a complete characterization for a rule to be strongly equivalent to another one, and checking whether a given set of rules is strongly equivalent to another, perhaps simpler set of rules. 1  </paper_abstract>
<query_num> 2301 </query_num>
<text>   ..., hk}, P sr the set {p1, ..., pm}, and Ngr the set {pm+1, ..., pn}. Thus a rule r can also be written as Hdr ← P sr, not Ngr. The semantics of these programs are given by answer sets as defined in =-=[3]-=-. As conventional in logic programming, we identify interpretations of L with sets of atoms in L. Let I be a set of atoms, and P a logic program. We say that I is closed under P if for any rule r in P   </text>
<query_num> 2302 </query_num>
<text>   P1 yet does not contain P1, i.e. P1 cannot be simplified using strong equivalence. As another example, consider P2 = {(a1 ← not a2), (a2 ← not a3), (a3 ← not a1)}. This is a program with an odd cycle =-=[8, 9]-=-, i.e. there is a simple cycle in the dependency graph of the program that has an odd number of negative edges. Odd cycles in a logic program act as constraints [9]. For instance, we have already seen   </text>
<query_num> 2303 </query_num>
<text>   another one, and checking whether a given set of rules is strongly equivalent to another, perhaps simpler set of rules. 1 Introduction The notion of strongly equivalent logic programs was proposed by =-=[5]-=-. It has been found useful for tasks such as program simplification (e.g. [2]). In this paper, we describe a system called SELP that can help us answer questions regarding this notion, from simple one this rule; otherwise, delete all the literals of the form not qi from this rule. Two logic programs P1 and P2 in L are said to be equivalent if they have the same answer sets, and strongly equivalent =-=[5]-=- (in the language L) if for any logic program P in L, P ∪P1 and P ∪P2 are equivalent. For instance, {p ← q} and {(p ← q), (q ← p)} are equivalent but not strongly equivalent: they both have the unique  them, the first one will have the answer set {p}, while the latter {p, q}. The notion of strongly equivalent logic programs is interesting for a variety of reasons. For instance, as Lifschitz et al. =-=[5]-=- noted, if two sets of rules are strongly equivalent, then one can be replaced by the other in any logic program regardless of the context. Thus knowing whether two sets of rules are strongly equivale ent, we mean that the two sets of rules, each consisting of exactly one of the rules, are strongly equivalent. 3 Checking strong equivalence between two logic programs Lifschitz, Pearce, and Valverde =-=[5]-=- showed that checking for strong equivalence between two logic programs can be done in the logic of here-and-there, a three-valued non-classical logic somewhere between classical logic and intuitionis  ← | p ∈ ML ′}. (2) If ML ′ is closed under P2, then P1 ∪ P and P2 ∪ P is not equivalent, where P = {p ← | p ∈ ML} ∪ {p ← q | p, q ∈ ML ′ \ ML, p �= q}. Proof. Follows from the proofs of Theorem 1 in =-=[5]-=- and Theorem 1 in [6]. For example, given P1 = {(a ← not b), (b ← not a)} and P2 = {a; b ←}, the answer of SELP will return the counter-example P = {(a ← b), (b ← a)}. However, SELP will confirm that   </text>
<query_num> 2304 </query_num>
<text>   ent in classical propositional logic. Thus, the problem of strong equivalence checking can be translated into a satisfiability problem in propositional logic, and solved using SAT solvers like zChaff =-=[10]-=-. In addition, when two programs are not strongly equivalent, we may want to find some witnesses. For example, P1 = {(a1 ← a2), (a1 ← not a2)} and P2 = {a1 ←} are not strongly equivalent, and P = {a2   P sr} ∪ {¬p ′ | p ∈ Ngr} ∪ {¬p | p ∈ Hdr}, {p ⊃ p ′ | p ∈ L} ∪ ∆(P1) ∪ {p ′ | p ∈ P sr} ∪ {¬p ′ | p ∈ Ngr ∪ Hdr}. Algorithm 1 makes precise this idea, and was implemented using the SAT solver zChaff =-=[10]-=-. If P1 and P2 are not strongly equivalent, then zChaff will return an assignment that is a counter-example to either (2) or (3), and from this assignment, we can construct a program P such that P ∪ P   </text>
<query_num> 2305 </query_num>
<text>   ing for strong equivalence between two logic programs can be done in the logic of here-and-there, a three-valued non-classical logic somewhere between classical logic and intuitionistic logic. Turner =-=[11]-=- provided a model-theoretic characterization of strong equivalence in terms of pairs of sets of atoms. Lin [6] provided a mapping from logic programs to propositional theories and showed that two logi   </text>
<query_num> 2306 </query_num>
<text>   ll return a program P , such that P1 ∪ P and P2 ∪ P have different answer sets. The most interesting part of SELP is that it allows us to study some properties of the notion of strong equivalence. In =-=[7]-=-, we described some results on classes of strongly equivalent logic programs discovered using the system. In this paper, we shall show how the system can help us answer questions of the following form use this way. Rather, we consider it a tool to systematically study the notion of strong equivalence. In the following, we discuss its use in discovering classes of strongly equivalent logic programs =-=[7]-=-, and in searching for simpler sets of rules that are strongly equivalent to a given one. 4 Discovering general theorems As we have mentioned, one possible use of strongly equivalent logic program is  le r is strongly equivalent to empty set if Hdr ∩ Ngr �= ∅. But is this the only case, i.e. is this both a sufficient and necessary condition for a rule to be strongly equivalent to the empty set? In =-=[7]-=-, we described a methodology and proved some general theorems for discovering theorems like this. More precisely, we were interested in the following so-called k-m-n theorem-discovery problem: Find so k that {r ′ , r} and {r ′ } are strongly equivalent.sSELP - A System for Studying Strong Equivalence between Logic Programs 135 We will not go into details of how we addressed the k-m-n problems (see =-=[7]-=-). The basic idea is to first find a condition that captures the strong equivalence between {r1, ..., rk, u1, ..., um} and {r1, ..., rk, v1, ..., vn} when all the rules are from a small language, say   ∪ Ngr3 and P sri \ {p} ⊆ P sr3 and Ngri \ {p} ⊆ Ngr3, where i = 1, 2 4.3 If p ∈ P sr1 ∩ Ngr2 , then Hdr1 ∩ Hdr3 = ∅ 4.4 If p ∈ P sr2 ∩ Ngr1 , then Hdr2 ∩ Hdr3 = ∅ Some general theorems are proved in =-=[7]-=- that help us verifying that Lemmas 1-3 in fact hold in the general case. Theorem 3. Lemma 1-3 hold in the general case, without any restriction on the number of atoms. An important consequence of thi   </text>
<query_num> 2307 </query_num>
<text>   lowing form: Given a set P of rules, is there another set of rules of certain property that is strongly equivalent to P ?sSELP - A System for Studying Strong Equivalence between Logic Programs 131 In =-=[4]-=-, a system called LPEQ was developed that can check if two normal programs are strongly equivalent, and was implemented using the answer set logic programming system smodels. Besides being implemented   </text>
<query_num> 2308 </query_num>
<text>   minimal set of atoms that is closed under P . Generally, I is an answer set of P iff I is an answer set of P I , where P I , the reduct of P on I, is obtained from P as follows: for any rule of form =-=(1)-=-, if there is an atom pi, m + 1 ≤ i ≤ n, such that pi ∈ I, then delete this rule; otherwise, delete all the literals of the form not qi from this rule. Two logic programs P1 and P2 in L are said to be in program simplification. In the following, for convenience, when we say a rule is strongly equivalent to the empty set, we mean the set that contains exactly this rule is strongly equivalent to the =-=(1)-=-s132 Yin Chen, Fangzhen Lin and Lei Li empty set. Similarly, when we say two rules are strongly equivalent, we mean that the two sets of rules, each consisting of exactly one of the rules, are strongl ⊃ p ′ |p ∈ L} ∪ ∆(P1) |= ∆(P2), (2) {p ⊃ p ′ |p ∈ L} ∪ ∆(P2) |= ∆(P1). (3) where for each p ∈ L, p ′ is a new atom, and for each program P , ∆(P ) = {∆(r) | r ∈ P }, where for each rule r of the form =-=(1)-=-, ∆(r) is the conjunction of the following two sentences: p1 ∧ · · · ∧ pm ∧ ¬p ′ m+1 ∧ · · · ∧ ¬p ′ n ⊃ h1 ∨ · · · ∨ hk, (4) p ′ 1 ∧ · · · ∧ p ′ m ∧ ¬p ′ m+1 ∧ · · · ∧ ¬p ′ n ⊃ h ′ 1 ∨ · · · ∨ h ′ k.  , M a model of {p ⊃ p ′ |p ∈ L} ∪ ∆(P1), and not ∆(P2). Let ML and ML ′ be the two sets of atoms defined as follows: Then we have ML = {p | p ∈ L and M |= p}, (6) ML ′ = {p | p ∈ L and M |= p′ }. (7) =-=(1)-=- If ML ′ is not closed under P2, then P1 ∪ P and P2 ∪ P is not equivalent, where P = {p ← | p ∈ ML ′}. (2) If ML ′ is closed under P2, then P1 ∪ P and P2 ∪ P is not equivalent, where P = {p ← | p ∈ ML   </text>
<query_num> 2309 </query_num>
<text>   to another, perhaps simpler set of rules. 1 Introduction The notion of strongly equivalent logic programs was proposed by [5]. It has been found useful for tasks such as program simplification (e.g. =-=[2]-=-). In this paper, we describe a system called SELP that can help us answer questions regarding this notion, from simple ones such as “are P and Q strongly equivalent” to more involved ones such as “ex   </text>
</top>
<top>
<paper_num> 24 </paper_num>
<paper_title>   Chosen-Ciphertext Secure Identity-Based Encryption in the Standard Model with short Ciphertexts.  </paper_title>
<paper_abstract>   We describe a practical identity-based encryption scheme that is secure in the standard model  against chosen-ciphertext (CCA2) attacks. Security is based on an assumption comparable to (but  slightly stronger than) Bilinear Decisonal Di#e-Hellman (BDDH). A comparison shows that our  construction outperforms all known identity-based encryption schemes in the standard model and  its performance is even comparable with the one from the random-oracle based Boneh/Franklin  IBE scheme. Our proposed IBE scheme has furthermore the property that it fulfills some notion of  &amp;quot;redundancy-freeness&amp;quot;, i.e. the encryption algorithm is not only a probabilistic injection but also a  surjection. As a consequence the ciphertext overhead is nearly optimal: to encrypt k bit messages  for k bit identities and with k bit randomness we get 3k bit ciphertexts to guarantee (roughly) k  bits of security.  </paper_abstract>
<query_num> 2401 </query_num>
<text>   3] mode of operation that avoid the overhead due to the MAC. Then 2sciphertexts of our IBE come with minimal overhead, i.e they are identity-preserving redundancy-free. Following Phan and Pointcheval =-=[34]-=- this property means that the IBE encryption algorithm (viewed as a mapping from randomness space, identity space, and message space into the ciphertext space) is a bijection. Consequently all possibl existence of identity-preserving redundancy-free IBE schemes in the standard model particularly remarkable since in the standard public-key encryption setting redundancy-free schemes (in the sense of =-=[34]-=-) are not known to exist. We further remark that the ciphertexts of our IBE scheme have the same message expansion as the most efficient standard public-key encryption schemes (like Kurosawa/Desmedt [ ur security definition we need a sufficiently large randomness space since otherwise the IBE scheme is not even indistinguishable against chosen-plaintext attacks [20]. Following Phan and Pointcheval =-=[34]-=- we say that an IBE scheme is redundancy-free if for any possible identity id the above encryption mapping IBEencid is a bijection, i.e. if all elements from the ciphertext space are “reachable”. This   </text>
<query_num> 2402 </query_num>
<text>   BDDHI (2 Biliear Decisional Diffie-Hellman Inversion) assumption was introduced by Boneh and Boyen [5] and its stronger variants (q-BDDHI for some polynomial q) already found numerous applications in =-=[5, 6, 7, 32]-=-. 1.2 Related Work and Comparison As we already pointed out, chosen-ciphertext secure IBE scheme were known to exist using generic reductions [12] based on Waters’ 2-level HIBE [40]. Recently the firs pporting hierarchical structures [24, 19]. By the relation to Waters IBE scheme it is easy to see that our technique can also be used to obtain a chosen-ciphertext secure HIBE. Using a technique from =-=[7]-=- it is furthermore possible to reduce the HIBE ciphertext size to three elements, i.e. it is independent of the hierarchy’s depth. As in [19, 40] the security reduction is only exponential in the dept   </text>
<query_num> 2403 </query_num>
<text>   H) assumption. However, Waters’ plain IBE scheme only guarantees chosen-plaintext security. From 2-level Hierarchical IBE to chosen-chipertext secure IBE. Hierarchical identitybased encryption (HIBE) =-=[24, 19]-=- is a generalization of IBE allowing for hierarchical delegation of decryption keys. Recent results from Canetti, Halevi, and Katz [12], further improved upon by Boneh and Katz [9] show a generic and  (Theorem 4.1). 3 6.2 Chosen-ciphertext secure Hierarchical Identity-Based Encryption Hierarchical identity-based encryption is a generalization of IBE to identities supporting hierarchical structures =-=[24, 19]-=-. By the relation to Waters IBE scheme it is easy to see that our technique can also be used to obtain a chosen-ciphertext secure HIBE. Using a technique from [7] it is furthermore possible to reduce   </text>
<query_num> 2404 </query_num>
<text>   d with a DEM to get a full IBE scheme. Kiltz/Galindo+DEM: The IB-KEM from [26] updated with a DEM to get a full IBE scheme. Hybrid Waters/BB+CHK: The IBE scheme obtained by the generic transformation =-=[12, 9]-=- applied to the 2-level hybrid HIBE consisting of Waters’ IBE scheme [40] at the first level and the Boneh/Boyen IBE scheme [5] at the second level (as proposed in [40]). Waters: Waters’ plain chosen-   </text>
<query_num> 2405 </query_num>
<text>   ext secure IBE. Hierarchical identitybased encryption (HIBE) [24, 19] is a generalization of IBE allowing for hierarchical delegation of decryption keys. Recent results from Canetti, Halevi, and Katz =-=[12]-=-, further improved upon by Boneh and Katz [9] show a generic and practical transformation from any chosen-plaintext secure 2-level HIBE scheme to a chosen-ciphertext secure IBE scheme. Since Waters’ I al q) already found numerous applications in [5, 6, 7, 32]. 1.2 Related Work and Comparison As we already pointed out, chosen-ciphertext secure IBE scheme were known to exist using generic reductions =-=[12]-=- based on Waters’ 2-level HIBE [40]. Recently the first direct chosen-ciphertext secure IBE scheme was constructed in [26]. Compared to Waters’ chosen-plaintext secure IBE scheme, the latter direct co omparisons in terms of the respective IBE schemes. The previously most efficient CCA-secure IBE scheme is the one from Kiltz and Galindo [26]. We also compare our scheme with the generic construction =-=[12]-=- obtained from a 2-level HIBE [40, 5] and with the original (only chosenplaintext secure) IBE scheme from Waters. Furthermore we compare or scheme with the reference random-oracle IBE scheme from Bone d with a DEM to get a full IBE scheme. Kiltz/Galindo+DEM: The IB-KEM from [26] updated with a DEM to get a full IBE scheme. Hybrid Waters/BB+CHK: The IBE scheme obtained by the generic transformation =-=[12, 9]-=- applied to the 2-level hybrid HIBE consisting of Waters’ IBE scheme [40] at the first level and the Boneh/Boyen IBE scheme [5] at the second level (as proposed in [40]). Waters: Waters’ plain chosen- l values for α are α = 6, 12, 24). (ii) An exponentiation in G2 takes about α as much time as an exponentiation in G1. We adapt the convention to count one multi-exponentiation as 1.5 exponentiations =-=[12]-=- and the ratio of two pairings as 1.5 pairings [10]. 4 Based on those assumptions in Table 2 we give three variants of our IBE scheme with different tradeoffs between ciphertext size and encryption/de   </text>
<query_num> 2406 </query_num>
<text>   in the latter an attacker is not given access to the decryption oracle. The drawback of the IBE scheme from Boneh-Franklin and Cocks is that security can only be guaranteed in the random oracle model =-=[2]-=-, i.e. in an idealized world where all parties magically get 1sblack-box access to a truly random function. Unfortunately a proof in the random oracle model can only serve as a heuristic argument and  lindo+DEM Hybrid Waters/BB+CHK (Waters) (Boneh/Franklin) secure? √ √ √ — √ Model? √ √ √ √ — |C | 2|G|+128 3|G|+128 3|G|+768 2|G| 1|G|+256 pk n + 3 n + 4 n + 4 n + 2 1 #pairings + #[multi,reg]-exp 0 + =-=[1, 2]-=- 2 + [1, 1] 0 + [0, 3] 0 + [1, 3] 3 + [1, 3] 0 + [0, 2] 0 + [1, 3] 3 + [1, 2] 0 + [0, 2] 0 + [0, 3] 2 + [0, 0] 0 + [0, 2] 1 + [0, 2] 1 + [0, 1] 0 + [0, 1] Table 1: Efficiency comparison for chosen-cip   </text>
<query_num> 2407 </query_num>
<text>   inct element y �= x such that TCR(x) = TCR(y). (In collision resistant hash functions the value x may be chosen by the adversary.) Such hash functions are also called universal one-way hash functions =-=[31]-=- and can be built from arbitrary one-way functions [31, 35]. We define (slightly informal) Adv hash-tcr TCR,H (k) = Pr[H finds a collision in TCR]. Hash function family is said to be a target collisio   </text>
<query_num> 2408 </query_num>
<text>   ision resistant hash functions the value x may be chosen by the adversary.) Such hash functions are also called universal one-way hash functions [31] and can be built from arbitrary one-way functions =-=[31, 35]-=-. We define (slightly informal) Adv hash-tcr TCR,H (k) = Pr[H finds a collision in TCR]. Hash function family is said to be a target collision resistant if the advantage function Adv hash-tcr TCR,H is   </text>
<query_num> 2409 </query_num>
<text>   its of security. The IBE scheme from Sakai and Kasahara [36]. We remark that in the random oracle model there also exists a more efficient IBE scheme that was recently proved secure by Chen and Cheng =-=[14]-=-, and further analyzed and implemented by Chen et al. [15]. Its encryption speed it roughly 1.5 times 13sfaster than ours (and therefore 6-30 times faster than the one from Boneh/Franklin), and decryp   </text>
<query_num> 2410 </query_num>
<text>   ll symmetric operations (like symmetric encryption, random oracle hashes, and MACs). For comparison we mention that relative timings for the various operations are as follows: regular pairing ≈ 3 − 5 =-=[33]-=-, multi-exponentiation ≈ 1.5, regular exponentiation = 1. Variant Element ... in group key decapsulation Encryption Decryption c1 c2 d1 d2 d3 #pairings + #exp in (G1, G2, GT ) V1 G2 G2 G1 G1 G1 V2 G1  ption it performs one exponentiation in G1, one exponentiation in GT , one pairing, and one call a “hash-to-point” hash function , modeled as a random oracle. The latter one was already identified in =-=[33, 15, 21]-=- H1 : {0, 1} n → G∗ 2 to be problematic to implement since it needs one “cofactor” exponentiation in G2. For decryption it needs one exponentiation in G1 and one pairing. The ciphertext space is G1 ×   </text>
<query_num> 2411 </query_num>
<text>   n form of a one-time signature or a MAC with their respective keys. The first “direct” (non 2-level HIBE based) chosen-ciphertext IBE construction in the standard model was given by Galindo and Kiltz =-=[26]-=-. Their construction is based on Waters’ IBE and adds one additional element to the ciphertext that is used for a consistency check in the decryption algorithm. However, in terms of ciphertext size an sen-ciphertext secure IBE scheme were known to exist using generic reductions [12] based on Waters’ 2-level HIBE [40]. Recently the first direct chosen-ciphertext secure IBE scheme was constructed in =-=[26]-=-. Compared to Waters’ chosen-plaintext secure IBE scheme, the latter direct construction adds one additional redundant element to the ciphertext. This element is used as a “check” to defend against in xt is needed to be verifiable, i.e. if one can (publicly) verify if an IBE ciphertext was indeed encrypted with some given identity. Applications of this property can be found in threshold IBE scheme =-=[26]-=-. Using the IB-KEM/DEM paradigm with our IB-KEM constriction and one of the DEMs based on the CMC/EME mode of operation we get an identity-preserving redundancy-free chosen-ciphertext secure IBE schem uction mentioned in the last subsection is possible. 6.3 IB-KEM with Non-Interactive Threshold Decryption Exploiting the public verifiability property of the ciphertext and using the same ideas as in =-=[26]-=- we are able to make key derivation and decapsulation of our IB-KEM construction “threshold”. The ciphertexts of the resulting threshold IB-KEM are shorter in comparison with [26]. 6.4 Selective-Ident chemes from the literature. For a uniform treatment we do all comparisons in terms of the respective IBE schemes. The previously most efficient CCA-secure IBE scheme is the one from Kiltz and Galindo =-=[26]-=-. We also compare our scheme with the generic construction [12] obtained from a 2-level HIBE [40, 5] and with the original (only chosenplaintext secure) IBE scheme from Waters. Furthermore we compare   </text>
<query_num> 2412 </query_num>
<text>   n identity-preserving redundancy-free IBE scheme in the standard model. It is furthermore possible to obtain a full IBE scheme with shorter ciphertexts by using the DEMs based on the CMC [22] and EME =-=[23]-=- mode of operation that avoid the overhead due to the MAC. Then 2sciphertexts of our IBE come with minimal overhead, i.e they are identity-preserving redundancy-free. Following Phan and Pointcheval [3 m our IB-KEM construction with an additional overhead of a DEM which consists of a (one-time secure) symmetric encryption plus additional 128 bits for the MAC. The modes of operation CMC [22] and EME =-=[23]-=- both give chosen-ciphertext secure DEMs provided that the underlying block-cipher is a strong pseudorandom permutation and avoid the usual overhead due to the MAC. We note that for the natural task o   </text>
<query_num> 2413 </query_num>
<text>   nts the first efficient Identity-Based Encryption scheme that is chosen-plaintext secure without random oracles. The proof of his scheme makes use of an algebraic method first used by Boneh and Boyen =-=[5]-=- and security of the scheme is based on the Bilinear Decisional Diffie-Hellman (BDDH) assumption. However, Waters’ plain IBE scheme only guarantees chosen-plaintext security. From 2-level Hierarchical articular we show that “2-BDDHI is at least as strong as mBDDH is at least as strong as BDDH”. The 2-BDDHI (2 Biliear Decisional Diffie-Hellman Inversion) assumption was introduced by Boneh and Boyen =-=[5]-=- and its stronger variants (q-BDDHI for some polynomial q) already found numerous applications in [5, 6, 7, 32]. 1.2 Related Work and Comparison As we already pointed out, chosen-ciphertext secure IBE such that the adversary has to commit to the target identity id ∗ before seeing the public key. Clearly, this is a weaker security requirement. We quickly note that (using an algebraic technique from =-=[5]-=-) by replacing Waters’ hash H with H(id) = h0 · hid 1 (for id ∈ Zp) we get a selective-id chosenciphertext secure IB-KEM. Note that the size of the public-key of this scheme drops to 3 elements. 6.5 C  Replacing Water’s hash H with H(id) = h0 · h R(id) 1 (where R : {0, 1} ∗ → Zp is a random oracle) we get (using the slective-identity secure scheme from the last subsection and a general result from =-=[5]-=-) a chosen-ciphertext secure IB-KEM in the random oracle model. By adding another random oracle to the symmetric key the scheme can then be proved chosen-ciphertext secure with respect to the computat ive IBE schemes. The previously most efficient CCA-secure IBE scheme is the one from Kiltz and Galindo [26]. We also compare our scheme with the generic construction [12] obtained from a 2-level HIBE =-=[40, 5]-=- and with the original (only chosenplaintext secure) IBE scheme from Waters. Furthermore we compare or scheme with the reference random-oracle IBE scheme from Boneh and Franklin [8]. In pairing based   </text>
<query_num> 2414 </query_num>
<text>   ny applications it is sufficient to let sender and receiver agree on a common random session key. This can be accomplished with an identity-based key encapsulation mechanism (IB-KEM) as formalized in =-=[17, 4]-=-. Any IB-KEM can be updated to a full IBE scheme by adding a symmetric encryption scheme. The latter one is also called a data encapsulation scheme (DEM) and the resulting identity-based encryption sc   </text>
<query_num> 2415 </query_num>
<text>   ption it performs one exponentiation in G1, one exponentiation in GT , one pairing, and one call a “hash-to-point” hash function , modeled as a random oracle. The latter one was already identified in =-=[33, 15, 21]-=- H1 : {0, 1} n → G∗ 2 to be problematic to implement since it needs one “cofactor” exponentiation in G2. For decryption it needs one exponentiation in G1 and one pairing. The ciphertext space is G1 ×  6]. We remark that in the random oracle model there also exists a more efficient IBE scheme that was recently proved secure by Chen and Cheng [14], and further analyzed and implemented by Chen et al. =-=[15]-=-. Its encryption speed it roughly 1.5 times 13sfaster than ours (and therefore 6-30 times faster than the one from Boneh/Franklin), and decryption speed is 2 − 3 times as fast (and therefore comparabl   </text>
<query_num> 2416 </query_num>
<text>   rity has emerged as the “right” notion of security for encryption schemes. We stress that, in general, chosen-ciphertext security is a much stronger security requirement than chosen-plaintext attacks =-=[1]-=-, where in the latter an attacker is not given access to the decryption oracle. The drawback of the IBE scheme from Boneh-Franklin and Cocks is that security can only be guaranteed in the random oracl lindo+DEM Hybrid Waters/BB+CHK (Waters) (Boneh/Franklin) secure? √ √ √ — √ Model? √ √ √ √ — |C | 2|G|+128 3|G|+128 3|G|+768 2|G| 1|G|+256 pk n + 3 n + 4 n + 4 n + 2 1 #pairings + #[multi,reg]-exp 0 + =-=[1, 2]-=- 2 + [1, 1] 0 + [0, 3] 0 + [1, 3] 3 + [1, 3] 0 + [0, 2] 0 + [1, 3] 3 + [1, 2] 0 + [0, 2] 0 + [0, 3] 2 + [0, 0] 0 + [0, 2] 1 + [0, 2] 1 + [0, 1] 0 + [0, 1] Table 1: Efficiency comparison for chosen-cip   </text>
<query_num> 2417 </query_num>
<text>   serve as a heuristic argument and has proved to possibly lead to insecure schemes when the random oracles are implemented in the standard model (see, e.g., [11]). Waters’ IBE. To fill this gap Waters =-=[40]-=- presents the first efficient Identity-Based Encryption scheme that is chosen-plaintext secure without random oracles. The proof of his scheme makes use of an algebraic method first used by Boneh and  ations in [5, 6, 7, 32]. 1.2 Related Work and Comparison As we already pointed out, chosen-ciphertext secure IBE scheme were known to exist using generic reductions [12] based on Waters’ 2-level HIBE =-=[40]-=-. Recently the first direct chosen-ciphertext secure IBE scheme was constructed in [26]. Compared to Waters’ chosen-plaintext secure IBE scheme, the latter direct construction adds one additional redu  be public system parameters obtained by running the group parameter algorithm G(1 k ). 4.1 Waters’ Hash We review the hash function H : {0, 1} n → G used in Waters’ identity based encryption schemes =-=[40]-=-. On input of an integer n, the randomized hash key generator HGen(G) chooses n + 1 random group elements h0, . . . , hn ∈ G and returns h = (h0, h1, . . . , hn) ∈ Gn+1 as the public description of th = TimeB−Ω(ɛ−2 ·ln(ɛ−1 )+q), where ɛ = Adv mbddh G,B (k) and q is an upper bound on the number of key derivation/decryption queries made by adversary A. The proof of Theorem 4.1 uses ideas from Waters =-=[40]-=- and will be given in Appendix B. 5 (Redundancy-free) Identity-Based Encryption Given an IB-KEM and a symmetric encryption scheme, a hybrid identity-based encryption scheme can be obtained by using th ain a chosen-ciphertext secure HIBE. Using a technique from [7] it is furthermore possible to reduce the HIBE ciphertext size to three elements, i.e. it is independent of the hierarchy’s depth. As in =-=[19, 40]-=- the security reduction is only exponential in the depth d of the hierarchy, i.e. it introduces, roughly, a multiplicative factor of (nq) d . The keysize of the HIBE scheme is O(nd), whereas the same   </text>
<query_num> 2418 </query_num>
<text>   tandard model. A comparison with the Boneh/Franklin random oracle IBE scheme. Using recent experimental data for atomic primitives (such as exponentiations and pairings) from Granger, Page, and Smart =-=[21]-=- we estimate the efficiency of a possible implementation of our scheme using asymmetric pairings over non-singular elliptic curves. We make a careful comparison with the well known IBE scheme from Bon nd G2. Depending on how this is done we can give different trade-offs between computational efficiency for encryption/decryption and ciphertext size. To this end we will use the following conventions =-=[21]-=-: (i) For general curves an element in G2 takes about α times as much space 11sScheme CCA Standard Size Encrypt Decrypt Key Der. Ours+DEM Kiltz/Galindo+DEM Hybrid Waters/BB+CHK (Waters) (Boneh/Frankli ption it performs one exponentiation in G1, one exponentiation in GT , one pairing, and one call a “hash-to-point” hash function , modeled as a random oracle. The latter one was already identified in =-=[33, 15, 21]-=- H1 : {0, 1} n → G∗ 2 to be problematic to implement since it needs one “cofactor” exponentiation in G2. For decryption it needs one exponentiation in G1 and one pairing. The ciphertext space is G1 ×  cient to guarantee security of k bits). We estimate the cost of encryption and decryption using the timings for each atomic primitive (exponentiations/hashes in G1, G2, GT and pairings) calculated in =-=[21]-=-, where we used the timings for the “pairing friendly curves” and the Tate pairing. Here we counted one hash-into-curve operation used in the Boneh/Franklin scheme (the random oracle H2) as one co-fac r exponentiation =-=[21]-=-. For completeness all used timing data for the atomic primitives is given in Table 5 of Appendix A. The comparison is done with respect to the different curves A-H considered in [21], the names correspond the ones given therein. Important parameters for the curves are the estimated bits of security they provide, the embedding degree α = 6, 12, 24, and the field size of the underl   </text>
<query_num> 2419 </query_num>
<text>   the verifiability property. Suppose an adversary attacking the IBE scheme makes at most q decryption/key derivation queries. A common estimate used here is q = 2 30 (suggested by Bellare and Rogaway =-=[3]-=-). According to Theorem 4.1, to encrypt k bit messages for n = k bit identities and with k bit randomness we get 3k bit ciphertexts to guarantee ≈ k − 30 bits of security. The 30 = log(q) bits of loss s/BB+CHK (Waters) (Boneh/Franklin) secure? √ √ √ — √ Model? √ √ √ √ — |C | 2|G|+128 3|G|+128 3|G|+768 2|G| 1|G|+256 pk n + 3 n + 4 n + 4 n + 2 1 #pairings + #[multi,reg]-exp 0 + [1, 2] 2 + [1, 1] 0 + =-=[0, 3]-=- 0 + [1, 3] 3 + [1, 3] 0 + [0, 2] 0 + [1, 3] 3 + [1, 2] 0 + [0, 2] 0 + [0, 3] 2 + [0, 0] 0 + [0, 2] 1 + [0, 2] 1 + [0, 1] 0 + [0, 1] Table 1: Efficiency comparison for chosen-ciphertext secure IBE sch   </text>
<query_num> 2420 </query_num>
<text>   y 2006 on the new IEEE P1363.3 standard for “Identity-Based Cryptographic Techniques using Pairings” [25] accepts submissions. An alternative but less efficient IBE construction was proposed by Cocks =-=[16]-=- based on quadratic residues. Both IBE schemes (Cocks’ scheme only through Fujisaki-Okamoto [18]) provide security against chosen-ciphertext attacks. In a chosen ciphertext attack, the adversary is gi   </text>
</top>
<top>
<paper_num> 25 </paper_num>
<paper_title>   Maintaining variance and k-medians over data stream windows.  </paper_title>
<paper_abstract>   The sliding window model is useful for discounting stale data in data stream applications. In this model, data elements arrive continually and only the most recent N elements are used when answering queries. We present a novel technique for solving two important and related problems in the sliding window model — maintaining variance and maintaining a k– median clustering. Our solution to the problem of maintaining variance provides a continually updated estimate of the variance of the last N values in a data stream with relative error of at most ɛ using O ( 1 ɛ 2 log N) memory. We present a constant-factor approximation algorithm which maintains an approximate k–median solution for the last N data points using O ( k τ 4 N 2τ log 2 N) memory, where τ &amp;lt; 1/2 is a parameter which trades off the space bound with the approximation factor of O(2 O(1/τ)). 1.  </paper_abstract>
<query_num> 2501 </query_num>
<text>   2τ log 2 N)). For k–median clustering we introduce a sliding-window algorithm that incorporates techniques from the one-pass data-stream clustering algorithm of Guha, Mishra, Motwani, and O’Callaghan =-=[12]-=-. Their algorithm uses O(n τ ) memory and provides a constant factor (2 O(1/τ) ) approximation to the k–median problem. Our sliding window algorithm also makes use of EH to estimate the value of the k  Hulten and Spencer [8, 9] study the problem of maintaining decision trees over sliding windows on data streams. Our results on k–median extends earlier work of Guha, Mishra, Motwani, and O’Callaghan =-=[12]-=-, on one-pass clustering of data streams, to the sliding window model. There is a large body of previous work on k–median clustering [1, 5, 15, 16, 17, 18, 19]. Datar, Gionis, Indyk and Motwani [7] ha dian; however, the techniques also apply to discrete k– median, although the approximation ratios will be different. Our solution incorporates the techniques of Guha, Mishra, Motwani, and O’Callaghan =-=[12]-=-, but ensures a highquality sliding-window clustering. Roughly speaking, the algorithm of Guha et al. is a divide-and-conquer algorithm that builds clusters incrementally and hierarchically as the str  general metric this could simply be the index of x in the point set. The value w(x) is the weight of a median x, i.e., the number of points that x represents. Similar to the algorithm of Guha et al. =-=[12]-=-, if x is of level 0, w(x) = 1, and if x is of level i, then w(x) is the sum of the weights of the level-(i − 1) points that were assigned to x when the level-i clustering was performed. Finally, c(x) members of x; and, if x is of level i &amp;gt; 1, c(x) is the sum over all members y of x, of c(y) + w(y) · ℓ(x, y). Thus, c(x) is an overestimate of the “true” cost of x. As in the algorithm of Guha et al. =-=[12]-=-, we maintain medians at intermediate levels and whenever there are N τ medians at the same level we cluster them into O(k) medians at the next higher level. Thus, each bucket Bi can be split into 1/τ   </text>
<query_num> 2502 </query_num>
<text>   database may help advertisers discover market segments, and clustering telephone-call records may expose fraudulent telephone use. In the systems community, k–median [4, 21] and clustering in general =-=[13, 20, 22]-=- have long been areas of active research. Previous stream clustering algorithms aim to maintain clusters that are valid for all points since the beginning of the stream; that is, old points do not exp   </text>
<query_num> 2503 </query_num>
<text>   e functions over sliding windows that cannot be estimated using their techniques. Babcock, Datar, and Motwani [3] study sampling in the sliding window model. In a recent paper, Gibbons and Tirthapura =-=[10]-=- improved the results from Datar et al. [7] for computing counts and sums over sliding windows. They present a new data structure called waves that has a worst-case update time of O(1) compared to O(l   </text>
<query_num> 2504 </query_num>
<text>   extends earlier work of Guha, Mishra, Motwani, and O’Callaghan [12], on one-pass clustering of data streams, to the sliding window model. There is a large body of previous work on k–median clustering =-=[1, 5, 15, 16, 17, 18, 19]-=-. Datar, Gionis, Indyk and Motwani [7] have considered the problem of maintaining simple statistics over sliding windows. Our work can be considered an extension of that work; we estimate functions ov   </text>
<query_num> 2505 </query_num>
<text>   extends earlier work of Guha, Mishra, Motwani, and O’Callaghan [12], on one-pass clustering of data streams, to the sliding window model. There is a large body of previous work on k–median clustering =-=[1, 5, 15, 16, 17, 18, 19]-=-. Datar, Gionis, Indyk and Motwani [7] have considered the problem of maintaining simple statistics over sliding windows. Our work can be considered an extension of that work; we estimate functions ov 2 ) local search algorithm of Charikar and Guha [5]. While this algorithm uses up to 2k centers, the number can be reduced to k for the final answer via the primal-dual algorithm of Jain and Vazirani =-=[16]-=-. Now consider the running time of the modified algorithm. Whenever a current clustering is desired, we cluster the medians in the oldest bucket Bm and the medians in the suffix buckets Bm∗. If bucket   </text>
<query_num> 2506 </query_num>
<text>   extends earlier work of Guha, Mishra, Motwani, and O’Callaghan [12], on one-pass clustering of data streams, to the sliding window model. There is a large body of previous work on k–median clustering =-=[1, 5, 15, 16, 17, 18, 19]-=-. Datar, Gionis, Indyk and Motwani [7] have considered the problem of maintaining simple statistics over sliding windows. Our work can be considered an extension of that work; we estimate functions ov original data points for the bucket, we can distinguish the expired elements from the active ones and cluster only the latter. We cluster the level-0 medians using the randomized algorithm from Indyk =-=[15]-=-, using the local search algorithm from Charikar and Guha [5] as a subroutine. This procedure requires linear space and takes time Õ(nk) (where n is the number of points that are clustered) while prov er the medians in the oldest bucket Bm and the medians in the suffix buckets Bm∗. If bucket Bm violates Invariant 3 then we cluster only the active elements in it using the Õ(nk) algorithm from Indyk =-=[15]-=- to produce k medians which are then clustered with the medians from the suffix bucket Bm∗. The total number of such medians is at most O( 1 τ2 N τ log N). The running time for clustering the active e   </text>
<query_num> 2507 </query_num>
<text>   extends earlier work of Guha, Mishra, Motwani, and O’Callaghan [12], on one-pass clustering of data streams, to the sliding window model. There is a large body of previous work on k–median clustering =-=[1, 5, 15, 16, 17, 18, 19]-=-. Datar, Gionis, Indyk and Motwani [7] have considered the problem of maintaining simple statistics over sliding windows. Our work can be considered an extension of that work; we estimate functions ov pired elements from the active ones and cluster only the latter. We cluster the level-0 medians using the randomized algorithm from Indyk [15], using the local search algorithm from Charikar and Guha =-=[5]-=- as a subroutine. This procedure requires linear space and takes time Õ(nk) (where n is the number of points that are clustered) while providing a constant factor approximation with high probability.   </text>
<query_num> 2508 </query_num>
<text>   ics over sliding windows. Our work can be considered an extension of that work; we estimate functions over sliding windows that cannot be estimated using their techniques. Babcock, Datar, and Motwani =-=[3]-=- study sampling in the sliding window model. In a recent paper, Gibbons and Tirthapura [10] improved the results from Datar et al. [7] for computing counts and sums over sliding windows. They present   </text>
<query_num> 2509 </query_num>
<text>   ments are associated with weights that decrease over time. In most algorithms that use aging, the weights decrease according to the computationally-simple exponential-decay model (e.g. Gilbert et al. =-=[11]-=-), although linear-decay models are also used. The sliding window model [2, 7] is the other commonly-used mechanism for discounting stale data. Here, only the last N elements to arrive in the stream a   </text>
<query_num> 2510 </query_num>
<text>   ng data in a customer-information database may help advertisers discover market segments, and clustering telephone-call records may expose fraudulent telephone use. In the systems community, k–median =-=[4, 21]-=- and clustering in general [13, 20, 22] have long been areas of active research. Previous stream clustering algorithms aim to maintain clusters that are valid for all points since the beginning of the   </text>
<query_num> 2511 </query_num>
<text>   s that use aging, the weights decrease according to the computationally-simple exponential-decay model (e.g. Gilbert et al. [11]), although linear-decay models are also used. The sliding window model =-=[2, 7]-=- is the other commonly-used mechanism for discounting stale data. Here, only the last N elements to arrive in the stream are considered relevant for answering queries, where N is the window size. Data ative error. 1.1 Related Work Algorithms for streaming data have been an area of much recent research interest. A detailed survey of the algorithmic and database research in data streams is available =-=[2]-=-. Domingos, Hulten and Spencer [8, 9] study the problem of maintaining decision trees over sliding windows on data streams. Our results on k–median extends earlier work of Guha, Mishra, Motwani, and O   </text>
<query_num> 2512 </query_num>
<text>   s that use aging, the weights decrease according to the computationally-simple exponential-decay model (e.g. Gilbert et al. [11]), although linear-decay models are also used. The sliding window model =-=[2, 7]-=- is the other commonly-used mechanism for discounting stale data. Here, only the last N elements to arrive in the stream are considered relevant for answering queries, where N is the window size. Data st), is not large as compared to all other intervals. Moreover—and this is the main technical innovation that distinguishes our algorithms from the ones described by Datar, Gionis, Indyk, and Motwani =-=[7]-=-—we can estimate the contribution of this interval by treating its expired points as though they were “typical” points from the interval. Our main results are as follows. We show how to estimate varia problem. Our sliding window algorithm also makes use of EH to estimate the value of the k–median objective function, though direct application of the techniques from Datar, Gionis, Indyk, and Motwani =-=[7]-=- is impossible (as in variance) because the k–median objective function does not satisfy Property 4 from above. If there are two sets of data points, each tightly clustered about its mean, then the va n [12], on one-pass clustering of data streams, to the sliding window model. There is a large body of previous work on k–median clustering [1, 5, 15, 16, 17, 18, 19]. Datar, Gionis, Indyk and Motwani =-=[7]-=- have considered the problem of maintaining simple statistics over sliding windows. Our work can be considered an extension of that work; we estimate functions over sliding windows that cannot be esti  the sum of assignment distances of the points in Xt is within a constant multiplicative factor 2 O(1/τ) of the optimal. 2. MAINTAINING VARIANCE OVER SLIDING WINDOWS Datar, Gionis, Indyk, and Motwani =-=[7]-=- presented a space lower bound of Ω( 1 log N(log N + log R)) bits for approxiɛ mately (with error at most ɛ) maintaining the sum, where N is the sliding window size and each data value is at most R. A   </text>
<query_num> 2513 </query_num>
<text>   the important issue of concept drift, a major problem in online learning caused when a model based on old data fails to correctly reflect the current state of the world. Domingos, Hulten and Spencer =-=[9]-=- have used sliding windows to deal with this problem in the context of learning decision trees over data streams. One contribution of our work is to begin to address the problem of concept drift by pr rithms for streaming data have been an area of much recent research interest. A detailed survey of the algorithmic and database research in data streams is available [2]. Domingos, Hulten and Spencer =-=[8, 9]-=- study the problem of maintaining decision trees over sliding windows on data streams. Our results on k–median extends earlier work of Guha, Mishra, Motwani, and O’Callaghan [12], on one-pass clusteri   </text>
</top>
<top>
<paper_num> 26 </paper_num>
<paper_title>   Automatic Generation of Subdivision Surface Head Models from Point Cloud Data.  </paper_title>
<paper_abstract>   An automatic procedure is presented to generate a multiresolution head model from sampled surface data. A generic control mesh serves as the starting point for a fitting algorithm that approximates the points in an unstructured set of surface samples, e.g. a point cloud obtained directly from range scans of an individual. A hierarchical representation of the model is generated by repeated refinement using subdivision rules and measuring displacements to the input data. Key features of our method are the fully automated construction process, the ability to deal with noisy and incomplete input data, and no requirement for further processing of the scan data after registering the range images into a single point cloud.  </paper_abstract>
<query_num> 2601 </query_num>
<text>   a large database of several hundred scanned faces. The resulting model has the same resolution as the scanned faces and cannot be readily animated. In the context of medical imaging, SZELISKI et al. =-=[21]-=- minimize the distance between two surfaces obtained from volume scans of human heads by applying local and global deformations in a hierarchical manner. The deformations are modelled by a combination   </text>
<query_num> 2602 </query_num>
<text>   displaced along triangle normal direction to lie on the surface sampled by the point cloud. In this manner, a hierarchy of surfaces is generated with locally encoded details, similar to normal meshes =-=[4]-=-. The construction process is detailed in Section 5. For animation, we rebuild the surface in the changed areas, such that the local detail will follow the deformation properly, see Section 6. 4 Fitti reas of the point cloud with no data we cannot measure displacements, but the subdivision operator generates a smooth surface in these regions. This construction technique is similar to normal meshes =-=[4]-=-, but we sample displacements to a set of sample points instead of to another mesh. By storing only one scalar displacement value per vertex introduced on each level, we achieve a storageefficient hie   </text>
<query_num> 2603 </query_num>
<text>   e a storageefficient hierarchical mesh structure [4, 12]. 6 Animating the Surface During animation, the control mesh vertices of the subdivision surface are displaced via simulated muscle contraction =-=[9]-=-. Since the subdivision hierarchy is built using an interpolating scheme, each refinement level including the control mesh can serve as an approximation to the limit surface for rendering. Depending o   </text>
<query_num> 2604 </query_num>
<text>   ecause of their conceptual simplicity and the availability of efficient graphics hardware for rendering. Adaptive refinement of arbitrary triangle meshes is an active topic in multiresolution editing =-=[26, 11]-=-. While these methods provide powerful tools for mesh deformations, the computational complexity is still considerably too high for real-time applications. In physics-based animation, the vertices of   </text>
<query_num> 2605 </query_num>
<text>   ents to a set of sample points instead of to another mesh. By storing only one scalar displacement value per vertex introduced on each level, we achieve a storageefficient hierarchical mesh structure =-=[4, 12]-=-. 6 Animating the Surface During animation, the control mesh vertices of the subdivision surface are displaced via simulated muscle contraction [9]. Since the subdivision hierarchy is built using an i   </text>
<query_num> 2606 </query_num>
<text>   es apply the subdivision operator uniformly to the surface, some recent results demonstrate adaptive refinement [26, 10]. The algorithms become considerably more complex in these cases. PIGHIN et al. =-=[18]-=- have used an approach based on radial basis functions to match a generic head mesh to several photographs of a head simultaneously. Their approach doesn’t need additional hardware besides a lowcost s   </text>
<query_num> 2607 </query_num>
<text>   mate or interpolate the control mesh after refinement [24]. While most subdivision schemes apply the subdivision operator uniformly to the surface, some recent results demonstrate adaptive refinement =-=[26, 10]-=-. The algorithms become considerably more complex in these cases. PIGHIN et al. [18] have used an approach based on radial basis functions to match a generic head mesh to several photographs of a head   </text>
<query_num> 2608 </query_num>
<text>   mputational complexity is still considerably too high for real-time applications. In physics-based animation, the vertices of a deformable mesh surface are often interpreted as nodes of a spring mesh =-=[13, 22]-=-. Adaptively refining such a massspring system is non-trivial [7]. An efficient method to smooth polygonal geometry proposed by VOLINO et al. [23] can be applied to the deformed geometry that results  al polynomial deformations and local free-form deformations [20]. The method does not require specification of corresponding features on the geometries. The goal in the method presented by LEE et al. =-=[13]-=- is the automated construction of animatable head models from range scans. They adapt a generic face mesh with embedded muscle vectors to range scans of human heads. This process is largely automated,   </text>
<query_num> 2609 </query_num>
<text>   n of the texture to the range data. The model created from the scan data is fully animatable. The generated mesh approximates the input geometry well on a rather coarse detail level. MARSCHNER et al. =-=[17]-=- match a subdivision surface to geometry measured by a range scanner. They use Loop subdivision rules [16] and a fitting algorithm based on the work of HOPPE et al. [5]. A continuous optimization proc tting step reduces the distance from the point cloud to the mesh by global energy minimization. The employed energy functional is essentially the same as used by HOPPE et al. [6] and MARSCHNER et al. =-=[17]-=-. In addition to minimization of the distance between point cloud and mesh, the function accounts for smoothness and constraints defined on the generic mesh. All three steps of this fitting procedure  igure 4). Econstraint(G) = � �v − ˜v� 2 , v∈V c G where Vc G is the set of constrained vertices defined on G, and ˜v denotes the original position of vertex v before the optimization. As described i=-=n [17]-=-, a sparse linear system can be built expressing the function Eglobal, which can be solved using the conjugate gradient method [19]. To be able to set up a linear system, Edist has to be linearized, w rtices of the previous refinement levels, so that coarser levels of refinement serve as an approximation to the head geometry, which is not necessarily the case with approximating subdivision schemes =-=[17]-=-. After fitting, the surface of the deformed generic mesh approximates the point cloud in a least squares sense, i.e. the distance between the points and the surface is minimized. The control mesh ver   </text>
<query_num> 2610 </query_num>
<text>   quires the manual specification of facial features in all of the views. Animation is limited, because each expression needs to be captured in advance. An optimization process proposed by BLANZ et al. =-=[2]-=- generates close approximations to even only a single photograph. This technique draws from a large database of several hundred scanned faces. The resulting model has the same resolution as the scanne   </text>
<query_num> 2611 </query_num>
<text>   sample data. The final fitting step reduces the distance from the point cloud to the mesh by global energy minimization. The employed energy functional is essentially the same as used by HOPPE et al. =-=[6]-=- and MARSCHNER et al. [17]. In addition to minimization of the distance between point cloud and mesh, the function accounts for smoothness and constraints defined on the generic mesh. All three steps   </text>
<query_num> 2612 </query_num>
<text>   ully automatically by exploiting graphics hardware to perform a silhouette-based geometry fitting. Our approach thus extends the idea of the 2D silhouette-based texture mapping technique presented in =-=[14, 15]-=- to threedimensional geometry. To evaluate the current T , we render both G and P into a common frame buffer for one of the three canonical viewing directions along the coordinate axes. The frame buff   </text>
</top>
<top>
<paper_num> 27 </paper_num>
<paper_title>   Gaussian Bounds for Noise Correlation of Functions and Tight Analysis of Long Codes.  </paper_title>
<paper_abstract>   In this paper we derive tight bounds on the expected value of products of low influence functions defined on correlated probability spaces. The proofs are based on extending Fourier theory to an arbitrary number of correlated probability spaces, on a generalization of an invariance principle recently obtained with O’Donnell and Oleszkiewicz for multilinear polynomials with low influences and bounded degree and on properties of multi-dimensional Gaussian distributions. We present two applications of the new bounds to the theory of social choice. We show that Majority is asymptotically the most predictable function among all low influence functions given a random sample of the voters. Moreover, we derive an almost tight bound in the context of Condorcet aggregation and low influence voting schemes on a large number of candidates. In particular, we show that for every low influence aggregation function, the probability that Condorcet voting on k candidates will result in a unique candidate that is preferable to all others is k−1+o(1). This matches the asymptotic behavior of the majority function for which the probability is k−1−o(1). A number of applications in hardness of approximation in theoretical computer science were  </paper_abstract>
<query_num> 2701 </query_num>
<text>   ami-Beckner operator T = Tρ is defined by T = ⊗n i=1T i ρ , where T i is the Bonami-Beckner operator on (Ωi × Ωi,µi). This noise operator is the one most commonly discussed in previous work, see e.g. =-=[10, 14, 17]-=-. In a more recent work [6] the case of Ωi × Ωi with Ti a reversible Markov operator with respect to a measure µi on Ωi was studied. 10Example 2.6 In the first social choice example the space Ω = {±1   </text>
<query_num> 2702 </query_num>
<text>   ami-Beckner operator T = Tρ is defined by T = ⊗n i=1T i ρ , where T i is the Bonami-Beckner operator on (Ωi × Ωi,µi). This noise operator is the one most commonly discussed in previous work, see e.g. =-=[10, 14, 17]-=-. In a more recent work [6] the case of Ωi × Ωi with Ti a reversible Markov operator with respect to a measure µi on Ωi was studied. 10Example 2.6 In the first social choice example the space Ω = {±1 her inequality is trivial. ✷ S ‖fS‖ 2 2 = max i ρ2 i. 3 Background: Influences and Hyper-Contractivity S S̸=∅ i∈S ρi‖fS‖ 2 2 In this section we recall and generalize some definitions and results from =-=[17]-=-. In particular, the generalizations allows the study non-reversible noise operators and correlated ensembles. For the reader that is familiar with [17] it suffices to look at subsections 3.3 and 3.5. n 3.1 The influence of the ith coordinate on f is Infi(f) = ∑ E [Var[fj]]. µ µi 3.2 Multi-linear Polynomials 1≤j≤d In this sub-section we recall and slightly generalize the setup and notation used in =-=[17]-=-. Recall that we are interested in functions on product of finite probability spaces, f : Ω1 × · · · × Ωn → R. For each i, the space of all functions Ωi → R can be expressed as the span of a finite se l only consider functions ψ that are smooth enough that the order of derivatives does not matter). We will also write Q i for the product Q i1 1 · · · Qid d . c j2 σ . 163.4 Hypercontractivity As in =-=[17]-=- the invariance principle requires that the ensembles involved are (2,q,η)-hypercontractive with some η ∈ (0,1) if and only if E[Y ] = 0 and E[|Y | q ] &amp;lt; ∞. Also, if Y is (2,q,η)-hypercontractive then tractivity for sets of random variables which considers all polynomials in the variables, not just multilinear polynomials; see, e.g., Janson [9]. We summarize some of the basic properties below, see =-=[17]-=- for details. Proposition 3.14 Suppose X is a sequence of n1 ensembles and Y is an independent sequence of n2 ensembles. Assume both are (p,q,η)-hypercontractive. Then the sequence of ensembles X ∪Y =   </text>
<query_num> 2703 </query_num>
<text>   ematical economics, from applications in the theory of hardness of approximation in theoretical computer science and from problems in additive number theory. We refer the reader to some recent papers =-=[14, 15, 16, 17, 6, 23, 8]-=- for motivation and general background. The main theorems established here provide tight bounds on the expected value of the product of functions defined on correlated probability spaces. These in tur   </text>
<query_num> 2704 </query_num>
<text>   ematical economics, from applications in the theory of hardness of approximation in theoretical computer science and from problems in additive number theory. We refer the reader to some recent papers =-=[14, 15, 16, 17, 6, 23, 8]-=- for motivation and general background. The main theorems established here provide tight bounds on the expected value of the product of functions defined on correlated probability spaces. These in tur 19]) that if Maj n(x1,...,xn) = sgn( ∑ n i=1 xi), then lim n→∞ E[Majnsgn(TMaj n)] = 2 π arcsin √ ρ. We note that the bound obtained in Theorem 1.1 is a reminiscent of the Majority is Stablest theorem =-=[16, 17]-=- as both involve the arcsin function. However, the two theorems are quite different. The Majority is Stablest theorem asserts that under the same condition as in Theorem 1.1 it holds that E[f(x)f(y)]   implying that for small values of ρ in the context of prediction the majority function is much more predictable than the dictator function. We also note that the “Ain’t over until it’s over” Theorem =-=[16, 17]-=- provides a bound under the same conditions on P[Tf &amp;gt; 1 − δ], for small δ. However, this bound is not tight and does not imply Theorem 1.1. Similarly, Theorem 1.1 does not imply the “Ain’t over until   f(x σ(1),... ,x σ(n)) for all (x1,... ,xn). Kalai conjectured that Majority is the transitive-symmetric function that maximized 3 4 (1+E[fTf]). This was proven using the Majority is Stablest Theorem =-=[16, 17]-=-. Here we obtain similar results for any value of k. Our result is not tight, but almost tight. More specifically we show that: Theorem 1.3 (“Majority is best for Condorcet”) Consider Condorcet voting unctions to their low degree parts when considering the expected value of the product of functions on correlated spaces. • In order to derive an invariance principle we need to extend the approach of =-=[16, 17]-=- to prove the joint invariance of a number of multi-linear polynomials. The proof of the extension appears in sections 3 and 4. The proof follows the same main steps as in [16, 17] but requires a numb   </text>
<query_num> 2705 </query_num>
<text>   ematical economics, from applications in the theory of hardness of approximation in theoretical computer science and from problems in additive number theory. We refer the reader to some recent papers =-=[14, 15, 16, 17, 6, 23, 8]-=- for motivation and general background. The main theorems established here provide tight bounds on the expected value of the product of functions defined on correlated probability spaces. These in tur ami-Beckner operator T = Tρ is defined by T = ⊗n i=1T i ρ , where T i is the Bonami-Beckner operator on (Ωi × Ωi,µi). This noise operator is the one most commonly discussed in previous work, see e.g. =-=[10, 14, 17]-=-. In a more recent work [6] the case of Ωi × Ωi with Ti a reversible Markov operator with respect to a measure µi on Ωi was studied. 10Example 2.6 In the first social choice example the space Ω = {±1 (TρQ)(x) = ∑ ρ |σ| cσxσ. We finally recall the notion of “low-degree influences”, a notion that has proven crucial in the analysis of PCPs in hardness of approximation in computer science (see, e.g., =-=[14]-=-). σ c 2 σ; |σ|&amp;gt;0 15Definition 3.7 The d-low-degree influence of the ith ensemble on Q(X) is Inf ≤d i (Q(X)) = Inf ≤d i (Q) = ∑ c σ:|σ|≤d,σi&amp;gt;0 2 σ . Note that this gives a way to define low-degree in   </text>
<query_num> 2706 </query_num>
<text>   ematical economics, from applications in the theory of hardness of approximation in theoretical computer science and from problems in additive number theory. We refer the reader to some recent papers =-=[14, 15, 16, 17, 6, 23, 8]-=- for motivation and general background. The main theorems established here provide tight bounds on the expected value of the product of functions defined on correlated probability spaces. These in tur s UG-hard to approximate within O(kq 2 )/q k + ǫ. Moreover, for the special case of q = 2, i.e., boolean variables, it gives hardness of (k+O(k 0.525 ))/2 k +ǫ, improving upon the best previous bound =-=[23]-=- of 2k/2 k +ǫ by essentially a factor 2. Finally, again for q = 2, assuming that the famous Hadamard Conjecture is true, this can be improved even further, and the O(k 0.525 ) term can be replaced by  hardness factor of 2 ⌈log 2 k+1⌉ /2 k , which is (k + 1)/2 k for k = 2 r − 1, but can be as large as 2k/2 k for general k. From the quantitative point of view [1] give stronger stronger hardness than =-=[23]-=- for Max k-CSPq, even in the already thoroughly explored q = 2 case. These improvements may seem very small, being an improvement only by a multiplicative factor 2. However, it is well known that it i   </text>
<query_num> 2707 </query_num>
<text>   ematical economics, from applications in the theory of hardness of approximation in theoretical computer science and from problems in additive number theory. We refer the reader to some recent papers =-=[14, 15, 16, 17, 6, 23, 8]-=- for motivation and general background. The main theorems established here provide tight bounds on the expected value of the product of functions defined on correlated probability spaces. These in tur we use in this paper. • We develop a Fourier theory on correlated spaces in Section 2. Previous work considered Fourier theory on one product space and reversible operators with respect to that space =-=[6]-=-. Our results here allows to study non-reversible operators which in turn allows to study products of k correlated spaces. An important fact we prove that is used repeatedly is that general noise oper ed by T = ⊗n i=1T i ρ , where T i is the Bonami-Beckner operator on (Ωi × Ωi,µi). This noise operator is the one most commonly discussed in previous work, see e.g. [10, 14, 17]. In a more recent work =-=[6]-=- the case of Ωi × Ωi with Ti a reversible Markov operator with respect to a measure µi on Ωi was studied. 10Example 2.6 In the first social choice example the space Ω = {±1} × {0, ±1} where element ( as needed. ✷ 256 Gaussian Bounds on Non-reversible Noise forms In this section we prove the main results of the paper: Theorem 1.12 and its relaxations proposition 1.13 and 1.14. As in previous work =-=[16, 17, 6]-=-, the proof idea is to use an invariance principle, in this case Theorem 4.2, together with the Gaussian bounds of Section 5. Since the invariance principle requires working either with low degree pol he relaxed conditions on the influences for k = 2 and for r-wise independent distributions are derived in subsection 6.4 using a “two-threshold” technique. A related technique has been used before in =-=[7, 6]-=-. However, the variant presented here is more elegant, gives more explicit dependency on the influences and allows to exploit s-wise independence. Finally, using a “weak Szemeredi type” type argument   </text>
<query_num> 2708 </query_num>
<text>   known when voter 1 is sampled and are ±1 with 2probability 1/2 otherwise. The notion of predictability is very natural in statistical contexts. It was also studied in a more combinatorial context in =-=[19]-=-. In the first application presented here we show that Theorem 1.1 (“Majority Is Most Predictable”) Let 0 ≤ ρ ≤ 1 and ǫ &amp;gt; 0 be given. Then there exists τ &amp;gt; 0 such that if f : {−1,1} n → [−1,1] satisfi τ for all i, then E[f sgn(Tf)] ≤ 2 π arcsin √ ρ + ǫ, (3) where T is defined in (2). Moreover, it follows from the central limit theorem (see Section 7.5; a version of this calculation also appears in =-=[19]-=-) that if Maj n(x1,...,xn) = sgn( ∑ n i=1 xi), then lim n→∞ E[Majnsgn(TMaj n)] = 2 π arcsin √ ρ. We note that the bound obtained in Theorem 1.1 is a reminiscent of the Majority is Stablest theorem [16   </text>
<query_num> 2709 </query_num>
<text>   n turn imply some new results in the theory of social choice and in the theory of hyper-graphs. Application to hardness of approximation in computer science were derived in subsequent work in [1] and =-=[21]-=-. In our main result we consider a probability measure P defined on a space ∏k i=1 Ωi . Letting fi : (Ωi) n → [0,1],1 ≤ i ≤ k be a collection of low influence functions we derive tight bounds on E[f1  ucts under certain types of correlation, putting it in the same general framework as many other UGC-based hardness results, in particular those for 2-CSPs. In a second beautiful result by Raghavendra =-=[21]-=- the results of the current paper were used to obtained very general hardness results for Max k-CSPP. In [21] it is shown that for every predicate 9P and for every approximation factor which is small he reduction from UG given the integrality gap of the corresponding convex optimization problem. We note that for most predicates the UG hardness of Max k-CSPP is unknown and therefore the results of =-=[21]-=- complement those of [1]. 1.10 Acknowledgments I would like to thank Noam Nisan for suggesting that generalization of the invariance principle should be useful in the social choice context and Gil Kal   </text>
<query_num> 2710 </query_num>
<text>   xiv, two applications of our results to hardness of approximation in computer science have been established. Both results are in the context of the Unique Games conjecture in computational complexity =-=[13]-=-. Furthermore, both result consider an important problem in computer science, that is - the problem of solving constraint satisfaction problems (CSP) . Given a predicate P : [q] k → {0,1}, we define M   </text>
</top>
<top>
<paper_num> 28 </paper_num>
<paper_title>   Group motion segmentation using a Spatio-Temporal Driving Force Model.  </paper_title>
<paper_abstract>   We consider the ‘group motion segmentation ’ problem and provide a solution for it. The group motion segmentation problem aims at analyzing motion trajectories of multiple objects in video and finding among them the ones involved in a ‘group motion pattern’. This problem is motivated by and serves as the basis for the ‘multi-object activity recognition ’ problem, which is currently an active research topic in event analysis and activity recognition. Specifically, we learn a Spatio-Temporal Driving Force Model to characterize a group motion pattern and design an approach for segmenting the group motion. We illustrate the approach using videos of American football plays, where we identify the offensive players, who follow an offensive motion pattern, from motions of all players in the field. Experiments using GaTech Football Play Dataset validate the effectiveness of the segmentation algorithm. 1.  </paper_abstract>
<query_num> 2801 </query_num>
<text>   el leading to nonlinear manifold clustering [32]. The group motion segmentation problem considered here has little in common with them. On the other hand, the non-rigid Structure-from-Motion problems =-=[4, 3, 34]-=- assume nonrigid shape to be linear combination of rigid ones, and nonrigid motion segmentation [35] makes use of local piecewise subspace model, while the group motion under our consideration does no   </text>
<query_num> 2802 </query_num>
<text>   nd point trajectories from consecutive frames of a video sequence, and aim to group them into two or more clusters. While this may appear to be similar to the traditional motion segmentation problems =-=[5, 7, 13, 16, 26, 33, 35, 36, 32]-=-, it is actually different in the following sense. During group motion, the participating objects/people have distinctive and varying motions but the group itself collectively demonstrates an underlyi  26, 33] and the problem eventually boils down to subspace clustering. Other works also exploit dependant/articulated rigid body motion [36] or a motion model leading to nonlinear manifold clustering =-=[32]-=-. The group motion segmentation problem considered here has little in common with them. On the other hand, the non-rigid Structure-from-Motion problems [4, 3, 34] assume nonrigid shape to be linear co   </text>
<query_num> 2803 </query_num>
<text>   nd point trajectories from consecutive frames of a video sequence, and aim to group them into two or more clusters. While this may appear to be similar to the traditional motion segmentation problems =-=[5, 7, 13, 16, 26, 33, 35, 36, 32]-=-, it is actually different in the following sense. During group motion, the participating objects/people have distinctive and varying motions but the group itself collectively demonstrates an underlyi d our method is able to handle them. Turning back to traditional motion segmentation problems we find the majority of them addressing trajectories of feature points from independent 3-D rigid objects =-=[5, 7, 13, 16, 26, 33]-=- and the problem eventually boils down to subspace clustering. Other works also exploit dependant/articulated rigid body motion [36] or a motion model leading to nonlinear manifold clustering [32]. Th   </text>
<query_num> 2804 </query_num>
<text>   nd point trajectories from consecutive frames of a video sequence, and aim to group them into two or more clusters. While this may appear to be similar to the traditional motion segmentation problems =-=[5, 7, 13, 16, 26, 33, 35, 36, 32]-=-, it is actually different in the following sense. During group motion, the participating objects/people have distinctive and varying motions but the group itself collectively demonstrates an underlyi feature points from independent 3-D rigid objects [5, 7, 13, 16, 26, 33] and the problem eventually boils down to subspace clustering. Other works also exploit dependant/articulated rigid body motion =-=[36]-=- or a motion model leading to nonlinear manifold clustering [32]. The group motion segmentation problem considered here has little in common with them. On the other hand, the non-rigid Structure-from-   </text>
<query_num> 2805 </query_num>
<text>   nd point trajectories from consecutive frames of a video sequence, and aim to group them into two or more clusters. While this may appear to be similar to the traditional motion segmentation problems =-=[5, 7, 13, 16, 26, 33, 35, 36, 32]-=-, it is actually different in the following sense. During group motion, the participating objects/people have distinctive and varying motions but the group itself collectively demonstrates an underlyi s little in common with them. On the other hand, the non-rigid Structure-from-Motion problems [4, 3, 34] assume nonrigid shape to be linear combination of rigid ones, and nonrigid motion segmentation =-=[35]-=- makes use of local piecewise subspace model, while the group motion under our consideration does not belong to either of these cases. In this work we employ Lie group theory [27] and in particular es   </text>
<query_num> 2806 </query_num>
<text>   r establish a statistical model over Lie algebra. Lie group and Lie algebra based approaches play roles in invariant visual modeling and recognition [9, 25], robotics [6], 3-D rigid motion estimation =-=[10, 1, 30]-=-, as well as dense flow field modeling [20]. In this work, we discuss a new application to group motion estimation. The proposed model is detailed in Section 2, and its application to group motion seg   </text>
<query_num> 2807 </query_num>
<text>   rk we employ Lie group theory [27] and in particular establish a statistical model over Lie algebra. Lie group and Lie algebra based approaches play roles in invariant visual modeling and recognition =-=[9, 25]-=-, robotics [6], 3-D rigid motion estimation [10, 1, 30], as well as dense flow field modeling [20]. In this work, we discuss a new application to group motion estimation. The proposed model is detaile   </text>
<query_num> 2808 </query_num>
<text>   velopment in the area of video analysis and activity recognition is the need for analyzing these motion patterns of the participating group, which are also called ‘multi-object’ or ‘group’ activities =-=[17, 31, 12, 14, 21, 11, 8, 37, 23, 19, 24, 28]-=-, and various approaches have been proposed to recognize the group motion pattern or detect a change or an anomaly. However, these works assume that all objects are involved in the activity, which is  e most challenging case is American football plays involving a greater number of participants, where the task is to recognize the play strategy of the offensive players from their moving trajectories =-=[14, 19]-=-. In a football play, the offensive players are the participants of the group motion of offense while the defensive players are non-participants. Different offensive participants will give rise to dif   </text>
<query_num> 2809 </query_num>
<text>   velopment in the area of video analysis and activity recognition is the need for analyzing these motion patterns of the participating group, which are also called ‘multi-object’ or ‘group’ activities =-=[17, 31, 12, 14, 21, 11, 8, 37, 23, 19, 24, 28]-=-, and various approaches have been proposed to recognize the group motion pattern or detect a change or an anomaly. However, these works assume that all objects are involved in the activity, which is  e most challenging case is American football plays involving a greater number of participants, where the task is to recognize the play strategy of the offensive players from their moving trajectories =-=[14, 19]-=-. In a football play, the offensive players are the participants of the group motion of offense while the defensive players are non-participants. Different offensive participants will give rise to dif ive) ones solely by their motion trajectories. The GaTech Football Play Dataset is a newly established dataset for group motion pattern/activity modeling and analysis. Recent works using this dataset =-=[19, 29]-=- reported results on play strategy recognition. The dataset consists of a collection of 155 NCAA football game videos (of which 56 are now available). Each video is a temporally segmented recording of   </text>
<query_num> 2810 </query_num>
<text>   velopment in the area of video analysis and activity recognition is the need for analyzing these motion patterns of the participating group, which are also called ‘multi-object’ or ‘group’ activities =-=[17, 31, 12, 14, 21, 11, 8, 37, 23, 19, 24, 28]-=-, and various approaches have been proposed to recognize the group motion pattern or detect a change or an anomaly. However, these works assume that all objects are involved in the activity, which is  ts interact with each other, and typical activities include approaching, chasing, hand-shaking, robbery, fighting, etc. [12, 21, 11, 37, 23, 24, 28]. Also, a group activity in an outdoor airport ramp =-=[8]-=- involves several individual activities occurring in a constrained area and in a specific temporal order. The most challenging case is American football plays involving a greater number of participant   </text>
</top>
<top>
<paper_num> 29 </paper_num>
<paper_title>   CoScan: cooperative scan sharing in the cloud.  </paper_title>
<paper_abstract>   We present CoScan, a scheduling framework that eliminates redundant processing in workflows that scan large batches of data in a map-reduce computing environment. CoScan merges Pig programs from multiple users at runtime to reduce I/O contention while adhering to soft deadline requirements in scheduling. This includes support for join workflows that operate on multiple data sources. Our solution maps well to workflows at many Internet companies which reuse data from a common set of inputs. Experiments on the PigMix data analytics benchmark exhibit orders of magnitude reduction in resource contention with minimal impact on latency.  </paper_abstract>
<query_num> 2901 </query_num>
<text>   L15 L16 L17 Figure 1: Fraction of time spent on loading and scanning data. crementally arriving data, and support for task versioning and data provenance. The system is built on top of the Pig/Hadoop =-=[16, 28]-=- platform. Hadoop is a scalable, fault-tolerant system for executing map-reduce [8] jobs and Pig provides a high-level language for describing relational algebra style operations on semi-structured da dlines even in the presence of cost estimation errors. 7. DISCUSSION We evaluated several techniques that demonstrate the effectiveness of scan sharing for join and non-join jobs using the Pig/Hadoop =-=[16, 28]-=- platform as a motivating application. These include greedy ordering and local improvement algorithms that balance the benefits of scan sharing with meeting soft deadlines, join extensions optimized f   </text>
<query_num> 2902 </query_num>
<text>   a manner that prevents query starvation and accounts for precedence constraints. Scan sharing is studied for workloads against large datasets stored on tertiary storage in order to minimize I/O cost =-=[25, 31, 37]-=-. Yu and Dewitt [37] explored this in Paradise by reordering queries over data stored on magnetic tape. The reordering achieves sequential I/O by collecting data requirements during a pre-execution ph   </text>
<query_num> 2903 </query_num>
<text>   also the first system that explores scan sharing for map and reduce-side joins that share one or more input files. Multiple works on scheduling theory study the problem of multicriteria optimization =-=[18, 19, 23]-=- in both single and parallel computing environments. Lai et al. [21] implement an auction-based resource allocation scheme in Tycoon in which users are allotted resources based on the amount they are   </text>
<query_num> 2904 </query_num>
<text>   employ a dynamic program to identify cases in which the cost of scan sharing outweigh potential benefits. In addition to the mechanisms for scan sharing, our paper also tackles policy. Agrawal et al. =-=[2]-=- provide a solution to maximize scan sharing in an online setting. Map-reduce jobs are grouped into batches so that sequential scans of large files are shared among as many simultaneous jobs as possib   </text>
<query_num> 2905 </query_num>
<text>   ess large amounts of semi-structured user content in order to correlate, mine, and extract valuable features. In turn, they have developed several distributed and scalable Cloud processing frameworks =-=[6, 9, 16, 22]-=- for large scale computations. Many of the data processing tasks involve multiple, inter-dependent steps that operate on large batches of continually arriving data performing data-intensive, disk-base   </text>
<query_num> 2906 </query_num>
<text>   ess large amounts of semi-structured user content in order to correlate, mine, and extract valuable features. In turn, they have developed several distributed and scalable Cloud processing frameworks =-=[6, 9, 16, 22]-=- for large scale computations. Many of the data processing tasks involve multiple, inter-dependent steps that operate on large batches of continually arriving data performing data-intensive, disk-base dlines with resource usage by maximizing the net reward after operating costs. We also plan to validate the generality of our approach in other data processing systems (such as Hive [33] and BigTable =-=[6]-=-). This includes implementing specific multi-query optimization mechanisms to support scan sharing and validating our cost model for scheduling. In addition, CoScan requires better cost estimators tha   </text>
<query_num> 2907 </query_num>
<text>   future works in Section 7. 2. RELATED WORK Prior works on multi-query optimization allow database queries that are executed together to share work and eliminate redundant data access and computation =-=[5, 12, 15, 17, 34, 38]-=-. These works describe mechanisms for scan sharing that include pipelining of results to queries with common sub-expressions [17] and reordering of queries to maximize the probability that data in the   </text>
<query_num> 2908 </query_num>
<text>   future works in Section 7. 2. RELATED WORK Prior works on multi-query optimization allow database queries that are executed together to share work and eliminate redundant data access and computation =-=[5, 12, 15, 17, 34, 38]-=-. These works describe mechanisms for scan sharing that include pipelining of results to queries with common sub-expressions [17] and reordering of queries to maximize the probability that data in the nner et. al. describe a scan sharing solution for main memory databases in which incoming queries are batched together and share a single cursor that continuously scans the data table. Candea et. al. =-=[5]-=- present a similar solution for concurrent analysis in data warehouses. Namely, incoming queries latch onto a single physical plan and share the output of continuous scans of a shared fact table. Stil   </text>
<query_num> 2909 </query_num>
<text>   ilar solution for concurrent analysis in data warehouses. Namely, incoming queries latch onto a single physical plan and share the output of continuous scans of a shared fact table. Still other works =-=[35, 36]-=- focus on the scheduling aspects. Queries that access the same portion of a data table are executed together in a manner that prevents query starvation and accounts for precedence constraints. Scan sh   </text>
<query_num> 2910 </query_num>
<text>   ins that share one or more input files. Multiple works on scheduling theory study the problem of multicriteria optimization [18, 19, 23] in both single and parallel computing environments. Lai et al. =-=[21]-=- implement an auction-based resource allocation scheme in Tycoon in which users are allotted resources based on the amount they are willing to pay. Dua et al. [10] put forth fraction of soft deadlines   </text>
<query_num> 2911 </query_num>
<text>   l factors contribute to these differences including the amount of network congestion and data skew. While better makespan estimates can be derived through in-depth analysis of the map-reduce pipeline =-=[7, 13, 24]-=-, these techniques are beyond the scope of this paper. Definition 5. The resource utilization of a merged job consisting of J1 and J2 is ∑ F ∈f(J1)∪f(J2) su(F ) + cu(J1) + cu(J2). Considering scan cos ilable reduce slots. CoScan relies on past execution of jobs for cost estimation, which may lead to errors (i.e. data skew in the input file or configuration change in the Hadoop cluster). Many works =-=[7, 13, 24]-=- exist that estimate the running time of jobs by sampling the input data or performing a static analysis of the execution pipeline. These techniques are complementary and can be applied in CoScan to i   </text>
<query_num> 2912 </query_num>
<text>   nned only once. This optimization proved effective in lowering resource footprint and improving overall system throughputfor workflows in Nova [27], a data management system developed at Yahoo. Hive =-=[33]-=- provides similar mechanisms for sharing the work of loading and parsing data that is accessed by multiple queries. Nykiel et al. [26] further extend these works by providing a costbased optimization  o balance soft deadlines with resource usage by maximizing the net reward after operating costs. We also plan to validate the generality of our approach in other data processing systems (such as Hive =-=[33]-=- and BigTable [6]). This includes implementing specific multi-query optimization mechanisms to support scan sharing and validating our cost model for scheduling. In addition, CoScan requires better co   </text>
<query_num> 2913 </query_num>
<text>   relational algebra style operations on semi-structured data that are compiled into map-reduce. Reuse of task modules and input data at the workflow abstraction layer enables multi-query optimization =-=[32]-=- opportunities such as sharing common input data. Typically, data processing tasks in Nova are scheduled either periodically (once per week) or triggered in a data-driven manner (on arrival of new dat   </text>
<query_num> 2914 </query_num>
<text>   s in Nova [27], a data management system developed at Yahoo. Hive [33] provides similar mechanisms for sharing the work of loading and parsing data that is accessed by multiple queries. Nykiel et al. =-=[26]-=- further extend these works by providing a costbased optimization to scan sharing in map-reduce. Specifically, they employ a dynamic program to identify cases in which the cost of scan sharing outweig   </text>
<query_num> 2915 </query_num>
<text>   vide non-sequential processing by partitioning the data into fragments that are physically contiguous on the tertiary device and scheduling concurrent queries on a per fragment basis. Andrade et. al. =-=[3]-=- describe a general framework for caching and reusing intermediate results to reduce I/O. Map-reduce jobs also benefit from scan sharing. For instance, Pig programs that share data are merged [14] int   </text>
</top>
<top>
<paper_num> 30 </paper_num>
<paper_title>   A new paradigm for public key identification.  </paper_title>
<paper_abstract>   The present article investigates the possibility of designing zero-knowledge identification schemes based on hard problems from coding theory. Zero-knowledge proofs were introduced in 1985, in a paper by Goldwasser, Micali and Rackoff ([16]). Their practical significance was soon demonstrated in the work of Fiat and Shamir ([11]), who turned zero-knowledge proofs of quadratic residuosity into efficient means of establishing user identities. In the present paper, we propose a new identification scheme, based on error-correcting codes, which is zero-knowledge and seems of practical value. Furthermore, we describe several variants, including one which has an identity based  character. The security of our schemes depends on the hardness of finding a word of given syndrome and prescribed (small) weight with respect to some randomly generated binary linear error-correcting code. This is, of course, not the first attempt to design a cryptographic scheme using tools from coding theory. The dif...  </paper_abstract>
<query_num> 3001 </query_num>
<text>   be recovered. A drastic way to ensure this property is to model the commit function hui by a truly random function. This hypothesis has already been used by various authors and is nicely developed in =-=[4]-=-. The underlying complexity-theoretic model is the Turing machine with random oracle. Another possibility is to add randomness to the hash function: instead of computing the image of u by a collision- any further assumption. A way to circumvent this difficulty is to assume that the hash function is actually a random function. This hypothesis has already been used in [11] and is nicely developed in =-=[4]-=-. The underlying complexity-theoretic model is the Turing machine with random oracle: both P and V are probabilistic oracle Turing machines, the oracle providing, upon request, specific values for the chines and over the oracle. As pointed out in [4], the simulator has to include a simulation of the oracle, which may appear difficult since the oracle is an infinite object. The solution proposed in =-=[4]-=- is to allow the simulator to prescribe a small (polynomial-time) piece of the oracle and have the rest filled at random. We thus get a proof of the following. Theorem 3 In the random oracle model, th   </text>
<query_num> 3002 </query_num>
<text>   e using tools from coding theory. The difference is that identification protocols do not follow the public key paradigm based on trap-door functions and described in the seminal Diffie-Hellman paper (=-=[8]-=-). Rather, they only require one-way functions, which opens the way to using, in a rather direct manner, simple combinatorial problems of the kind provided by coding theory. The resulting schemes comp ing where the transmitter and the receiver share a common key, whose secrecy is requested for proper operation. A major breakthrough took place in 1976 with the appearance of public-key cryptography (=-=[8]-=-). In their paper, Diffie and Hellman proposed a new concept, allowing the use of two matching keys, one for encryption and a different one for decryption. The main novel character of the concept is t techniques have continued to rely on number theory, even though the new protocols do not exactly follow the basic public key paradigm invented by Diffie and Hellman and requiring trap-door functions (=-=[8]-=-). Rather, they are based on one-way functions, which is a less stringent requirement and which opens the way to using simpler techniques, more combinatorial in spirit. In 1989, there were two attempt   </text>
<query_num> 3003 </query_num>
<text>   g acceptable keys contradicts our intractability assumption SD(`; ffi). Thus, our approach already provides a strong argument towards the security of our scheme. Furthermore, the reader familiar with =-=[9, 10, 3]-=- will easily translate our statements into the respective formalism of these papers. In order to carry the proofs through, we will make use of a property of the family of hash functions we use, namely   </text>
<query_num> 3004 </query_num>
<text>   m coding theory. Zero-knowledge proofs were introduced in 1985, in a paper by Goldwasser, Micali and Rackoff ([16]). Their practical significance was soon demonstrated in the work of Fiat and Shamir (=-=[11]-=-), who turned zero-knowledge proofs of quadratic residuosity into efficient means of establishing user identities. In the present paper, we propose a new identification scheme, based on error-correcti ge proofs, introduced in 1985, in a paper by Goldwasser, Micali and Rackoff ([16]) and whose practical significance for public key identification was soon demonstrated in the work of Fiat and Shamir (=-=[11]-=-). Still, zero-knowledge based techniques have continued to rely on number theory, even though the new protocols do not exactly follow the basic public key paradigm invented by Diffie and Hellman and  to be made available by means of a directory or that they have to be certified by the issuing authority. We will consider below a variant of this scheme with an identity-based character. 7. Following =-=[11]-=-, our identification scheme can be turned into a signature scheme as follows: ffl prepare r commitments c j 1 ; c j 2 ; c j 3 , j = 1; \Delta \Delta \Delta r according to the instructions for step 1 a  be performed unless some specific arithmetical co-processor is added. With such a device on board, 768 modular multiplications are done in approximately 500 milliseconds. ffl The schemes proposed in =-=[11, 9]-=- and usually called the Fiat-Shamir schemes, only use modular multiplication. They are based on one or several secret keys, whose respective square modulo some fixed large number n is public. When a s ting strategy, at least without any further assumption. A way to circumvent this difficulty is to assume that the hash function is actually a random function. This hypothesis has already been used in =-=[11]-=- and is nicely developed in [4]. The underlying complexity-theoretic model is the Turing machine with random oracle: both P and V are probabilistic oracle Turing machines, the oracle providing, upon r   </text>
<query_num> 3005 </query_num>
<text>   ntification protocols from hard combinatorial problems. Besides Shamir&amp;apos;s PKP and the present SD, other proposals have been made that belong to the same family: one was put forward by Pointcheval (see =-=[28]-=-) using the so-called perceptron&amp;apos;s problem; another one, due to the author ([36]), is based on the problem of solving linear equations modulo a small prime, the unknowns being subject to the condition   </text>
<query_num> 3006 </query_num>
<text>   pre-processing (based on the parity matrix only), it is still hard to produce a minimum-weight word leading to this syndrome. Non-approximability results for the minimum distance of a code appear in =-=[1]-=-. As for the the hardness of random instances, the question has been investigated by various researchers, especially for families of random codes with a constant information rate. Such codes can be ob   </text>
<query_num> 3007 </query_num>
<text>   rds achieving the minimum distance or to find words of given syndrome whose weight is close to the minimum distance. Several probabilistic algorithms have been proposed that solve these problems (see =-=[19, 34, 6]-=-) but their running time is exponential. In practical terms, it appears relatively easy to design efficient probabilistic algorithms which find words of very low weight and given syndrome, and similar l parameters that guarantee that the concrete problem of finding short codewords is beyond the limits of current computing technology. A survey of known algorithms for solving this problem appears in =-=[6]-=- with a discussion of their possible implementations and of their actual performances. We refer the reader to this paper and we only briefly comment on some figures taken from [6], for the case ` = 1= ord of weight p in the code consisting of all words x such that H(x) is 0 or i. We next use the precise asymptotic evaluation of best algorithms computing codewords of small weight, recently given in =-=[6]-=- and confirmed by experiments in moderate sizes. This leads to the following possible sizes: ffl n = 512; m = 256; p = 56 ffl n = 768; m = 384; p = 84 ffl n = 1024; m = 512; p = 110 These values corre  bounding t. Now, using the attacker repeatedly, we find an acceptable key in time T + 10t ffl 3 . This can be compared with the time needed to attack the SD problem by the best known algorithms (see =-=[6]-=-). The figures should be convincing enough for codes of size 1024 although they cannot really justify the smaller parameter size that we suggest. But the same is true of all proofs that support variou   </text>
<query_num> 3008 </query_num>
<text>   s) = i consists entirely of zeros and ones. Thus the underlying difficult problem is a modular knapsack. Although it is known that knapsacks can be attacked by methods based on lattice reduction (see =-=[21, 7]-=-), it is clear also that these methods do not apply to the modular case, at least when the modulus q is very small. Possible values for the scheme are (with the same notations as above) ffl n = 196, m   </text>
<query_num> 3009 </query_num>
<text>   t one for decryption. The main novel character of the concept is that the encryption key need not be kept secret. Shortly afterwards, Rivest, Shamir and Adleman invented the celebrated RSA algorithm (=-=[29]-=-). This algorithm is a public key system making heavy use of operations modulo a large integer n obtained by multiplying together two prime numbers and whose security is related to difficulty of facto   </text>
<query_num> 3010 </query_num>
<text>   ut treats provers more uniformly: for any given a and large enough I, it outputs an S such that \Pi(I; S) holds, with probability ? P rfACC(P;V; I)g \Gamma jIj \Gammaa . Another definition appears in =-=[3]-=-. It also treats the prover uniformly but allows a knowledge errorsto appear in the success probability for the extractor: namely, the extractor outputs an S such that \Pi(I; S) holds, within an expec h r = 1 is 2=3. We feel that the subtleties involved in the various definitions of soundness will only be of interest to the zero-knowledge experts and we do not go further on this topic referring to =-=[9, 10, 3]-=- for more information. We will take a much simpler path here by only considering the case of a polynomial time probabilistic machine ~ P which operates with an empty knowledge tape, i.e. without the s g acceptable keys contradicts our intractability assumption SD(`; ffi). Thus, our approach already provides a strong argument towards the security of our scheme. Furthermore, the reader familiar with =-=[9, 10, 3]-=- will easily translate our statements into the respective formalism of these papers. In order to carry the proofs through, we will make use of a property of the family of hash functions we use, namely we suggest. But the same is true of all proofs that support various number-theoretic schemes from the literature. Lemma 1 can be read as proving soundness with knowledge error (2=3) r in the sense of =-=[3]-=-. The following result achieves soundness in the sense of [9], provided that the number of rounds is not too small. Theorem 2 Assume that some probabilistic polynomial time adversary ~ P is accepted w   </text>
</top>
<top>
<paper_num> 31 </paper_num>
<paper_title>   A link-indexed statistical traffic prediction approach to improving IEEE 802.11 PSM.  </paper_title>
<paper_abstract>   IEEE 802.11 power save mode (PSM) is a representative of energy-saving protocols which put wireless network interfaces into sleep during idleness. To save energy, part of the performance of IEEE 802.11 is sacrificed attributed to the wake-up latency thus introduced. This paper proposes a complementary mechanism, called link-indexed statistical traffic predictor (LISP) to improve IEEE 802.11 PSM. LISP employs a simple, light-weight traffic prediction method to speed up the delivery of packets along the end-to-end path. By seeking the inherent correlation between ATIM_ACKs and incoming traffic, nodes en route stay awake in the beacon intervals in which packets are anticipated to arrive. As the result, a ‘‘freeway’ ’ is bridged for packets to rapidly traverse the route. Meanwhile, the number of duty cycles is reduced and more energy is conserved. We have conducted analytical and simulation studies and demonstrated the effectiveness of LISP. The impact of various factors is investigated, including traffic load, number of hops (of routes which connections traverse), ATIM window size and packet size, in both tandem networks and networks of arbitrary topologies.  </paper_abstract>
<query_num> 3101 </query_num>
<text>   . Geographic adaptive fidelity (GAF) [13] and Span [2] are two representative mechanisms that employ the notion of virtual routing core. Ondemand power management (on-demand) [15], STEM [9] and S-MAC =-=[14]-=- are representatives of the third category. Both virtual routing core-based mechanisms and on-demand mechanisms share the similarity that both attempt to keep a subset of nodes awake in the course of   </text>
<query_num> 3102 </query_num>
<text>   . In the first category, Woesner et al. [12] investigate, via simulation, the optimal ratio of the ATIM window size to the beacon interval, and suggest the value of approximately 1/4. Jung and Vaidya =-=[5]-=- propose a mechanism of adjusting the ATIM window size dynamically to accommodate varying traffic loads. Tseng et al. [11], on the other hand, address the problems caused by in-synchronization and pro   </text>
<query_num> 3103 </query_num>
<text>   02.11 PSM that consumes energy of intermediate nodes much more than that of the source and destination nodes, LISP balances energy consumption over the nodes en route. LISP differs from previous work =-=[13,2,15]-=- in that it does not trade energy for better end-toend performance at low to moderate traffic loads, i.e., it reduces the end-to-end delay without consuming more power. Instead, energy is further save of exercising PSM remains. Geographic adaptive fidelity (GAF) [13] and Span [2] are two representative mechanisms that employ the notion of virtual routing core. Ondemand power management (on-demand) =-=[15]-=-, STEM [9] and S-MAC [14] are representatives of the third category. Both virtual routing core-based mechanisms and on-demand mechanisms share the similarity that both attempt to keep a subset of node simulation study. The code for IEEE 802.11 PSM was originally developed by Span [16], and has been adapted to accommodate LISP. We compare LISP against IEEE 802.11 with and without PSM, and on-demand =-=[15]-=-, with respect to the following four metrics: (i) end-to-end delay, D; (ii) energy efficiency, EE, defined as the end-to-end throughput (bits) over the energy (Joule) consumed in the duration of a sim   </text>
<query_num> 3104 </query_num>
<text>   device. The latter leverages the interfaces thus provided and tunes the operations of higher protocol layers in a power-aware manner. For example, the operational mode of the device can be controlled =-=[7]-=- and/or the various power management strategies can be exercised in the MAC layer. Topology control—determination of the adequate transmission power of each node so as to maintain network connectivity   </text>
<query_num> 3105 </query_num>
<text>   e beacon interval, and suggest the value of approximately 1/4. Jung and Vaidya [5] propose a mechanism of adjusting the ATIM window size dynamically to accommodate varying traffic loads. Tseng et al. =-=[11]-=-, on the other hand, address the problems caused by in-synchronization and propose the notion of asynchronous wake-up. In spite of all the efforts, the problem of prolonged end-to-end delay as a resul   </text>
<query_num> 3106 </query_num>
<text>   easing energy efficiency, and spans the physical, MAC, and network layers. Several energy efficient routing protocols have also been proposed, e.g., minimum energy routing [8] and power aware routing =-=[10]-=-, in the network layer. In this paper, we focus on the issue of improving power efficiency in the MAC layer, and will provide a comprehensive overview of existing work that pertains to our work in Sec may be more sensitive to the ATIM window size because it takes time for the prediction chain to finish within the ATIM windows. To investigate all these effects, we vary the ATIM window size TATIM in =-=[10,40]-=- ms and compare the performance of the various mechanisms in the same tandem topology: 1 ! 2 ! ! (H + 1). Fig. 7 gives the performance of the various mechanisms in the tandem network of 4 hops (H = 4)   </text>
<query_num> 3107 </query_num>
<text>   ies can be exercised in the MAC layer. Topology control—determination of the adequate transmission power of each node so as to maintain network connectivity while consuming the minimum possible power =-=[2,6,13]-=-—is another (orthogonal) strategy for increasing energy efficiency, and spans the physical, MAC, and network layers. Several energy efficient routing protocols have also been proposed, e.g., minimum e   </text>
<query_num> 3108 </query_num>
<text>   ies can be exercised in the MAC layer. Topology control—determination of the adequate transmission power of each node so as to maintain network connectivity while consuming the minimum possible power =-=[2,6,13]-=-—is another (orthogonal) strategy for increasing energy efficiency, and spans the physical, MAC, and network layers. Several energy efficient routing protocols have also been proposed, e.g., minimum e 02.11 PSM that consumes energy of intermediate nodes much more than that of the source and destination nodes, LISP balances energy consumption over the nodes en route. LISP differs from previous work =-=[13,2,15]-=- in that it does not trade energy for better end-toend performance at low to moderate traffic loads, i.e., it reduces the end-to-end delay without consuming more power. Instead, energy is further save pose the notion of asynchronous wake-up. In spite of all the efforts, the problem of prolonged end-to-end delay as a result of exercising PSM remains. Geographic adaptive fidelity (GAF) [13] and Span =-=[2]-=- are two representative mechanisms that employ the notion of virtual routing core. Ondemand power management (on-demand) [15], STEM [9] and S-MAC [14] are representatives of the third category. Both v   </text>
<query_num> 3109 </query_num>
<text>   ies can be exercised in the MAC layer. Topology control—determination of the adequate transmission power of each node so as to maintain network connectivity while consuming the minimum possible power =-=[2,6,13]-=-—is another (orthogonal) strategy for increasing energy efficiency, and spans the physical, MAC, and network layers. Several energy efficient routing protocols have also been proposed, e.g., minimum e 02.11 PSM that consumes energy of intermediate nodes much more than that of the source and destination nodes, LISP balances energy consumption over the nodes en route. LISP differs from previous work =-=[13,2,15]-=- in that it does not trade energy for better end-toend performance at low to moderate traffic loads, i.e., it reduces the end-to-end delay without consuming more power. Instead, energy is further save zation and propose the notion of asynchronous wake-up. In spite of all the efforts, the problem of prolonged end-to-end delay as a result of exercising PSM remains. Geographic adaptive fidelity (GAF) =-=[13]-=- and Span [2] are two representative mechanisms that employ the notion of virtual routing core. Ondemand power management (on-demand) [15], STEM [9] and S-MAC [14] are representatives of the third cat   </text>
<query_num> 3110 </query_num>
<text>   ng PSM remains. Geographic adaptive fidelity (GAF) [13] and Span [2] are two representative mechanisms that employ the notion of virtual routing core. Ondemand power management (on-demand) [15], STEM =-=[9]-=- and S-MAC [14] are representatives of the third category. Both virtual routing core-based mechanisms and on-demand mechanisms share the similarity that both attempt to keep a subset of nodes awake in   </text>
<query_num> 3111 </query_num>
<text>   on in the sleep state is, in general, an order of magnitude lower than that in the idle and active (transmitting or receiving) states. As a matter of fact, as indicted in several experimental results =-=[3]-=-, the energy consumed in the idle state is only slightly smaller than that in the active state. This is because a wireless device in the idle state has to continuously listen to the wireless medium fo eld. The simulation environment has been set up as follows. All the nodes communicate with each other with half-duplex radio, and have a uniform communication range of 250m. The energy model given in =-=[3]-=- is used, i.e., it takes 1400mW, 1000mW, 830mW and 130mW for a node to transmit, receive, stay idle and sleep, respectively. The energy consumed for switching between idle and sleep states is assumed   </text>
<query_num> 3112 </query_num>
<text>   rthogonal) strategy for increasing energy efficiency, and spans the physical, MAC, and network layers. Several energy efficient routing protocols have also been proposed, e.g., minimum energy routing =-=[8]-=- and power aware routing [10], in the network layer. In this paper, we focus on the issue of improving power efficiency in the MAC layer, and will provide a comprehensive overview of existing work tha   </text>
</top>
<top>
<paper_num> 32 </paper_num>
<paper_title>   2PSM: An Efficient Framework for Searching Video Information in a Limited-Bandwidth Environment.  </paper_title>
<paper_abstract>   We present a novel technique, called 2-Phase Service Model, for streaming videos to home users in a limited-bandwidth environment. This scheme first delivers some number of non-adjacent data fragments to the client in Phase 1. The missing fragments are then transmitted in Phase 2 as the client is playing back the video. This approach offers many benefits. The isochronous bandwidth required for Phase 2 can be controlled within the capability of the transport medium. The data fragments received during Phase 1 can be used to provide an excellent preview of the video. They can also be used to facilitate VCR-style operations such as fast-forward and fast-reverse. Systems designed based on this method are less expensive because the fast-forward and fast-reverse versions of the video files are no longer needed. Eliminating these files also improves system performance because mapping between the regular files and their fast-forward and fast-reverse versions is no longer part of the VCR operations. Furthermore, since each client machine handles its own VCR-style interaction, this technique is very scalable. We provide simulation results to show that 2-Phase Service Model is able to handle VCR functions efficiently. We also implement a video player called FRV-player. With this prototype, we are able to judge that the visual quality of the previews and VCR-style operations is excellent. These features are essential to many important applications. We discuss the application of FRV-player in the design of a video management system, called VideoCenter. This system is intended for Internet applications such as digital video libraries.  </paper_abstract>
<query_num> 3201 </query_num>
<text>   ponding video frame in the fast#forward \Thetale. The mapping can also add delay to the VCR operations. We note that although band# width renegotiation is another way to provide the VCR functionality =-=[4, 5, 18]-=-, this approach is not suitable for a limited#bandwidth environment. We recall that the band# width is not even enough to support the normal playback. To address the \Thetarst problem, one must be abl   </text>
<query_num> 3202 </query_num>
<text>   ponding video frame in the fast#forward \Thetale. The mapping can also add delay to the VCR operations. We note that although band# width renegotiation is another way to provide the VCR functionality =-=[4, 5, 18]-=-, this approach is not suitable for a limited#bandwidth environment. We recall that the band# width is not even enough to support the normal playback. To address the \Thetarst problem, one must be abl we have discussed in Section 1, bandwidth renegotiation [4] can# not be used for a limited#bandwidth environment.On the other hand, using fast#forward and fast reverse versions of the video \Thetales =-=[18, 2]-=- is expensive and makes the system more complex. The proposed 2PSM offers a more natural environment for supporting such interaction. As soon as the L#fragments have been downloaded, fast#forward and   </text>
<query_num> 3203 </query_num>
<text>   streaming technique, the current version of the video server is quite simple. We will enhance the video server in our VideoCenter project. Some of the sophisticated server technology are presented in =-=[3, 6, 7, 9, 10, 13, 16, 20, 25, 24]-=-. 6.2 System Architecture We implemented our system using the following packages: # Continuous Media Toolkit (CMT) [22] version 3.03b3: CMT provides a exible and complete programming environ# ment for   </text>
<query_num> 3204 </query_num>
<text>   ve our concluding remarks in Section 7. 2 Conventional Pipelining We \Thetarst review the concept of pipelining, and summarize some analyses relevant to this paper. Interested readers are referred to =-=[8, 26]-=- for more detail. A video \Thetale is logically divided into a sequence of data seg# ments (S 0 ; S 1 ; : : : ; Sn\Gamma1 ; Sn ), where the playback duration of S i\Gamma1 must eclipse the time requir   </text>
</top>
<top>
<paper_num> 33 </paper_num>
<paper_title>   Complexity results for security protocols with Diffie-Hellman exponentiation and commuting public key encryption.  </paper_title>
<paper_abstract>   We show that the insecurity problem for protocols with modular exponentiation and arbitrary products allowed in exponents is NP-complete. This result is based on a protocol and intruder model which is powerful enough to uncover known attacks on the Authenticated Group Diffie-Hellman (A-GDH.2) protocol suite. To prove our results, we develop a general framework in which the Dolev-Yao intruder is extended by generic intruder rules. This framework is also applied to obtain complexity results for protocols with commuting public key encryption.  </paper_abstract>
<query_num> 3301 </query_num>
<text>   control the communication network and may combine messages from different protocol sessions. Also, protocol This paper is a full version of work previously published in [=-=Chevalier et al. 2003a-=-] and [=-=Chevalier et al. 2004-=-]. The authors have partially been supported by PROCOPE and IST AVISPA. The second author was also supported by the DFG. Author’s address: Y. Chevalier, IRIT, 31068 Toulouse, France; R. Küsters, ETH Z   </text>
<query_num> 3302 </query_num>
<text>   d w.r.t. a bounded number sessions. We illustrate that our protocol and intruder model is powerful enough to uncover attacks first pointed out by Pereira and Quisquater on the A-GDH.2 protocol suite [=-=Pereira and Quisquater 2001-=-]. The NP-completeness result is also shown for protocols employing commuting public key encryption (such as RSA with common modulus). As a consequence of our proofs, we in addition obtain that the de istic polynomial time, respectively. To illustrate our model and results, in Section 6 we formally specify the A-GDH.2 protocol and present an attack on it first discovered by Pereira and Quisquater [=-=Pereira and Quisquater 2001-=-]. Finally, in Section 7 we apply our method to protocols with commuting public key encryption. We conclude in Section 8. Some parts of our proofs are moved to the appendix. 2. THE PROTOCOL AND INTRUD =-=Steiner et al. 1998-=-] allows a group of people who share pairwise long-term keys to establish a shared secret key using Diffie-Hellman exponentiation. We refer the reader to [=-=Steiner et al. 1998-=-] and [=-=Pereira and Quisquater 2001-=-] for a more detailed description of this protocol. Let P = {1, . . . , n, I} be the set of principals that may be involved in a run of a A-GDH.2 protocol where I is the name of the intruder (who can   the secrets returned by p or p ′ in the second session. Note that since the intruder is not a member of the group of the second session, he should not be able to obtain secret. However, as shown in [=-=Pereira and Quisquater 2001-=-], there exists an attack on P . It is easy to verify that this attack will be found by our decision procedure. 7. TRANSFERRING THE RESULTS TO COMMUTATIVE PUBLIC-KEY ENCRYPTION In this section, we tra   </text>
<query_num> 3303 </query_num>
<text>   hout considering such properties. For example, the Recursive Authentication Protocol by Bull and Otway [=-=Bull and Otway 1997-=-] was proven to be secure when perfect cryptographic functions are employed [=-=Paulson 1997-=-] and was shown to be insecure when the protocol is implemented using the XOR operator [=-=Ryan and Schneider 1998-=-] and its nilpotency property. Section 6 contains another example, the A-GDH.2 protocol s   </text>
</top>
<top>
<paper_num> 34 </paper_num>
<paper_title>   Combining many alignments for speech to speech translation.  </paper_title>
<paper_abstract>   Alignment combination (symmetrization) has been shown to be useful for improving Machine Translation (MT) models. Most existing alignment combination techniques are based on heuristics, and can combine only two sets of alignments at a time. Recently in [1], we proposed a power mean based algorithm that can be optimized to combine an arbitrary number alignment tables simultaneously. In this paper we present an empirical investigation of the merits of the approach for combining a large number of alignments (more than 200 in total before pruning). The results of the study suggest that the algorithm can often improve the performance of speech to speech translation systems for low resource languages. 1.  </paper_abstract>
<query_num> 3401 </query_num>
<text>   EM) algorithm. There have been numerous such algorithms proposed, including [2], [3], [4], [5]. The estimated alignment pairs are used to build the core models of the MT engine, such as phrase tables =-=[6]-=-, hierarchical rules [7], or tree-to-string mappings [8]. Thus, it is crucial that the estimated alignment links are as accurate as possible. One common technique for improving alignment accuracy is t precision but low recall and produces fewer alignments, while their intersection, A∩ = A1 ∩ A2, has high recall but low precision. Various heuristic methods for estimating Ao have been proposed ([3], =-=[6]-=-). The method presented in [3], for example, interpolates between the intersection and union of two asymmetric alignment tables by adding links that are adjacent to intersection links, and connect at   </text>
<query_num> 3402 </query_num>
<text>   alignment algorithm finds links between source and target words using some variation of Expectation Maximization (EM) algorithm. There have been numerous such algorithms proposed, including [2], [3], =-=[4]-=-, [5]. The estimated alignment pairs are used to build the core models of the MT engine, such as phrase tables [6], hierarchical rules [7], or tree-to-string mappings [8]. Thus, it is crucial that the operties were varied is described in the following three subsections. 3.2. Alignment algorithms Alignments were generated using GIZA++ algorithm [13] and an HMM aligner similar to the one proposed in =-=[4]-=-. The GIZA++ uses an EM algorithm based on the IBM Models [2] to generate alignments that maximize the log likelihood of ˆθ ∏S ∑ = arg maxθ s=1 a pθ(fs, a|es), where the form of the model is further r   </text>
<query_num> 3403 </query_num>
<text>   e been numerous such algorithms proposed, including [2], [3], [4], [5]. The estimated alignment pairs are used to build the core models of the MT engine, such as phrase tables [6], hierarchical rules =-=[7]-=-, or tree-to-string mappings [8]. Thus, it is crucial that the estimated alignment links are as accurate as possible. One common technique for improving alignment accuracy is to estimate (one-to-many)   </text>
<query_num> 3404 </query_num>
<text>   e generated in the two E2F and F2E directions have been widely used. Besides the basic methods of Intersection (I) and Union (U), several heuristic methods have been proposed and some are widely used =-=[15, 16]-=-. The method presented in [11] relies less on heuristics and has been shown to perform well. Here we consider all of the aforementioned symmetrization methods,and tag alignments generated with them wi   </text>
<query_num> 3405 </query_num>
<text>   easure would lead to some improvement in overall MT quality with respect to BLEU scores. However, how well alignment Fmeasures actually correlate with BLEU scores is an open question, as explained in =-=[17]-=-. While there is no mathematical problem with optimizing the parameters of the presented PM-based combination algorithm w.r.t. BLEU scores, computationally it is not practical to do so because each it   </text>
<query_num> 3406 </query_num>
<text>   gnment training process. In [9], asymmetric models are jointly trained to maximize the similarity of their alignments by iteratively optimizing an objective function based on agreement heuristics. In =-=[10]-=-, the authors present a technique for combining alignments based on various linguistic resources such as parts of speech, dependency parses, or bilingual dictionaries, and use machine learning techniq   </text>
<query_num> 3407 </query_num>
<text>   interpolated with in-domain data to produce an English language model. For the Pashto LM, we simply used the Pashto side of bilingual corpus. MT models were trained using minimum error rate training =-=[19]-=- with a stack based decoder that uses an A ∗ search. We can see in Table 6 that ”comb.17”, which combines 17 different alignments, has the best BLEU scores with respect to dev set, test1 and test2. Fo   </text>
<query_num> 3408 </query_num>
<text>   ment algorithm finds links between source and target words using some variation of Expectation Maximization (EM) algorithm. There have been numerous such algorithms proposed, including [2], [3], [4], =-=[5]-=-. The estimated alignment pairs are used to build the core models of the MT engine, such as phrase tables [6], hierarchical rules [7], or tree-to-string mappings [8]. Thus, it is crucial that the esti   </text>
<query_num> 3409 </query_num>
<text>   ning alignments based on various linguistic resources such as parts of speech, dependency parses, or bilingual dictionaries, and use machine learning techniques to do alignment combination. Recently, =-=[11]-=- presented a method for combining two alignment tables that is effective and relies minimally on heuristics during the combination process. [12] extended this algorithm by integrating confidence score  directions have been widely used. Besides the basic methods of Intersection (I) and Union (U), several heuristic methods have been proposed and some are widely used [15, 16]. The method presented in =-=[11]-=- relies less on heuristics and has been shown to perform well. Here we consider all of the aforementioned symmetrization methods,and tag alignments generated with them with the following acronyms: Int hich combines 17 different alignments, has the best BLEU scores with respect to dev set, test1 and test2. For our baseline (O68), we combined alignments based on the combination algorithm proposed by =-=[11]-=-. We observe that comb.17 is better than baseline method by 2.18 BLEU score on test2 with 4 references. It is also bet-ter on test1 by 0.81 and on dev set by 0.77 BLEU. We can see from Table 5 that m   </text>
<query_num> 3410 </query_num>
<text>   of 6 for English and 8 for Pashto. We trained the lexicalized reordering model that produced distortion costs based on the number of words that are skipped on the target side, in a manner similar to =-=[18]-=-. We had significant amount of out of domain English sentences (1.4 million) that we interpolated with in-domain data to produce an English language model. For the Pashto LM, we simply used the Pashto   </text>
<query_num> 3411 </query_num>
<text>   proposed, including [2], [3], [4], [5]. The estimated alignment pairs are used to build the core models of the MT engine, such as phrase tables [6], hierarchical rules [7], or tree-to-string mappings =-=[8]-=-. Thus, it is crucial that the estimated alignment links are as accurate as possible. One common technique for improving alignment accuracy is to estimate (one-to-many) alignment tables in both the so   </text>
<query_num> 3412 </query_num>
<text>   to be useful for improving Machine Translation (MT) models. Most existing alignment combination techniques are based on heuristics, and can combine only two sets of alignments at a time. Recently in =-=[1]-=-, we proposed a power mean based algorithm that can be optimized to combine an arbitrary number alignment tables simultaneously. In this paper we present an empirical investigation of the merits of th ng the combination process. [12] extended this algorithm by integrating confidence scores into the framework of [11], and further showed that combining more than 2 alignments can be useful. Recently, =-=[1]-=- introduced a power mean based algorithm for alignment combination. The method not only avoids the use of heuristics, but can also simultaneously combine an arbitrary number of alignment tables, and h rs that can be used to optimize any chosen objective function. The power mean is defined by equation (1) below, where p is a real number in (−∞, ∞). Sp(a1, a2, ..., an) = ( 1 n n∑ k=1 a p k ) 1 p (1) =-=[1]-=- showed that as p → 0, the combination process is equivalent to the logical intersection of the input alignments when the alignments are represented as binary variable tables, and as p → ∞, the combin olate between these extremes. In this paper, we empirically investigate the merit of generating large numbers of different alignments and combining them using the power mean algorithm as presented by =-=[1]-=-. 2. Data We performed all of our experiments on English-Pashto data provided by DARPA for TRANSTAC (Spoken Language Communication and Translation System for Tactical Use) Evaluation of 2010. The TRAN rit of interpolating over many alignments to, in essence, approximate the expected value of the alignments when averaged over all possible alignment generation systems, using the power mean algorithm =-=[1]-=-. 3.1. Generating sets of alignments The first task is to generate a large set of alignments by varying the properties of the aligner, as alluded to above. In the following we describe how we generate   </text>
</top>
<top>
<paper_num> 35 </paper_num>
<paper_title>   Social Turing Tests: Crowdsourcing Sybil Detection  </paper_title>
<paper_abstract>   As popular tools for spreading spam and malware, Sybils (or fake accounts) pose a serious threat to online communities such as Online Social Networks (OSNs). Today, sophisticated attackers are creating realistic Sybils that effectively befriend legitimate users, rendering most automated Sybil detection techniques ineffective. In this paper, we explore the feasibility of a crowdsourced Sybil detection system for OSNs. We conduct a large user study on the ability of humans to detect today’s Sybil accounts, using a large corpus of ground-truth Sybil accounts from the Facebook and Renren networks. We analyze detection accuracy by both “experts ” and “turkers ” under a variety of conditions, and find that while turkers vary significantly in their effectiveness, experts consistently produce near-optimal results. We use these results to drive the design of a multi-tier crowdsourcing Sybil detection system. Using our user study data, we show that this system is scalable, and can be highly effective either as a standalone system or as a complementary technique to current tools. 1  </paper_abstract>
<query_num> 3501 </query_num>
<text>   . 1 http://www.bbc.com/news/technology-19093078 2 http://www.bbc.com/news/technology-18813237 The research community has produced a substantial number of techniques for automated detection of Sybils =-=[4, 32, 33]-=-. However, with the exception of SybilRank [3], few have been successfully deployed. The majority of these techniques rely on the assumption that Sybil accounts have difficulty friending legitimate us  they assume that Sybil accounts create many edges amongst themselves. This leads to the formation of well-defined Sybil communities that have a small quotient-cut from the honest region of the graph =-=[4, 28, 29, 32, 33]-=-. Although similar Sybil community detectors have been shown to work well on the Tuenti OSN [3], other studies have demonstrated limitations of this approach. For example, a study by Yang et al. showe   </text>
<query_num> 3502 </query_num>
<text>   cial networks (OSNs). Sybil accounts represent fake identities that are often controlled by a small number of real users, and are increasingly used in coordinated campaigns to spread spam and malware =-=[6, 30]-=-. In fact, measurement studies have detected hundreds of thousands of Sybil accounts in different OSNs around the world [3,31]. Recently, Facebook revealed that up to 83 million of its users may be fa   </text>
<query_num> 3503 </query_num>
<text>   cial networks (OSNs). Sybil accounts represent fake identities that are often controlled by a small number of real users, and are increasingly used in coordinated campaigns to spread spam and malware =-=[6, 30]-=-. In fact, measurement studies have detected hundreds of thousands of Sybil accounts in different OSNs around the world [3,31]. Recently, Facebook revealed that up to 83 million of its users may be fa e that creators of Sybil accounts are using advanced techniques to create more realistic profiles, either by copying profile data from existing accounts, or by recruiting real users to customize them =-=[30]-=-. Malicious parties are willing to pay for these authentic-looking accounts to better befriend real users. These observations motivate us to search for a new approach to detecting Sybil accounts. Our  attackers have also begun to leverage crowdsourcing. Two recent studies have uncovered crowdsourcing websites where malicious users pay crowdworkersto create Sybil accounts on OSNs and generate spam =-=[21, 30]-=-. These Sybils are particularly dangerous because they are created and managed by real human beings, and thus appear more authentic than those created by automated scripts. Crowdsourced Sybils can als ” Sybils are capable of passively gathering legitimate friends and penetrating the social graph [13]. Similarly, some attackers pay users to create fake profiles that bypass current detection methods =-=[30]-=-. As Sybil creators adopt more sophisticated strategies, current techniques are likely to become less effective. 2.3 Crowdsourcing Sybil Detection In this study, we propose a crowdsourced Sybil detect e, our system requires 500 turkers. Our own experience showed that recruiting this many turkers is not difficult (Table 1). In fact, following our crowdsourcing experiments on this and other projects =-=[30]-=-, we received numerous messages from crowd requesting more tasks to perform. Finally, we estimate the monetary cost of our system. Facebook pays turkers from oDesk $1 per hour to moderate images [10].   </text>
<query_num> 3504 </query_num>
<text>   experience and low education level. 5.2 Temporal Factors and Survey Fatigue It is known that turkers try to finish tasks as quickly as possible in order to earn more money in a limited amount of time =-=[16]-=-. This leads to our next question: do turkers spend less time evaluating profiles than experts, and does this lead to lower accuracy? The issue of time is also related to survey fatigue: does the accu es have measured aspects of Amazon’s Mechanical Turk, including worker demographics [12, 24] and task pricing [5, 11, 19]. There are studies that explore the pros and cons to use MTurk for user study =-=[16]-=-. Many studies address the problem of how to maximize accuracy from inherently unreliable turkers. The most common approach is to use majority voting [17, 25], although this scheme is vulnerable to co   </text>
<query_num> 3505 </query_num>
<text>   have observed malicious HITs asking turkers to send social spam [30], per3 http://www.glassdoor.com/Salary/ Tuenti-Salaries-E245751.htm form search engine optimization (SEO) [21], write fake reviews =-=[23]-=-, and even install malware on their systems [15]. 8 Conclusion and Open Questions Sybil accounts challenge the stability and security of today’s online social networks. Despite significant efforts fro   </text>
<query_num> 3506 </query_num>
<text>   ipants in our user study. (c) Gender sourcing websites [11]. Although we could have paid more, prior work has shown that paying more money does not yield higher quality results on crowdsourcing sites =-=[19]-=-. Sociology Undergraduates. The final group of test subjects are undergraduate students from the Department of Communications at UCSB (Social Science major). These students were asked to take our stud latforms on the web has generated a great deal of interest from researchers. Several studies have measured aspects of Amazon’s Mechanical Turk, including worker demographics [12, 24] and task pricing =-=[5, 11, 19]-=-. There are studies that explore the pros and cons to use MTurk for user study [16]. Many studies address the problem of how to maximize accuracy from inherently unreliable turkers. The most common ap   </text>
<query_num> 3507 </query_num>
<text>   latforms on the web has generated a great deal of interest from researchers. Several studies have measured aspects of Amazon’s Mechanical Turk, including worker demographics [12, 24] and task pricing =-=[5, 11, 19]-=-. There are studies that explore the pros and cons to use MTurk for user study [16]. Many studies address the problem of how to maximize accuracy from inherently unreliable turkers. The most common ap   </text>
<query_num> 3508 </query_num>
<text>   nd, social-Turing tests are resilient to changing attacker strategies, because they are not reliant on specific features. Third, crowdsourcing is much cheaper than hiring full-time content moderators =-=[9, 25]-=-. However, there are several questions that we must answer to verify that this system will work in practice: • How accurate are users at distinguishing between real and fake profiles? Trained content   </text>
<query_num> 3509 </query_num>
<text>   send social spam [30], per3 http://www.glassdoor.com/Salary/ Tuenti-Salaries-E245751.htm form search engine optimization (SEO) [21], write fake reviews [23], and even install malware on their systems =-=[15]-=-. 8 Conclusion and Open Questions Sybil accounts challenge the stability and security of today’s online social networks. Despite significant efforts from researchers and industry, malicious users are   </text>
<query_num> 3510 </query_num>
<text>   sumption that Sybil accounts have difficulty friending legitimate users, and thus tend to form their own communities, making them visible to community detection techniques applied to the social graph =-=[29]-=-. Unfortunately, the success of these detection schemes is likely to decrease over time as Sybils adopt more sophisticated strategies to ensnare legitimate users. First, early user studies on OSNs suc  they assume that Sybil accounts create many edges amongst themselves. This leads to the formation of well-defined Sybil communities that have a small quotient-cut from the honest region of the graph =-=[4, 28, 29, 32, 33]-=-. Although similar Sybil community detectors have been shown to work well on the Tuenti OSN [3], other studies have demonstrated limitations of this approach. For example, a study by Yang et al. showe   </text>
<query_num> 3511 </query_num>
<text>   t this point, we don’t have ground-truth about these profiles, i.e. are they really Sybils? To determine groundtruth, we use the methodology pioneered by Thomas et al. to locate fake Twitter accounts =-=[27]-=-. We monitored the suspicious Facebook profiles for 6 weeks, and observed 573 became inaccessible. Attempting to browse these profiles results in the message “The page you requested was not found,” in   </text>
<query_num> 3512 </query_num>
<text>   them vulnerable to sophisticated attack strategies. For example, Irani et al. demonstrate that “honeypot” Sybils are capable of passively gathering legitimate friends and penetrating the social graph =-=[13]-=-. Similarly, some attackers pay users to create fake profiles that bypass current detection methods [30]. As Sybil creators adopt more sophisticated strategies, current techniques are likely to become   </text>
<query_num> 3513 </query_num>
<text>   they assume that Sybil accounts create many edges amongst themselves. This leads to the formation of well-defined Sybil communities that have a small quotient-cut from the honest region of the graph =-=[4, 28, 29, 32, 33]-=-. Although similar Sybil community detectors have been shown to work well on the Tuenti OSN [3], other studies have demonstrated limitations of this approach. For example, a study by Yang et al. showe   </text>
<query_num> 3514 </query_num>
<text>   upt a real crowdsourced Sybil detector, there is little for them to gain by disrupting our study. Related work on detecting crowdsourcing abuse may be helpful in mitigating this problem in the future =-=[7]-=-. 3.3 Test Subjects In order to thoroughly investigate how accurate different types of users are at detecting Sybils, we ran user studies on three different groups of test subjects. Each individual te   </text>
</top>
<top>
<paper_num> 36 </paper_num>
<paper_title>   Execution suppression: An automated iterative technique for locating memory errors.  </paper_title>
<paper_abstract>   By studying the behavior of several programs that crash due to memory errors, we observed that locating the errors can be challenging because significant propagation of corrupt memory values can occur prior to the point of the crash. In this paper, we present an automated approach for locating memory errors in the presence of memory corruption propagation. Our approach leverages the information revealed by a program crash: when a crash occurs, this reveals a subset of the memory corruption that exists in the execution. By suppressing (nullifying) the effect of this known corruption during execution, the crash is avoided and any remaining (hidden) corruption may then be exposed by subsequent crashes. The newly-exposed corruption can then be suppressed in turn. By iterating this process until no further crashes occur, the first point of memory corruption – and the likely root cause of the program failure – can be identified. However, this iterative approach may terminate prematurely, since programs may not crash even when memory corruption is present during execution. To address this, we show how crashes can be exposed in an execution by manipulating the relative ordering of particular variables within memory. By revealing crashes through this variable re-ordering, the effectiveness and applicability of the execution suppression approach can be improved. We describe a set of experiments illustrating the effectiveness of  </paper_abstract>
<query_num> 3601 </query_num>
<text>   008] allocates heap objects far apart in virtual address space to combat buffer overflows, and protects against dangling pointer errors by preserving freed objects after they are freed. Exterminator [=-=Novark et al. 2007-=-] pinpoints heap-based memory errors and derives runtime patches to avoid them in the current and subsequent executions. Unlike these approaches that are targeted toward heap-based memory errors, our   </text>
<query_num> 3602 </query_num>
<text>   90; Korel and Laski 1988; Zhang et al. 2006b] finds the statements that actually do influence a variable value in a particular execution. The related concept of Relevant Slicing [=-=Agrawal et al. 1993; Gyimothy et al. 1999-=-] has also been studied. In general, program slicing identifies a set of statements that can potentially represent many chains of dependencies in a program, from which it may take considerable time to   </text>
<query_num> 3603 </query_num>
<text>   Agrawal and Horgan 1990; Korel and Laski 1988; Zhang et al. 2006b] finds the statements that actually do influence a variable value in a particular execution. The related concept of Relevant Slicing [=-=Agrawal et al. 1993; Gyimothy et al. 1999-=-] has also been studied. In general, program slicing identifies a set of statements that can potentially represent many chains of dependencies in a program, from which it may tak   </text>
<query_num> 3604 </query_num>
<text>   DIDUCE [=-=Hangal and Lam 2002-=-] automatically extract program invariants and monitor for violations during execution. Other error-exposing approaches are designed for specific kinds of errors. CP-Miner [=-=Li et al. 2006-=-] searches for copy-paste errors in large-scale software, and EXPLODE [=-=Yang et al. 2006-=-] identifies data integrity errors in storage systems. Hardware Support for Debugging. There has been some work o   </text>
<query_num> 3605 </query_num>
<text>   The Nearest Neighbor approach [=-=Renieris and Reiss 2003-=-] compares the spectra for two similar executions (one successful and one failing) to identify the most suspicious parts of a program. Tarantula [=-=Jones et al. 2002-=-] is a statistical approach that ranks program statements according to suspiciousness values determined by how many failing versus passing tests exercise each statement. In general, these approaches a   </text>
<query_num> 3606 </query_num>
<text>   do not write to unintended storage locations, and control does not transfer to unintended targets; the average space and runtime overhead of their approach is around 10%. Ruwase and Lam’s CRED tool [=-=Ruwase and Lam 2004-=-] performs bounds-checking in order to detect buffer overflow attacks, incurring an overhead of 1% to 130%. While these techniques incur relatively low overhead, the main difference compared ACM Trans   </text>
<query_num> 3607 </query_num>
<text>   emory leaks and bloat in C and C++ applications, that results in no false positives. There has been recent work on protecting against heap-based memory errors to improve program reliability. DieHard [=-=Berger and Zorn 2006-=-] provides memory safety with high probability by randomizing the location of objects in a large heap and by replicating execution. The goal of DieHard is the opposite of variable re-ordering: whereas   </text>
<query_num> 3608 </query_num>
<text>   errors to motivate the development of our approach for automatically locating the first point of memory corruption in an execution. The programs used in our study were obtained from [=-=Lu et al. 2005; Narayanasamy et al. 2005; Zhou et al. 2004-=-] and are described in Table I. The first column in the table shows the program name and version number. The second column shows the number of lines of code (in thousands). In the th   </text>
<query_num> 3609 </query_num>
<text>   found. For example, static slicing [=-=Weiser 1984-=-] identifies a subset of program statements that may influence the value of a variable at a program location. Dynamic slicing [=-=Agrawal and Horgan 1990;Korel and Laski 1988; Zhang et al. 2006b-=-] finds the statements that actually do influence a variable value in a particular execution. The related concept of Relevant Slicing [=-=Agrawal et al. 1993; Gyimothy et al. 1999-=-] ha   </text>
<query_num> 3610 </query_num>
<text>   ftware, in which it is often vital to avoid program failures. The Ph.D. dissertation of Michael Bond [=-=Bond 2008-=-] describes two techniques for tolerating memory leak errors. The first technique, Melt [=-=Bond and McKinley 2008-=-], identifies stale objects that a program is not accessing, stores these stale objects to disk, and activates these objects only if a program subsequently accesses them. The second technique, leak pr   </text>
<query_num> 3611 </query_num>
<text>   ging, failure-inducing input is identified [=-=Zeller and Hildebrandt 2002-=-] that allows for the computation of cause-effect chains for failures [=-=Zeller 2002-=-], which can in turn be linked to faulty code [=-=Cleve and Zeller 2005-=-]. This approach involves substituting state (the values of variables) between passing and failing runs. A related Value Replacement idea was proposed [=-=Jeffrey et al. 2008a-=-] that attempts to replace t   </text>
<query_num> 3612 </query_num>
<text>   ing memory errors. AccMon [=-=Zhou et al. 2004-=-] describes hardware support for an invariant-based approach that identifies program instructions that typically access different memory locations. HeapMon [=-=Shetty et al. 2006-=-] takes advantage of extra cores to improve the efficiency of error monitoring for heap memory errors. PathExpander [=-=Lu et al. 2006-=-] provides support to increase the path coverage of dynamic error-det   </text>
<query_num> 3613 </query_num>
<text>   ks to locate errors after they cause a failure, rather than to tolerate their effects during runtime. While our current work attempts to isolate memory corruption in an execution, the Samurai system [=-=Pattabiraman et al. 2008-=-] provides safeguards against corruption of critical data through a memory model called critical memory. Their system uses replication and forward error correction to ensure that non-critical updates   </text>
<query_num> 3614 </query_num>
<text>   neral because it can be used to locate any errors involving corrupted memory. On the other hand, Valgrind and Purify can detect some errors that may not lead to crashes, such as memory leaks. CCured [=-=Necula et al. 2002-=-] is an approach for verifying type-safety of pointers both statically and during runtime, which can be used to find potential memory errors. However, their approach requires modifications to program   </text>
<query_num> 3615 </query_num>
<text>   ng hardware support for dynamic information flow tracking (DIFT). In our work, we also evaluate hardware support for the execution suppression approach, which uses ideas similar to DIFT. Recent work [=-=Chen et al. 2008-=-] describes how deferred exception handling in the Itanium processor can be used to perform DIFT efficiently. 7. CONCLUSIONS This paper presented an automated approach for assisting developers in loca   </text>
<query_num> 3616 </query_num>
<text>   overhead. Finally, we describe a hardware-intensive implementation of execution suppression, using hardware similar to that in dynamic information flow tracking (DIFT) approaches [=-=Dalton et al. 2007; Venkataramani et al. 2008-=-], to further reduce overhead. We then compare the overheads for these different implementations. All implementations consider program executions at the binary instruction level. 4.1 General Implement path coverage of dynamic error-detection tools by executing non-taken paths in a sandbox environment. This allows for error detection in paths that would have otherwise not been analyzed. FlexiTaint [=-=Venkataramani et al. 2008-=-] and Raksha [=-=Dalton et al. 2007-=-] are recent works describing hardware support for dynamic information flow tracking (DIFT). In our work, we also evaluate hardware support for the execution suppressio   </text>
<query_num> 3617 </query_num>
<text>   rey et al. behavior and identifies inputs that violate this model. ESC/Java [=-=Flanagan et al. 2002-=-] identifies certain programming errors at compile-time using an annotation language. Check ’n’ Crash [=-=Csallner and Smaragdakis 2005-=-] derives error conditions statically and then attempts to generate test cases to dynamically verify the existence of errors. Daikon [=-=Ernst et al. 2001-=-] and DIDUCE [=-=Hangal and Lam 2002-=-] automatically   </text>
<query_num> 3618 </query_num>
<text>   tifies stale objects that a program is not accessing, stores these stale objects to disk, and activates these objects only if a program subsequently accesses them. The second technique, leak pruning [=-=Bond and McKinley 2009-=-], predicts leaked objects based on data structure usage patterns, and then reclaims these objects at runtime; an error is thrown if any reclaimed object is later accessed. Other recent work on detect   </text>
<query_num> 3619 </query_num>
<text>   tterns, and then reclaims these objects at runtime; an error is thrown if any reclaimed object is later accessed. Other recent work on detecting memory leaks has resulted in the development of Hound [=-=Novark et al. 2009-=-], a runtime system that helps identify the sources of memory leaks and bloat in C and C++ applications, that results in no false positives. There has been recent work on protecting against heap-based   </text>
<query_num> 3620 </query_num>
<text>   viewed as a technique for exposing errors in software (with a specific focus on causing a program to crash). The problem of exposing software errors has been extensively studied. For example, Eclat [=-=Pacheco and Ernst 2005-=-] infers an operational model of correct program ACM Transactions on Programming Languages and Systems, Vol. V, No. N, Month 20YY.32 · Dennis Jeffrey et al. behavior and identifies inputs that violat   </text>
<query_num> 3621 </query_num>
<text>   y hand until the error is found. For example, static slicing [=-=Weiser 1984-=-] identifies a subset of program statements that may influence the value of a variable at a program location. Dynamic slicing [=-=Agrawal and Horgan 1990; Korel and Laski 1988; Zhang et al. 2006b-=-] finds the statements that actually do influence a variable value in a particular execution. The related concept of Relevant Slicing [=-=Agrawal et al. 1993; Gy -=-  </text>
</top>
<top>
<paper_num> 37 </paper_num>
<paper_title>   CSX: an extended compression format for spmv on shared memory systems.  </paper_title>
<paper_abstract>   The Sparse Matrix-Vector multiplication (SpMV) kernel scales poorly on shared memory systems with multiple processing units due to the streaming nature of its data access pattern. Previous research has demonstrated that an effective strategy to improve the kernel’s performance is to drastically reduce the data volume involved in the computations. Since the storage formats for sparse matrices include metadata describing the structure of non-zero elements within the matrix, we propose a generalized approach to compress metadata by exploiting substructures within the matrix. We call the proposed storage format Compressed Sparse eXtended (CSX). In our implementation we employ runtime code generation to construct specialized SpMV routines for each matrix. Experimental evaluation on two shared memory systems for 15 sparse matrices demonstrates significant performance gains as the number of participating cores increases. Regarding the cost of CSX construction, we propose several strategies which trade performance for preprocessing cost making CSX applicable both to online and offline preprocessing.  </paper_abstract>
<query_num> 3701 </query_num>
<text>   dices for CSR and BCSR. The matrix suite that we used for the performance evaluation of our proposed technique consists of 15 matrices selected from the University of Florida sparse matrix collection =-=[7]-=-. We made an effort to include matrices that arise from different kind of problems. Table 4 presents our matrix suite and the characteristics of every matrix. For the sake of presentation, we have arr   </text>
<query_num> 3702 </query_num>
<text>   duce the index data of non-zero elements across the same matrix row by applying compression. The approach is effective as it can significantly benefit the performance of the multithreaded SpMV kernel =-=[14]-=-. CSR-DU employs a coarse grain delta encoding technique; the sparse matrix is divided into areas, called units, with a variable number of elements, and for each of these areas the minimum size for re   </text>
<query_num> 3703 </query_num>
<text>   ent, where multiple processing cores access the main memory. An approach for alleviating this problem is the reduction of the data volume accessed during the execution of the kernel (working set). In =-=[15]-=-, we proposed the CSR-DU (CSR with Delta Units) storage format as a way to reduce the index data of non-zero elements across the same matrix row by applying compression. The approach is effective as i Since delta values are positive and less or equal than their corresponding column indices, they can be stored in smaller size integers, leading to index data size reduction. The CSR-DU storage format =-=[15]-=- is based on a coarse grain delta encoding scheme. The matrix is divided into areas called units, each of which is characterized by the minimum integer size able to represent the unit’s delta-encoded  requires two-byte integers for storing the column index. 2 2010/12/16Figure 3. Example of the CSR-DU storage format, where a row is split into two units. In the implementation of CSR-DU presented in =-=[15]-=-, unit information is stored in a single byte-array called ctl, which consists of a header with the properties of the unit and the main body that includes the delta-encoded column indices. In the most o two units. The first unit has 5 elements, 1-byte delta sizes and a designator for a new row (nr), while the second unit has 3 elements and 2-byte delta sizes. The actual implementation of CSR-DU in =-=[15]-=- also includes a column index offset from the previous unit in the header. The offset is called ujmp and is stored as a (positive) variable-length integer at the end of the header. 3. Compressed Spars exploit the highly redundant nature of the col ind array and RPCSR, which generates matrix-specific dynamic code by applying aggressive compression on column indices patterns for the whole matrix. In =-=[15]-=- we propose the CSR-DU format, which employs a delta encoding scheme to group areas of non-zero elements. CSR-DU is restricted to non-zero elements along the same row. Another recent work that targets   </text>
<query_num> 3704 </query_num>
<text>   onale as in the horizontal case. Two-dimensional substructures Two-dimensional substructures are common in sparse matrices, especially in those that arise from problems with underlying 2D/3D geometry =-=[1, 11, 12]-=-. Storage formats that exploit these structures, e.g., BCSR, can provide significant speedups over the standard CSR implementation in many cases, since apart from reducing the SpMV working set, they e   </text>
<query_num> 3705 </query_num>
<text>   orage formats traditionally try to exploit contiguous elements, either in one or two dimensions. Examples include the BCSR format [20] and the variable length onedimensional block format described in =-=[19]-=-. The index data compression approach of the CSR-DU storage format is based on the general premise that sparse matrices have dense areas, which do not necessarily contain contiguous non-zero elements. y more aggressive index compression by employing run-length encoding in the delta values in multiple directions. Drawing inspiration from the variable length one-dimensional block format described in =-=[19]-=-, we generalize the notion of sequential elements to elements with a constant distance. As with sequential elements (α, α + 1, α + 2, . . .), elements with a constant distance (α, α + δ, α + 2δ, . . .   </text>
<query_num> 3706 </query_num>
<text>   r 8 threads on Harpertown and 12 threads for Dunnington. (e.g., register and cache blocking) on SMPs or examine reordering techniques to improve locality of references and minimize communication cost =-=[6, 8, 10, 18]-=-. Williams et. al [27] presented an evaluation of SpMV on a set of emerging multicore architectures. Their study covers a wide and diverse range of high-end chip multiprocessors, including recent mult   </text>
<query_num> 3707 </query_num>
<text>   se efforts [11,19,22–24] aim at the optimization of the irregular and indirect accesses on the input vector using methods such as matrix reordering, register blocking, and cache blocking. Other works =-=[17,25]-=- are concerned with the performance problems that arise in matrices with a large number of rows with small length. A significant part of the SpMV optimization techniques reported in the related litera   </text>
<query_num> 3708 </query_num>
<text>   threads for Dunnington. (e.g., register and cache blocking) on SMPs or examine reordering techniques to improve locality of references and minimize communication cost [6, 8, 10, 18]. Williams et. al =-=[27]-=- presented an evaluation of SpMV on a set of emerging multicore architectures. Their study covers a wide and diverse range of high-end chip multiprocessors, including recent multicores from AMD (Opter   </text>
<query_num> 3709 </query_num>
<text>   tions. Thus, the common approach is to store only the nonzero values of the matrix and employ additional indexing information representing the position of these values (index data). Our previous work =-=[2, 9]-=- has identified the memory subsystem as the main performance bottleneck of the SpMV kernel. Obviously, this problem becomes more severe in a multithreaded environment, where multiple processing cores   </text>
</top>
<top>
<paper_num> 38 </paper_num>
<paper_title>   PreDatA - preparatory data analytics on peta-scale machines.  </paper_title>
<paper_abstract>   Abstract—Peta-scale scientific applications running on High End Computing (HEC) platforms can generate large volumes of data. For high performance storage and in order to be useful to science end users, such data must be organized in its layout, indexed, sorted, and otherwise manipulated for subsequent data presentation, visualization, and detailed analysis. In addition, scientists desire to gain insights into selected data characteristics ‘hidden ’ or ‘latent ’ in these massive datasets while data is being produced by simulations. PreDatA, short for Preparatory Data Analytics, is an approach to preparing and characterizing data while it is being produced by the large scale simulations running on peta-scale machines. By dedicating additional compute nodes on the machine as ‘staging ’ nodes and by staging simulations ’ output data through these nodes, PreDatA can exploit their computational power to perform select data manipulations with lower latency than attainable by first moving data into file systems and storage. Such intransit manipulations are supported by the PreDatA middleware through asynchronous data movement to reduce write latency, application-specific operations on streaming data that are able to discover latent data characteristics, and appropriate data reorganization and metadata annotation to speed up subsequent data access. PreDatA enhances the scalability and flexibility of the current I/O stack on HEC platforms and is useful for data pre-processing, runtime data analysis and inspection, as well as for data exchange between concurrently running simulations. I.  </paper_abstract>
<query_num> 3801 </query_num>
<text>   A middleware resides on both staging nodes and the compute nodes on which the application runs. When the application performs I/O actions, PreDatA acquires output data through the ADIOS I/O interface =-=[28]-=-, stages data from compute nodes to staging nodes and performs in-transit data processing along the data flow. Its current implementation exploits the computational resources of compute nodes for data encoding facility is used for in-transit data to provide PreDatA operations access to buffered data with rich meta-data information. A. Data Extraction and Movement PreDatA uses the ADIOS I/O library =-=[28]-=- for integration with the HEC I/O stack and in addition, for PreDatA operations to access the data output by simulations. With ADIOS, PreDatA processing can be added without requiring changes to appli   </text>
<query_num> 3802 </query_num>
<text>   ] applies the MapReduce model to analyze molecular dynamics simulation trajectories and shows some efficiency at tera-byte scale. In contrast, experiences from implementing materialized ground models =-=[40]-=- show poor performance of MapReduce because some of the features provided by MapReduce are unnecessary for its target application. AllPairs [33] gains similar insights in that a mismatch between the a   </text>
<query_num> 3803 </query_num>
<text>   and ‘online’ approach to data output and manipulation can be used to implement the model-to-model comFigure 5. Stream Processing in the Staging Area munications used in coupled high performance codes =-=[3]-=-, [51]. Toward that end, we have integrated into PreDatA the ‘DataSpaces’ data indexing and querying services. DataSpaces provides high level programmable and managed services for (1) data sharing – b  to directly feed data to existing streaming processing systems. Code Coupling. Memory-to-memory code coupling addresses some of the issues faced by PreDatA, such as data movement and re-distribution =-=[3]-=-, [26]. PreDatA provides the underpinnings for supporting the rich model-model communications needed for inter-application interactions [15]. Interactive Computational Steering. Runtime steering can a   </text>
<query_num> 3804 </query_num>
<text>   anges to application code and providing improved flexibility in composing the simulation’s output and analysis pipeline. Scientific Workflows. Scientific workflow systems like Pegasus [14] and Kepler =-=[31]-=- are used to automate scientific data and simulation management. Unlike the end-to-endapproach used in In-situ visualization mentioned above, components in the workflow are usually connected via a fi   </text>
<query_num> 3805 </query_num>
<text>   ation technologies to facilitate better understanding of their data. Specialized data preparation, such as sorting, filtering, and indexing, is needed before data can be understood or visualized [9], =-=[39]-=-, [43]. Our work extends the I/O middleware stack to exploit computational power along the output data flow to perform data preparation, characterization, and re-organization, which would facilitate s   </text>
<query_num> 3806 </query_num>
<text>   data generated by scientific simulations presents daunting challenges to both computational and computer scientists [19], [36]. Recent work in parallel file systems [8], [35], [48] and I/O middleware =-=[22]-=-, [30], [38], [50], [52] aims at optimizing data storage and access for scientific application workloads. Beyond pure high I/O bandwidth, however, scientists also require complex data analysis, search   </text>
<query_num> 3807 </query_num>
<text>   data is moved to disks, to reduce the data volumes to be processed in subsequent workflow steps. Scientific Data Stream Processing. Scientific data stream processing, such as filtering [6], sampling =-=[47]-=-, query [27], and transformation [23] complements our work. This is because PreDatA can be used either as an in-transit data processing framework for implementing streaming processing tasks, or as a d   </text>
<query_num> 3808 </query_num>
<text>   enerated by scientific simulations presents daunting challenges to both computational and computer scientists [19], [36]. Recent work in parallel file systems [8], [35], [48] and I/O middleware [22], =-=[30]-=-, [38], [50], [52] aims at optimizing data storage and access for scientific application workloads. Beyond pure high I/O bandwidth, however, scientists also require complex data analysis, search, and   </text>
<query_num> 3809 </query_num>
<text>   ent of voluminous and complex data generated by scientific simulations presents daunting challenges to both computational and computer scientists [19], [36]. Recent work in parallel file systems [8], =-=[35]-=-, [48] and I/O middleware [22], [30], [38], [50], [52] aims at optimizing data storage and access for scientific application workloads. Beyond pure high I/O bandwidth, however, scientists also require   </text>
<query_num> 3810 </query_num>
<text>   ently scheduling data movement from compute nodes to the Staging Area. The EVPath [17] high performance event system is used for efficient data buffering and manipulation in the Staging Area. The FFS =-=[18]-=- binary data encoding facility is used for in-transit data to provide PreDatA operations access to buffered data with rich meta-data information. A. Data Extraction and Movement PreDatA uses the ADIOS d filtering out undesired regions. All output data (scalars, local arrays, partial chunks of global arrays) are then packed into a contiguous buffer, termed a packed partial data chunk, using the FFS =-=[18]-=- binary data encoding facility (shown as Stage 1b in Fig. 4). The structure of each packed partial data chunk is compatible with the ADIOS output data group definition, and metadata about the data str   </text>
<query_num> 3811 </query_num>
<text>   eration before data is moved to disks, to reduce the data volumes to be processed in subsequent workflow steps. Scientific Data Stream Processing. Scientific data stream processing, such as filtering =-=[6]-=-, sampling [47], query [27], and transformation [23] complements our work. This is because PreDatA can be used either as an in-transit data processing framework for implementing streaming processing t   </text>
<query_num> 3812 </query_num>
<text>   inal results to disk, feeds data to other consumers, and/or performs necessary cleanup. From this description, it should be apparent that the PreDatA data processing model is similar to the MapReduce =-=[12]-=- paradigm, with four notable differences. First, PreDatA’s data processing model requires operations to read data only once, meaning that data is processed in a streaming fashion, as done in other str   </text>
<query_num> 3813 </query_num>
<text>   iring minimal changes to application code and providing improved flexibility in composing the simulation’s output and analysis pipeline. Scientific Workflows. Scientific workflow systems like Pegasus =-=[14]-=- and Kepler [31] are used to automate scientific data and simulation management. Unlike the end-to-endapproach used in In-situ visualization mentioned above, components in the workflow are usually co   </text>
<query_num> 3814 </query_num>
<text>   ization, and re-organization, which would facilitate subsequent data analysis. Data Staging and Offloading in supercomputers. Previous work on data staging and asynchronous I/O [4], [16], [24], [25], =-=[32]-=-, [34], [41] derives substantial performance advantages from hiding I/O latency with asynchronous data movement. Our recent work [1], [2] shows the importance of minimizing interference of asynchronou   </text>
<query_num> 3815 </query_num>
<text>   measures that can be used to validate the veracity of the ongoing simulation, gain understanding of the simulation progress, and potentially, take early action when the simulation operates improperly =-=[20]-=-. The object of our research and topic of this paper is the development of efficient methods that properly prepare data for subsequent inspection, storage, analytics, and even for input into concurren e rich model-model communications needed for inter-application interactions [15]. Interactive Computational Steering. Runtime steering can aid scientists in debugging and monitoring their simulations =-=[20]-=-, [45]. The capability of extracting and inspecting data from running simulation with small overhead and interference makes PreDatA a potential infrastructure for online steering of running applicatio   </text>
<query_num> 3816 </query_num>
<text>   mong applications, so one potential problem with Active Storage is how to manage such resources to meet deadlines for multiple applications and minimize performance downgrade of storage nodes. Abacus =-=[5]-=- demonstrates the benefit of flexible, dynamic function placement in Active Storage, and we are going to investigate similar mechanisms for PreDatA. In-situ Data Analytics and Visualization. Hercules   </text>
<query_num> 3817 </query_num>
<text>   n the workflow are usually connected via a filebased interface, so that the performance of the workflow is very sensitive to data placement and movement and is easily affected by poor I/O performance =-=[13]-=-. PreDatA can serve as an early stage in output pipeline to apply application-specific data reduction, validation, and filtering operation before data is moved to disks, to reduce the data volumes to   </text>
<query_num> 3818 </query_num>
<text>   nces. First, PreDatA’s data processing model requires operations to read data only once, meaning that data is processed in a streaming fashion, as done in other streaming implementations of MapReduce =-=[11]-=-. This is because of limited memory space on staging nodes, which means that this assumption can be removed if additional memory (e.g., SSD or other disk storage) were to be made available on those ma   </text>
<query_num> 3819 </query_num>
<text>   ndexing, and annotation. For example, some analysis tools prefer data to be laid out as contiguous arrays for quick loading [49], and queries can be accelerated if data is properly sorted and indexed =-=[42]-=-. In other words, appropriate data preparation is critical for data analytics, inspection, or visualization to operate efficiently. Finally, ‘hidden’ in the voluminous data sets generated by running s  for our example are) sorted by their labels before searching. The second task performs a range query to discover the particles whose coordinates fall into certain ranges. A bitmap indexing technique =-=[42]-=- is used to avoid scanning the whole particle array, and multiple array chunks are merged to speed up bulk loading. The third task is to generate 1D histograms and 2D histograms on attributes of parti   </text>
<query_num> 3820 </query_num>
<text>   nodes for a diversity of data operations to achieve not only high write performance, but also high read performance and timely monitoring of output data and simulation. Active Storage. Active Storage =-=[37]-=- deploys data processing operations directly on the storage nodes where the data are buffered to reduce the amount of data movement between storage and compute nodes. The storage nodes have limited co   </text>
<query_num> 3821 </query_num>
<text>   preparation, characterization, and re-organization, which would facilitate subsequent data analysis. Data Staging and Offloading in supercomputers. Previous work on data staging and asynchronous I/O =-=[4]-=-, [16], [24], [25], [32], [34], [41] derives substantial performance advantages from hiding I/O latency with asynchronous data movement. Our recent work [1], [2] shows the importance of minimizing int   </text>
<query_num> 3822 </query_num>
<text>   racterization, and re-organization, which would facilitate subsequent data analysis. Data Staging and Offloading in supercomputers. Previous work on data staging and asynchronous I/O [4], [16], [24], =-=[25]-=-, [32], [34], [41] derives substantial performance advantages from hiding I/O latency with asynchronous data movement. Our recent work [1], [2] shows the importance of minimizing interference of async   </text>
<query_num> 3823 </query_num>
<text>   re-organization, which would facilitate subsequent data analysis. Data Staging and Offloading in supercomputers. Previous work on data staging and asynchronous I/O [4], [16], [24], [25], [32], [34], =-=[41]-=- derives substantial performance advantages from hiding I/O latency with asynchronous data movement. Our recent work [1], [2] shows the importance of minimizing interference of asynchronous data movem   </text>
<query_num> 3824 </query_num>
<text>   s operations running in the staging area can implement customized data shuffling and synchronization methods, in our case using the highlyoptimized MPI routines present on the peta-scale machine (see =-=[53]-=-). This is not only to take advantage of the high end communication support available on the peta-scale machine but also to be able to deploy and leverage existing parallel analysis codes written for  cies and to place the data chunks present within the data stream into some desired order to ease implementing such data analysis services. More details about the programming interface can be found in =-=[53]-=-. The staging area is running as a separate MPI program launched independently from the simulation. Each MPI process runs on one staging node. Within each staging node, there are multiple threads in e dge service is also evaluated with GTC. This demonstrates the feasibility of building higher-level data services with PreDatA. For brevity, the implementation details about those operations appear in =-=[53]-=-. A. Experimental Environment Experiments are run on the Oak Ridge National Laboratory’s Cray XT4/XT5 Jaguar platform. The XT5 partition contains 18,688 compute nodes. Each compute node contains two q   </text>
<query_num> 3825 </query_num>
<text>   s used as a common interface for application and PreDatA operations to coordinate sharing data.Data is extracted from compute nodes and moved to the Staging Area via the scheduled, asynchronous RDMA =-=[7]-=- operations. As explained in [2], using asynchronous RDMA reduces the write latency visible at compute nodes. Carefully scheduling such RDMA operations eliminates the potential interference between co   </text>
<query_num> 3826 </query_num>
<text>   s work on data staging and asynchronous I/O [4], [16], [24], [25], [32], [34], [41] derives substantial performance advantages from hiding I/O latency with asynchronous data movement. Our recent work =-=[1]-=-, [2] shows the importance of minimizing interference of asynchronous data movement with the application to achieve overall improvements in simulation time. One observation is that the computational r   </text>
<query_num> 3827 </query_num>
<text>   technologies to facilitate better understanding of their data. Specialized data preparation, such as sorting, filtering, and indexing, is needed before data can be understood or visualized [9], [39], =-=[43]-=-. Our work extends the I/O middleware stack to exploit computational power along the output data flow to perform data preparation, characterization, and re-organization, which would facilitate subsequ   </text>
<query_num> 3828 </query_num>
<text>   xperiences from implementing materialized ground models [40] show poor performance of MapReduce because some of the features provided by MapReduce are unnecessary for its target application. AllPairs =-=[33]-=- gains similar insights in that a mismatch between the application workload and the available MapReduce abstractions can result in poor performance. The two-pass streaming model used by PreDatA appear   </text>
</top>
<top>
<paper_num> 39 </paper_num>
<paper_title>   Dynamic Network Reconfiguration Support for Mobile Computers.  </paper_title>
<paper_abstract>   Hot swapping technology combined with pervasive heterogeneous networks empowers mobile laptop users to select the best network device for their current environment. Unfortunately, the majority of system software remains &amp;quot;customized&amp;quot;for a particular network configuration, and assumes that many characteristics associated with the network environment remain invariant over the runtime of the software. Mobility causes changes in the environment and nullifies many of these assumptions. This leads to severe loss in system functionality when resources are lost, and failure to benefit when resources are gained. This paper describes Physical Media Independence (PMI), an architecture for dynamically diverse network interface management. PMI addressesthree issuesconcerningdynamicnetwork configuration. First, a model for device availability is proposed to accurately determine when a network device is operational. Second, a structured methodology is used to construct adapters that reconfigure the sy...  </paper_abstract>
<query_num> 3901 </query_num>
<text>   0.13 Bandwidth (bps) N/A N/A 736 this time. We intend to reset local retransmission timers that have begun exponential backoff because of disconnection and use Caceres and Iftode&amp;apos;s &amp;quot;triple ack hack&amp;quot; =-=[3]-=- where multiple acknowledgements are sent to the correspondent host to encourage the remote TCP state machine to enter fast retransmit and recovery mode. We allow application-level adaptation through   </text>
<query_num> 3902 </query_num>
<text>   1024 5 model airfone device `tun0&amp;apos; label `GTE Airfone&amp;apos; circuit connection 627 60 328 connection limit 1500 Figure 4: Sample Cost Models enters the disconnectand uses the Berkeley Packet Filter (BPF) =-=[15]-=- to promiscuously listen for all ICMP echo replies. This allows the implementation to distinguish between link-level disconnections (no response at all) and network migrations (responses with incorrec   </text>
<query_num> 3903 </query_num>
<text>   ard is removed from a system and an application is attempting to send data without finding an available interface. 8 Related Work Our project goals are very similar to those of Stanford&amp;apos;s MosquitoNet =-=[1]-=- project and Berkeley&amp;apos;s Infopad [20] and Daedalus [14] projects. While these groups have focused more on Mobile IP implementations than dynamic reconfiguration policies, they do address many issues th   </text>
<query_num> 3904 </query_num>
<text>   er lossy wireless networks [2]. Stemm&amp;apos;s thesis mentions ongoing work toward policies for supporting &amp;quot;vertical handoffs&amp;quot; that determine when to pass an IP address from one type of interface to another =-=[26]-=-. Multiple interfaces are not available at any point in time, just the &amp;quot;best&amp;quot; interface which is selected according to a specified policy. Dynamic reconfiguration treats Mobile IP as a form of adaptat   </text>
<query_num> 3905 </query_num>
<text>   ion probably requires user input about the characteristics (ttl, bandwidth, security, etc.) of the stream. 6.4 Streaming UDP For unicast UDP we selected vcr, a real-time distributed MPEG video player =-=[4]-=- that we adapted to use our notification mechanism [11]. This video client resides on MH while the video server runs on a workstation attached to the Ethernet network. A reliable TCP connection is use DSL) focused on describing application assumptions in a manner meaningful to the application, but also transformable to a DSL more meaningful to a lower layer. For example, vcr uses software feedback =-=[4]-=- to monitor the bandwidth between itself and the video server. While the device layer has no understanding of path characteristics, it does understand link-layer characteristics. Allowing the feedback   </text>
<query_num> 3906 </query_num>
<text>   new system calls expose resources to application programs by permitting programmers to register a bound of tolerance for scarce resources such as network bandwidth, disk space, battery power and cost =-=[21]-=-. When the availability of the resource moves outside the tolerance window, the application is informed via an upcall. Physical Media Independence follows the Odyssey policy of registering interest wi   </text>
<query_num> 3907 </query_num>
<text>   sender never has to know where the receiver(s) are. The network figures this out and routes packets accordingly. To examine IP multicast, we use vic, a video player developed at LBL and U.C. Berkeley =-=[16]-=-. We modified vic to rebind the send and receive sockets upon the receipt of an asynchronous mobility notification. This addition of less than 30 lines of code allows vic to continue to receive video   </text>
<query_num> 3908 </query_num>
<text>   the IP address to another available interface, using the available interface&amp;apos;s IP address as a co-located care-of-address. This raises some interesting routing policy issues as described by Cheshire =-=[5]-=-. Normal routing policies using metrics related to hop counts may be inappropriate for selecting network routes where path connectivity, cost, and power management are important concerns. One of Mosqu   </text>
<query_num> 3909 </query_num>
<text>   uch as assumptions about wireless network characteristics and handoff policies. Daedalus has developed a snoop protocol that provides significantly better TCP performance over lossy wireless networks =-=[2]-=-. Stemm&amp;apos;s thesis mentions ongoing work toward policies for supporting &amp;quot;vertical handoffs&amp;quot; that determine when to pass an IP address from one type of interface to another [26]. Multiple interfaces are   </text>
<query_num> 3910 </query_num>
<text>   will perform more efficiently if they cooperate with the operating system. This was reinforced by our experiencesdeveloping vcr. CMU&amp;apos;s Odysseyarchitecture also supports application-aware applications =-=[25]-=-. Odyssey extends the system call interface, allowing applications to collaborate with the operating system. The new system calls expose resources to application programs by permitting programmers to   </text>
<query_num> 3911 </query_num>
<text>   with intermittent connectivity or low bandwidth. POP mailers 8 enable users to retrieve mail during periods of connectivity and read mail after disconnection. Application-level toolkits such as Rover =-=[13]-=- and Wit [29] allow existing applications to be re-engineered for mobility. We believe adaptive applications will perform more efficiently if they cooperate with the operating system. This was reinfor   </text>
</top>
<top>
<paper_num> 40 </paper_num>
<paper_title>   Abstraction Refinement for Quantified Array Assertions.  </paper_title>
<paper_abstract>   We present an abstraction refinement technique for the verification of universally quantified array assertions such as “all elements in the array are sorted”. Our technique can be seamlessly combined with existing software model checking algorithms. We implemented our technique in the ACSAR software model checker and successfully verified quantified array assertions for both text book examples and real-life examples taken from the Linux operating system kernel.  </paper_abstract>
<query_num> 4001 </query_num>
<text>   . Problems arise when this approach is used together with automated abstraction refinement. Standard techniques for extracting predicates from spurious counterexamples such as (weakest) preconditions =-=[1,6]-=- and interpolants [12] are insufficient. The reason is that these techniques do not infer predicates that allow the analysis to perform the necessary widening, i.e., to compute an invariant that state g of predicate abstraction refinement. The method is parameterized by the procedure extract that takes a formula and returns a set of predicates. We use a minimal notational setting (following, e.g., =-=[1]-=-) and ignore details (in particular, the concrete programming language and the use of concrete counterexamples for refinement). These details are irrelevant for our main purpose, which is to introduce  most basic version, the procedure extract0 returns the set of conjuncts. extract0(ϕ1 ∧ . . . ∧ ϕn) = {ϕ1, . . . , ϕn} The rationale for our extension of the procedure extract0 stems from a result in =-=[1]-=-. This result formally evaluates the power of the refinement scheme with the procedure extract0 above (the power as a proof method for program correctness). The evaluation uses an idealized oracle-bas n ϕ (for example, applied to the interval constraint 0 &amp;lt; x ∧ x &amp;lt; 1 it may result in 0 &amp;lt; x). It is the oracle which judiciously chooses what conjuncts to drop and what conjuncts to keep. The result in =-=[1]-=- states that the (realistic) refinement scheme with the procedure extract0 achieves the same power as the idealized oracle-based method with syntactic widening. In our setting, with programs over arra . i∈I extract1(ϕ) = extract0(saturate(ϕ)) Our proof method is the instantiation of the refinement scheme of Figure 5 with the predicate extraction procedure extract1. By the above-mentioned result in =-=[1]-=-, this proof method has the same power as the idealized oracle-based method with semantic widening. I.e., if the unrealistic oracle-based method succeeds in proving a program correct, then so does our   </text>
<query_num> 4002 </query_num>
<text>   These index variables are implicitly universally quantified at each program location. Heuristics for inferring index predicates based on counterexample-guided abstraction refinement are described in =-=[18]-=-. This approach is more general than an approach based on ghost variables because the index variables occuring in the computed invariant are quantified per program location rather than globally for th   </text>
<query_num> 4003 </query_num>
<text>   are still limited due to the missing inbuilt support for arithmetic theories in the underlying theorem provers. Abstract domains that are used in shape analyses such as in three-valued shape analysis =-=[23]-=- and Boolean heaps [22] can express quantified properties of unbounded data structures (namely, shape analysis constraints [16,26] in the case of [23] and their universal fragment in the case of [22])   </text>
<query_num> 4004 </query_num>
<text>   g. Software model checking offers a high degree of automation and has been successfully applied to non-trivial programs such as device drivers. Existing software model checkers (e.g., SLAM [2], BLAST =-=[13]-=-, MAGIC [5], and ARMC [21]) have shown to be suitable for the verification of control-oriented properties, but they are limited when it comes to richer properties that involve data structures. A promi ntation. We integrated our technique into the ACSAR software model checker [24]. The system implements a backward reachability analysis based on predicate abstraction refinement with lazy abstraction =-=[13]-=-. The implementation is done in C++. We performed tests using an X41 Thinkpad laptop with 1 GB of RAM and a 1.6 GHz CPU, running Linux. ACSAR uses the Yices theorem prover [7] for computing the abstra “Iter” refers to the number of refinement steps performed until a safe invariant is computed. Finally, column “Pred” refers to the number of inferred predicates. Our tool is based on lazy abstraction =-=[13]-=-, we therefore provide the average number of predicates per location instead of the total number of predicates. The size of examples varies from 10 to 200 lines of code. Although scalability is an imp   </text>
<query_num> 4005 </query_num>
<text>   have used three-valued shape analysis to verify properties of arrays. However, the abstract domains in these shape analyses are exponentially more succinct than the one used in predicate abstraction =-=[19]-=-. While this additional precision is needed for the analysis of programs manipulating linked data structures, our experience shows that it is not necessarily required for the verification of array rel   </text>
<query_num> 4006 </query_num>
<text>   his approach is used together with automated abstraction refinement. Standard techniques for extracting predicates from spurious counterexamples such as (weakest) preconditions [1,6] and interpolants =-=[12]-=- are insufficient. The reason is that these techniques do not infer predicates that allow the analysis to perform the necessary widening, i.e., to compute an invariant that states properties of unboun   </text>
<query_num> 4007 </query_num>
<text>   instance, consider the precondition pre(π,pc = ℓE) which is given by 0 ≤ k &amp;lt; n ∧ 0 ≤ l &amp;lt; n ∧ l &amp;lt; k ∧ i + 1 &amp;lt; n ∧ a(i) ≤ a(i + 1) ∧ n ≤ i + 2 ∧ a[i := a(i), i := a(i)](k) &amp;lt; a[i := a(i), i := a(i)](l) =-=(2)-=- Note that the updated function a[i := a(i), i := a(i)] is equal to a. Furthermore, it is easy to see that the implication k = i + 1 ∧ l = i ⇒ a(i) &amp;gt; a(i + 1) ∨ a(k) ≥ a(l) is valid. Thus, by contrapo   </text>
<query_num> 4008 </query_num>
<text>   model checking offers a high degree of automation and has been successfully applied to non-trivial programs such as device drivers. Existing software model checkers (e.g., SLAM [2], BLAST [13], MAGIC =-=[5]-=-, and ARMC [21]) have shown to be suitable for the verification of control-oriented properties, but they are limited when it comes to richer properties that involve data structures. A prominent class   </text>
<query_num> 4009 </query_num>
<text>   offers a high degree of automation and has been successfully applied to non-trivial programs such as device drivers. Existing software model checkers (e.g., SLAM [2], BLAST [13], MAGIC [5], and ARMC =-=[21]-=-) have shown to be suitable for the verification of control-oriented properties, but they are limited when it comes to richer properties that involve data structures. A prominent class of such propert   </text>
<query_num> 4010 </query_num>
<text>   rating system kernel. 1 Introduction Among the most promising approaches to the verification of software systems is the combination of predicate abstraction [10] with automated abstraction refinement =-=[6]-=-. This approach is commonly referred to as software model checking. Software model checking offers a high degree of automation and has been successfully applied to non-trivial programs such as device  . Problems arise when this approach is used together with automated abstraction refinement. Standard techniques for extracting predicates from spurious counterexamples such as (weakest) preconditions =-=[1,6]-=- and interpolants [12] are insufficient. The reason is that these techniques do not infer predicates that allow the analysis to perform the necessary widening, i.e., to compute an invariant that state   </text>
<query_num> 4011 </query_num>
<text>   real-life examples taken from the Linux operating system kernel. 1 Introduction Among the most promising approaches to the verification of software systems is the combination of predicate abstraction =-=[10]-=- with automated abstraction refinement [6]. This approach is commonly referred to as software model checking. Software model checking offers a high degree of automation and has been successfully appli   </text>
<query_num> 4012 </query_num>
<text>   rious techniques have been developed that either generalize or extend existing abstract domains (including the predicate abstraction domain) to abstract domains that can express quantified properties =-=[3,11,17,22,25]-=-. However, none of these approaches can be easily ⋆ The first author was supported in part by the German Federal Ministry of Education and Research (BMBF) in the framework of the VerisoftXT project un  There have been various attempts to account for the verification of quantified properties including approaches based on predicate abstraction [8,14,17], firstorder theorem provers [15,20], templates =-=[3,11,25]-=-, and shape analysis [9,22]. Our approach is able to handle all array related examples that have been analyzed in [3,8,9,11,14,17,20]. Some of the examples in [15,25] involve properties withalternati ate abstraction and requires theorem provers that can effectively deal with quantified formulas. Several template-based techniques for generation of quantified invariants have been developed recently =-=[3, 11, 25]-=-. The common idea behind these approaches is that the user provides templates that fix the structure of potential invariants. The analysis then searches for an invariant that instantiates the template   </text>
<query_num> 4013 </query_num>
<text>   rious techniques have been developed that either generalize or extend existing abstract domains (including the predicate abstraction domain) to abstract domains that can express quantified properties =-=[3,11,17,22,25]-=-. However, none of these approaches can be easily ⋆ The first author was supported in part by the German Federal Ministry of Education and Research (BMBF) in the framework of the VerisoftXT project un  There have been various attempts to account for the verification of quantified properties including approaches based on predicate abstraction [8,14,17], firstorder theorem provers [15,20], templates =-=[3,11,25]-=-, and shape analysis [9,22]. Our approach is able to handle all array related examples that have been analyzed in [3,8,9,11,14,17,20]. Some of the examples in [15,25] involve properties withalternati ate abstraction and requires theorem provers that can effectively deal with quantified formulas. Several template-based techniques for generation of quantified invariants have been developed recently =-=[3, 11, 25]-=-. The common idea behind these approaches is that the user provides templates that fix the structure of potential invariants. The analysis then searches for an invariant that instantiates the template [8]. Our tool automatically proves the postcondition specified in their paper. The example array init is a simple array initialization program which is considered in most papers on array verification =-=[3,11,14]-=-. Programs num index and part init were proposed by Gopan etal. [9]. The first one illustrates numeric constraints on the value of array elements. The second one aims to show the handling of multiple   </text>
<query_num> 4014 </query_num>
<text>   rious techniques have been developed that either generalize or extend existing abstract domains (including the predicate abstraction domain) to abstract domains that can express quantified properties =-=[3,11,17,22,25]-=-. However, none of these approaches can be easily ⋆ The first author was supported in part by the German Federal Ministry of Education and Research (BMBF) in the framework of the VerisoftXT project un pts to account for the verification of quantified properties including approaches based on predicate abstraction [8,14,17], firstorder theorem provers [15,20], templates [3,11,25], and shape analysis =-=[9,22]-=-. Our approach is able to handle all array related examples that have been analyzed in [3,8,9,11,14,17,20]. Some of the examples in [15,25] involve properties withalternating universal and existentia o the missing inbuilt support for arithmetic theories in the underlying theorem provers. Abstract domains that are used in shape analyses such as in three-valued shape analysis [23] and Boolean heaps =-=[22]-=- can express quantified properties of unbounded data structures (namely, shape analysis constraints [16,26] in the case of [23] and their universal fragment in the case of [22]). In particular, Gopan   </text>
<query_num> 4015 </query_num>
<text>   rious techniques have been developed that either generalize or extend existing abstract domains (including the predicate abstraction domain) to abstract domains that can express quantified properties =-=[3,11,17,22,25]-=-. However, none of these approaches can be easily ⋆ The first author was supported in part by the German Federal Ministry of Education and Research (BMBF) in the framework of the VerisoftXT project un ting system kernel and the Xen hypervisor. 2 Related Work There have been various attempts to account for the verification of quantified properties including approaches based on predicate abstraction =-=[8,14,17]-=-, firstorder theorem provers [15,20], templates [3,11,25], and shape analysis [9,22]. Our approach is able to handle all array related examples that have been analyzed in [3,8,9,11,14,17,20]. Some of  ch as properties of multidimensional arrays. Our approach does not have these restrictions. Lahiri and Bryant proposed an extension of predicate abstraction to infer universally quantified invariants =-=[17]-=-. Their technique is based on index predicates which are predicates that contain free index variables. These index variables are implicitly universally quantified at each program location. Heuristics   </text>
<query_num> 4016 </query_num>
<text>   s that are used in shape analyses such as in three-valued shape analysis [23] and Boolean heaps [22] can express quantified properties of unbounded data structures (namely, shape analysis constraints =-=[16,26]-=- in the case of [23] and their universal fragment in the case of [22]). In particular, Gopanet al. [9] have used three-valued shape analysis to verify properties of arrays. However, the abstract doma   </text>
<query_num> 4017 </query_num>
<text>   sor. 2 Related Work There have been various attempts to account for the verification of quantified properties including approaches based on predicate abstraction [8,14,17], firstorder theorem provers =-=[15,20]-=-, templates [3,11,25], and shape analysis [9,22]. Our approach is able to handle all array related examples that have been analyzed in [3,8,9,11,14,17,20]. Some of the examples in [15,25] involve prop developed. Another interesting direction is the recent deployment of resolution-based first-order theorem provers for inferring quantified invariants over arrays. Existing approaches include [20] and =-=[15]-=-. McMillan’s approach is based on the computation of quantified interpolants. The idea in [15] is to generate a set of clauses from quantified formulas that encode changes to arrays in the analyzed pr   </text>
<query_num> 4018 </query_num>
<text>   sor. 2 Related Work There have been various attempts to account for the verification of quantified properties including approaches based on predicate abstraction [8,14,17], firstorder theorem provers =-=[15,20]-=-, templates [3,11,25], and shape analysis [9,22]. Our approach is able to handle all array related examples that have been analyzed in [3,8,9,11,14,17,20]. Some of the examples in [15,25] involve prop yet been developed. Another interesting direction is the recent deployment of resolution-based first-order theorem provers for inferring quantified invariants over arrays. Existing approaches include =-=[20]-=- and [15]. McMillan’s approach is based on the computation of quantified interpolants. The idea in [15] is to generate a set of clauses from quantified formulas that encode changes to arrays in the an   </text>
<query_num> 4019 </query_num>
<text>   t with lazy abstraction [13]. The implementation is done in C++. We performed tests using an X41 Thinkpad laptop with 1 GB of RAM and a 1.6 GHz CPU, running Linux. ACSAR uses the Yices theorem prover =-=[7]-=- for computing the abstraction and analyzing spurious counterexamples. The communication with Yices is performed through its API Lite. The input to ACSAR is a C program annotated with assertions to be   </text>
<query_num> 4020 </query_num>
<text>   ting system kernel and the Xen hypervisor. 2 Related Work There have been various attempts to account for the verification of quantified properties including approaches based on predicate abstraction =-=[8,14,17]-=-, firstorder theorem provers [15,20], templates [3,11,25], and shape analysis [9,22]. Our approach is able to handle all array related examples that have been analyzed in [3,8,9,11,14,17,20]. Some of  ing universal and existential quantifiers such as permutation of arrays. These properties are outside the scope of our approach. In the following, we make a more detailed comparison. Range predicates =-=[14]-=- describe properties of unbounded array segments which enables the verification of universally quantified array assertions using predicate abstraction with abstraction refinement. In the refinement ph [8]. Our tool automatically proves the postcondition specified in their paper. The example array init is a simple array initialization program which is considered in most papers on array verification =-=[3,11,14]-=-. Programs num index and part init were proposed by Gopan etal. [9]. The first one illustrates numeric constraints on the value of array elements. The second one aims to show the handling of multiple   </text>
</top>
<top>
<paper_num> 41 </paper_num>
<paper_title>   OMNI: Introducing Social Structure, Norms and Ontologies into Agent Organizations.  </paper_title>
<paper_abstract>   Abstract. In this paper, we propose a framework for modelling agent organizations, Omni, that allows the balance of global organizational requirements with the autonomy of individual agents. It specifies global goals of the system independently from those of the specific agents that populate the system. Both the norms that regulate interaction between agents, as well as the contextual meaning of those interactions are important aspects when specifying the organizational structure. Omni integrates all this aspects in one framework. In order to make design of the multi-agent system manageable, we distinguish three levels of abstraction with increasing implementation detail. All dimensions of Omni have a formal logical semantics, which ensures consistency and possibility of verification of the different aspects of the system. Omni is therefore utmost suitable for the modelling of all types of MAS from open to closed environments. 1  </paper_abstract>
<query_num> 4101 </query_num>
<text>   action objectives can be more or less restrictive, giving the agent enacting the role more or less freedom to decide how to achieve the role objectives and interpret its norms. Following the ideas of =-=[15]-=-, we call such expressions landmarks, that is, conjunctions of logical expressions that are true in a state. Figure 3 shows the informal landmark pattern for the Review Process. Several different spec   </text>
<query_num> 4102 </query_num>
<text>   ations, further than the specification of constraints for scene transition and enactment (the only allowed interactions are those explicitly represented by arcs in scenes). TROPOS. TROPOS methodology =-=[2]-=- spans the overall development process. It distinguishes between an early and a late requirements phase, and between architectural design and detailed design. The models are implemented using Jack Int   </text>
<query_num> 4103 </query_num>
<text>   d requirements of organizations but, on the other hand, cannot assume that participating agents will act according to the needs and expectations of the system design. Concepts as organizational rules =-=[20]-=-, norms and institutions [6], [7], and social structures [13] arise from the idea that the effective engineering of MAS needs high-level, agent-independent concepts and abstractions that explicitly de   </text>
<query_num> 4104 </query_num>
<text>   ed independently from the society to model goals and capabilities of a given entity. Individual agents will join a society as enactors of organizational role(s), as a means to realize their own goals =-=[3]-=-. Social Model. Agent populations of the organizational model are described in the social model (SM) in terms of commitments regulating the enactment of roles by individual agents. Depending of the sp   </text>
<query_num> 4105 </query_num>
<text>   n specified in the OM. Social contracts in Omni are a two-sided agreement between agents and roles instead of a one-sided API description of role enactment, as have been proposed by other researchers =-=[16, 12]-=-. In the extreme, if all is expressed in the role definition and no room is left for negotiation, Omni social contracts can function as these API’s. Interaction Model. Omni provides two levels of spec   </text>
<query_num> 4106 </query_num>
<text>   ns but, on the other hand, cannot assume that participating agents will act according to the needs and expectations of the system design. Concepts as organizational rules [20], norms and institutions =-=[6]-=-, [7], and social structures [13] arise from the idea that the effective engineering of MAS needs high-level, agent-independent concepts and abstractions that explicitly define the organization in whi   </text>
<query_num> 4107 </query_num>
<text>   posed by three dimensions: Normative, Organizational and Ontological that describe different characterizations of the environment. The model is based on two recent MAS models, OperA [5], and HarmonIA =-=[18]-=-. Figure 1 depicts the different modules that compose our proposed framework organized into three levels of abstraction: – the Abstract Level: where the statutes of the organization to be modelled are text of this organization. This function is based on the counts-as operator as developed in [8]. There are several ways in which norms can be abstract and thus several ways to make them more concrete =-=[18]-=-. As an illustration of this process, in the following we describe two kinds of abstractness. Abstract actions: Actions that can be implemented in many ways. For example: “submitting a paper”. The tra   </text>
<query_num> 4108 </query_num>
<text>   s covered by the Omni framework. In the remainder of this section, we briefly discuss how some well known models support the social and normative concepts introduced by the Omni framework. GAIA. Gaia =-=[19]-=- is one of the first agent-oriented software engineering methodologies that explicitly takes into account social concepts. Gaia models are situated in at the Concrete Level of Omni (cf. figure 1). Whi   </text>
<query_num> 4109 </query_num>
<text>   t assume that participating agents will act according to the needs and expectations of the system design. Concepts as organizational rules [20], norms and institutions [6], [7], and social structures =-=[13]-=- arise from the idea that the effective engineering of MAS needs high-level, agent-independent concepts and abstractions that explicitly define the organization in which agents live [21]. These are th   </text>
<query_num> 4110 </query_num>
<text>   t, on the other hand, cannot assume that participating agents will act according to the needs and expectations of the system design. Concepts as organizational rules [20], norms and institutions [6], =-=[7]-=-, and social structures [13] arise from the idea that the effective engineering of MAS needs high-level, agent-independent concepts and abstractions that explicitly define the organization in which ag ted to interaction protocols, and SODA provides no explicit representation for the domain ontology. Furthermore, SODA also does not have a clear and formal semantics. ISLANDER. The ISLANDER formalism =-=[7]-=- provides a formal framework for institutions [14] and has proven to be well-suited to model practical applications (e.g. electronic auction houses). This formalism views an agent-based institution as   </text>
<query_num> 4111 </query_num>
<text>   terested agents, as it does not distinguish between organizational and individual aspects, and does not provide capabilities for agent interpretation of society objectives, norms or plans. SODA. SODA =-=[11]-=-, is actually an extension to Gaia that enables open societies to be designed around suitably-designed coordination media, and social rules to be designed and enforced in terms of coordination rules.   </text>
<query_num> 4112 </query_num>
<text>   the design of MAS can suffice with the idea that agents are mere performers of organizational roles or functions, interacting according to fixed protocols and unable to deviate from expected behavior =-=[21]-=-. As such, agent autonomy is rather limited. In open domains, agents are self-governed autonomous entities that pursue their own individual goals based only on their own beliefs and capabilities [1].  al structures [13] arise from the idea that the effective engineering of MAS needs high-level, agent-independent concepts and abstractions that explicitly define the organization in which agents live =-=[21]-=-. These are the rules and global objectives that govern the activity of an enterprise, group, organization or nation. Given that agents might deviate from expected behavior, open societies need mechan   </text>
</top>
<top>
<paper_num> 42 </paper_num>
<paper_title>   Boiling down information retrieval test collections.  </paper_title>
<paper_abstract>   Constructing large-scale test collections is costly and timeconsuming, and a few relevance assessment methods have been proposed for constructing “minimal ” information retrieval test collections that may still provide reliable experimental results. In contrast to building up such test collections, we take existing test collections constructed through the traditional pooling approach and empirically investigate whether they can be “boiled down. ” More specifically, we report on experiments with test collections from both NT-CIR and TREC to investigate the effect of reducing both the topic set size and the pool depth on the outcome of a statistical significance test between two systems, starting with (approximately) 100 topics and depth-100 pools. We define cost (of manual relevance assessment) as the pool depth multiplied by the topic set size, and error as a system pair whose outcome of statistical significance testing differs from the original result based on the full test collection. Our main findings are: (a) Cost and the number of errors are negatively correlated, and any attempt at substantially reducing cost introduces some errors; (b) The NTCIR-7 IR4QA and the TREC 2004 robust track test collections all yield a comparable and considerable number of errors in response to cost reduction, and this is true despite the fact that the TREC relevance assessments relied on more than twice as many runs as the NTCIR ones; (c) Using 100 topics with depth-30 pools generally yields fewer errors than using 30 topics with depth-100 pools; and (d) Even with depth-100 pools, using fewer than 100 topics results in false alarms, i.e. two systems are declared significantly different even though the full topic set would declare otherwise.  </paper_abstract>
<query_num> 4201 </query_num>
<text>   IR-7 ACLIA IR4QA test collections [15, 17] 2 . Although a few methods for allocating different amount of manual effort to different topics or to different systems were proposed more than a decade ago =-=[6, 29]-=-, they have not been adopted at venues such as TREC due to bias con1 Manual interactive searches can be used to augment the pools, in the hope of improving the coverage of relevant documents [9]. 2 AC  Zobel [29] suggested focussing the assessment effort on topics for which many relevant documents have been found so far, based on his experiments with data from TRECs 3-5. Cormack, Palmer and Clarke =-=[6]-=- suggested focussing the assessment effort on runs that have found many relevant documents so far, based on their experiments with the TREC-6 data. Voorhees [22] expressed concerns for these approache   </text>
<query_num> 4202 </query_num>
<text>   IR-7 ACLIA IR4QA test collections [15, 17] 2 . Although a few methods for allocating different amount of manual effort to different topics or to different systems were proposed more than a decade ago =-=[6, 29]-=-, they have not been adopted at venues such as TREC due to bias con1 Manual interactive searches can be used to augment the pools, in the hope of improving the coverage of relevant documents [9]. 2 AC t for test collection construction to clarify the contributions of the present study. Over a decade ago, some methods for better managing the relevance assessment process at TREC were proposed. Zobel =-=[29]-=- suggested focussing the assessment effort on topics for which many relevant documents have been found so far, based on his experiments with data from TRECs 3-5. Cormack, Palmer and Clarke [6] suggest he swap method and can directly examine the errors for the full topic set size, the present study adopts a variant of this method. In contrast to previous studies which examined the case of 25 topics =-=[19, 23, 29]-=- or 50 topics [25], our experiments directly examine the case of (approximately) 100 topics and fewer with each of our four test collections, including the aforementioned TREC data used by Voorhees [2 rst case scenario for each topic set size. Another related line of research is on the reusability of test collections, often examined by removing one system’s contribution to the pool at a time (e.g. =-=[16, 29]-=-). But reusability is beyond the scope of the present study. There are also approaches to assessment cost reduction that go outside of the basic methodology of pooling followed by relevance assessment   </text>
<query_num> 4203 </query_num>
<text>   [15]. We omit Average Precision (AP), but it is known that AP and Q are very highly correlated. Let L be a relevance level, and let gain(L) denotethe gain value for retrieving an L-relevant document =-=[8]-=-. For our graded-relevance data shown in Table 1, we let gain(L1) = 1, gain(L2) = 2 and gain(L3) = 3. Let I(r) = 1 if the document retrieved at rank r for a particular topic in a given run is L-releva   </text>
<query_num> 4204 </query_num>
<text>   are manually judged 1 . Although this methodology has introduced new problems to IR evaluation such as incompleteness (i.e. not all relevant documents in the document collection have been identified) =-=[1, 3, 14]-=- and biases (i.e. relevance assessments favour some particular classes of systems or retrieved documents) [16, 27], the IR research community still relies heavily on test collections built through poo   </text>
<query_num> 4205 </query_num>
<text>   as using 25 topics with 166 judgments for each for their particular data set. However, these 249 topics originate from different TRECs, and used different runs, pool depths, and even relevance scales =-=[24]-=-. Hence we follow [25] and use only topics 351-450 with the TREC 2004 robust data. Webber, Moffat and Zobel [26] used the TREC 2004 robust data with topics 301-450 (along with other TREC data), and ar tions, i.e. reducing either the topic set size, the pool depth, or both. “IR4QA-CS,” “IR4QA-CT” and “IR4QA-JA” from the NTCIR-7 ACLIA IR4QA task [15] and “ROBUST04OLD” from the TREC 2004 robust track =-=[24]-=-, each with (approximately) 100 topics, are the four data sets that we mainly use. According to Voorhees [25], the 100 topics of ROBUST04OLD “were developed using the same methodology, using mostly th   </text>
<query_num> 4206 </query_num>
<text>   e β makes Q more forgiving to relevant documents found near the bottom of the ranked list [13]. We let β = 1 in this paper. Let l be a document cut-off value. The version of nDCG we use is defined as =-=[2]-=-: nDCG = ∑l r=1 We let l = 1000 in this paper. g(r)/ log(r +1) ∑l r=1 g∗ (r)/ log(r +1) . (2) 4.2 Measuring Statistical Significance There are several ways to test statistical significance given paire   </text>
<query_num> 4207 </query_num>
<text>   ile the TREC million query track has recently reported on successfully obtaining relevance assessments for over 1,800 queries by means of statistical techniques for selecting which documents to judge =-=[4]-=-, the traditional pooling approach still offers several advantages, including its simplicity, its independence to any particular evaluation metric, and its “off-line” nature (i.e. document sets to be  boration across geographically-distributed task organisers (such as those for IR4QA, which covers multiple languages), and later analyses. In contrast to trying to build up “minimal” test collections =-=[4]-=-, we take the existing NTCIR-7 IR4QA and the TREC 2004 robust track test collections and empirically investigate whether they can be “boiled down.” More specifically, we investigate the effect of redu rom the viewpoint of the power of the t-test, that “shallow evaluation of many topics is preferable to deep evaluation of afew.” The analysis of the TREC million query track data by Carterette et al. =-=[4]-=- also supports these suggestions. For reducing assessment cost, it is probably useful to consider not only how many topics to use, but exactly which topics or which combination of topics to use for IR proximately) 100 topics, and also in that we use metrics designed for graded relevance. Note also that the test collection construction methods employed in the aforementioned TREC million query track =-=[4]-=- were designed specifically for evaluation with Average Precision – a binary relevance metric. Reducing the pool size and/or the topic set size has also been tried outside TREC and NTCIR, for example, riments, we investigate the relationship between the number of errors with cost, which we define as the number of topics multiplied by the pool depth. We do not consider the cost of topic development =-=[4]-=-, although we acknowledge that this in practice has a considerable impact on the overall cost of test collection construction. 5. RESULTS AND DISCUSSIONS 5.1 Main Results with 100 Topics Table 2 summa   </text>
<query_num> 4208 </query_num>
<text>   m TRECs 3-5. Cormack, Palmer and Clarke [6] suggested focussing the assessment effort on runs that have found many relevant documents so far, based on their experiments with the TREC-6 data. Voorhees =-=[22]-=- expressed concerns for these approaches from the viewpoint of judgment biases, and neither of these methods was adopted at TREC, as mentioned earlier. Using the TRECs 3-8 ad hoc track and the TRECs 9   </text>
<query_num> 4209 </query_num>
<text>   s (i.e. not all relevant documents in the document collection have been identified) [1, 3, 14] and biases (i.e. relevance assessments favour some particular classes of systems or retrieved documents) =-=[16, 27]-=-, the IR research community still relies heavily on test collections built through pooling. Even with pooling, however, the relevance assessment process for test collection construction is costly and   </text>
<query_num> 4210 </query_num>
<text>   s at ranking systems without relevance assessments by forming “pseudo-qrels” from the pools (e.g. [17, 21]), and those at ranking systems based on a single system, thereby avoiding pooling altogether =-=[18]-=-. It should be noted that there is very little work reported in the literature that uses data from both TREC and NTCIR and compares across these two forums for the purpose of IR evaluation. Exceptions   </text>
<query_num> 4211 </query_num>
<text>   ts these suggestions. For reducing assessment cost, it is probably useful to consider not only how many topics to use, but exactly which topics or which combination of topics to use for IR evaluation =-=[7, 28]-=-. However, it is not yet clear how best to “boil down” an existing topic set in this way. Therefore, in our present study, we reduce the topic set size by first sorting the topics with the performance topic set and multiple evaluation metrics to minimise the risk of jumping to wrong conclusions. One important direction for future research is selecting a minimal topic set for reliable IR evaluation =-=[7, 28]-=-. Which combination of topics to include is probably more important than just how many. Another direction is selecting which documents to judge for relevance without introducing any biases towards run   </text>
</top>
<top>
<paper_num> 43 </paper_num>
<paper_title>   Peer-to-Peer Clustering for Semantic Overlay Network Generation.  </paper_title>
<paper_abstract>   Abstract. The peer-to-peer (P2P) paradigm presents an attractive solution for applications that require scalability, fault-tolerance and autonomy. P2P systems in their basic unstructured form suffer high costs when it comes to efficiently locating content, mainly due to the lack of global knowledge. It is therefore crucial to organize content in an unsupervised way by creating groups of peers with similar content, in order to support efficient search mechanisms. In this paper, we discuss the need for content organization in unstructured P2P networks and present the requirements that must be fulfilled by any approach. We propose P2P clustering as a potential solution to Semantic Overlay Network (SON) generation for organizing P2P networks, and we present our unsupervised approach for decentralized SON creation towards this end. 1  </paper_abstract>
<query_num> 4301 </query_num>
<text>   -sharing being the most prominent, have already proved their merit and are extensively used. Other, more ambitious approaches have been recently proposed in the literature, for example P2P web search =-=[22]-=-. P2P systems are classified in unstructured and structured systems. Unstructured P2P systems do not impose any constraints to the participating peers, other than establishing a limited number of neig  that have provided answers to previous queries. Also they use possession rule overlays, formed by having peers maintaining a list of other peers, with which they index the same item. Parreira et al. =-=[22]-=- propose SONs for P2P web search. Their method is based on rearranging the connections between peers to link friend peers to each other. A similar approach is followed in [4], where the notion of acqu   </text>
<query_num> 4302 </query_num>
<text>   17], where existing classification of peers into taxonomy concepts is exploited to improve query routing. Several other approaches for SON creation over structured P2P systems have also been proposed =-=[2, 9, 15, 19, 30]-=-. Since the focus of this paper is in unstructured P2Pssystems, we confine to merely mention these approaches, but we will not describe their functionality in more detail. 3 Requirements for SON Gener   </text>
<query_num> 4303 </query_num>
<text>   e proposed to generate semantic clusters in super-peer networks. Particular emphasis is put on managing heterogeneous data schemes. Clustering peers based on schemas is also studied in [21], while in =-=[1]-=-, GridVine is presented, which is about SONs based on schema mappings. An approach for distributed document clustering based on k-means is presented in [12]. In [24], an approach for connectivity-base   </text>
<query_num> 4304 </query_num>
<text>   er autonomy, but presents some drawbacks as well, such as high search costs with no guarantees of locating content. In order to solve some of these problems, structured P2P systems have been proposed =-=[25, 27, 29, 34]-=-. These systems are based on distributed hash tables (DHTs) that can support efficient key-based lookups, with predictable logarithmic cost. However, structured P2P systems impose restrictions on data   </text>
<query_num> 4305 </query_num>
<text>   er autonomy, but presents some drawbacks as well, such as high search costs with no guarantees of locating content. In order to solve some of these problems, structured P2P systems have been proposed =-=[25, 27, 29, 34]-=-. These systems are based on distributed hash tables (DHTs) that can support efficient key-based lookups, with predictable logarithmic cost. However, structured P2P systems impose restrictions on data 15], the authors present HSPIR, an approach to index documents in the network hierarchically, in order to support efficient distributed information retrieval. HSPIR uses a structured P2P network (CAN =-=[25]-=-) to organize the nodes, while support for semantics is guaranteed by the use of Latent Semantic Indexing (LSI). A different focus is given in [28], where hierarchical summary indices for content sear   </text>
<query_num> 4306 </query_num>
<text>   es that all peers in the network will be contacted, as long as they are connected to the network. This is essential, otherwise there may exist peers whose content will never be retrieved. We refer to =-=[10]-=- for more details on initiator selection and zone creation. Phase 4: Intra-zone Clustering. After the zones and their initiators have been determined, global clustering starts by collecting feature ve ed to a super-zone, then neighboring super-zones are combined to a larger super-zone, etc. Note that level-i initiators are a subset of the level-(i − 1) initiators. Due to lack of space, we refer to =-=[10]-=- for the actual details of inter-zone clustering. We emphasize that even though parts of this process resemble a centralized approach, this is not the case: initiators are chosen at random and perform   </text>
<query_num> 4307 </query_num>
<text>   ing and fault-tolerant systems, we focus in this paper on unstructured P2P architectures. To improve the efficiency and quality of search in unstructured P2P systems, Semantic Overlay Networks (SONs) =-=[8]-=- have been proposed. The basic idea behind SONs is to group together peers that contain similar contents, so that at search time, queries can be forwarded to only those peers containing content that s ections for neighbor selection. A similar approach utilizing taxonomy-based routing indices is proposed in [23].sThe concept of Semantic Overlay Networks (SONs) is introduced in the P2P literature in =-=[8]-=-. The authors recognize the following challenges when building SONs: a) classification of queries and peers, b) level of granularity for each classification, c) the condition(s) that should be satisfi   </text>
<query_num> 4308 </query_num>
<text>   ing policies are proposed to generate semantic clusters in super-peer networks. Particular emphasis is put on managing heterogeneous data schemes. Clustering peers based on schemas is also studied in =-=[21]-=-, while in [1], GridVine is presented, which is about SONs based on schema mappings. An approach for distributed document clustering based on k-means is presented in [12]. In [24], an approach for con   </text>
<query_num> 4309 </query_num>
<text>   k-means is presented in [12]. In [24], an approach for connectivity-based clustering that creates topological clusters, which can be used as starting points for flooding, is presented. Tempich et al. =-=[31]-=- present an approach where peers join overlay networks based on observations about queries that were successfully answered by other peers. This information is later used to direct searches only to pee   </text>
<query_num> 4310 </query_num>
<text>   lina [7]. Each peer maintains local routing indices that help choosing the most promising directions for neighbor selection. A similar approach utilizing taxonomy-based routing indices is proposed in =-=[23]-=-.sThe concept of Semantic Overlay Networks (SONs) is introduced in the P2P literature in [8]. The authors recognize the following challenges when building SONs: a) classification of queries and peers,   </text>
<query_num> 4311 </query_num>
<text>   lusions of our work and identify future research directions related to SON generation. 2 Related Work Performance and scalability problems in unstructured P2P networks, like Gnutella [14] and Freenet =-=[5]-=-, are well-known [20, 26] and approaches that try to rectify the search performance have been previously proposed [3, 7, 13]. Another study [11], has pointed the problem of free riding in P2P networks   </text>
<query_num> 4312 </query_num>
<text>   me item. Parreira et al. [22] propose SONs for P2P web search. Their method is based on rearranging the connections between peers to link friend peers to each other. A similar approach is followed in =-=[4]-=-, where the notion of acquaintances is proposed. In [32], a P2P architecture where nodes are logically organized into a fixed number of clusters is presented. The main focus of the paper is fairness w   </text>
<query_num> 4313 </query_num>
<text>   one-hop replication and biased random walks. Gkantsidis et al. [13] study hybrid search schemes for unstructured P2P networks, including normalized flooding and random walks with shallow flooding. In =-=[33]-=-, broadcast policies are proposed for improving search and three families of techniques are proposed: a) iterative deepening, b) directed breadth-first search, and c) local indices. Another approach b   </text>
<query_num> 4314 </query_num>
<text>   retrieval. HSPIR uses a structured P2P network (CAN [25]) to organize the nodes, while support for semantics is guaranteed by the use of Latent Semantic Indexing (LSI). A different focus is given in =-=[28]-=-, where hierarchical summary indices for content search are created, following a super-peer approach. Taxonomy-based overlays are studied in [17], where existing classification of peers into taxonomy   </text>
<query_num> 4315 </query_num>
<text>   rs is presented. The main focus of the paper is fairness with respect to the load of individual nodes. The allocation of documents to clusters is done by classification, so it is not unsupervised. In =-=[18]-=-, clustering policies are proposed to generate semantic clusters in super-peer networks. Particular emphasis is put on managing heterogeneous data schemes. Clustering peers based on schemas is also st   </text>
<query_num> 4316 </query_num>
<text>   s is also studied in [21], while in [1], GridVine is presented, which is about SONs based on schema mappings. An approach for distributed document clustering based on k-means is presented in [12]. In =-=[24]-=-, an approach for connectivity-based clustering that creates topological clusters, which can be used as starting points for flooding, is presented. Tempich et al. [31] present an approach where peers   </text>
<query_num> 4317 </query_num>
<text>   s to the participating peers, other than establishing a limited number of neighbors for each peer. The basic search mechanisms are flooding [14] and its variants, like directed or normalized flooding =-=[13]-=-. This pure P2P architecture has several advantages, like resilience to failures and peer autonomy, but presents some drawbacks as well, such as high search costs with no guarantees of locating conten  scalability problems in unstructured P2P networks, like Gnutella [14] and Freenet [5], are well-known [20, 26] and approaches that try to rectify the search performance have been previously proposed =-=[3, 7, 13]-=-. Another study [11], has pointed the problem of free riding in P2P networks and in particular for Gnutella the authors reached the conclusions that: a) nearly 70% of Gnutella users shared no files an  In Gia [3] the combination of several techniques are proposed to effectively improve searches: topology adaptation, hot-spot avoidance, one-hop replication and biased random walks. Gkantsidis et al. =-=[13]-=- study hybrid search schemes for unstructured P2P networks, including normalized flooding and random walks with shallow flooding. In [33], broadcast policies are proposed for improving search and thre   </text>
<query_num> 4318 </query_num>
<text>   scalability problems in unstructured P2P networks, like Gnutella [14] and Freenet [5], are well-known [20, 26] and approaches that try to rectify the search performance have been previously proposed =-=[3, 7, 13]-=-. Another study [11], has pointed the problem of free riding in P2P networks and in particular for Gnutella the authors reached the conclusions that: a) nearly 70% of Gnutella users shared no files an d by the top 1% of sharing hosts. All these results bring out the problems of search using unstructured P2P networks in their basic form and motivate the development of more efficient methods. In Gia =-=[3]-=- the combination of several techniques are proposed to effectively improve searches: topology adaptation, hot-spot avoidance, one-hop replication and biased random walks. Gkantsidis et al. [13] study   </text>
<query_num> 4319 </query_num>
<text>   scalability problems in unstructured P2P networks, like Gnutella [14] and Freenet [5], are well-known [20, 26] and approaches that try to rectify the search performance have been previously proposed =-=[3, 7, 13]-=-. Another study [11], has pointed the problem of free riding in P2P networks and in particular for Gnutella the authors reached the conclusions that: a) nearly 70% of Gnutella users shared no files an osed: a) iterative deepening, b) directed breadth-first search, and c) local indices. Another approach based on directed searches that improves on blind flooding is presented Crespo and Garcia-Molina =-=[7]-=-. Each peer maintains local routing indices that help choosing the most promising directions for neighbor selection. A similar approach utilizing taxonomy-based routing indices is proposed in [23].sTh   </text>
<query_num> 4320 </query_num>
<text>   search. Their method is based on rearranging the connections between peers to link friend peers to each other. A similar approach is followed in [4], where the notion of acquaintances is proposed. In =-=[32]-=-, a P2P architecture where nodes are logically organized into a fixed number of clusters is presented. The main focus of the paper is fairness with respect to the load of individual nodes. The allocat   </text>
<query_num> 4321 </query_num>
<text>   u et al. [16] create groups of peers that are topologically near each other, which they call clusters, and within each cluster specific peers are assigned a set of predefined categories. Cohen et al. =-=[6]-=- propose associative overlays, which are formed by peers that have provided answers to previous queries. Also they use possession rule overlays, formed by having peers maintaining a list of other peer   </text>
</top>
<top>
<paper_num> 44 </paper_num>
<paper_title>   Secure Remote Authentication Using Biometric Data.  </paper_title>
<paper_abstract>   Abstract. Biometric data offer a potential source of high-entropy, secret information that can be used in cryptographic protocols provided two issues are addressed: (1) biometric data are not uniformly distributed; and (2) they are not exactly reproducible. Recent work, most notably that of Dodis, Reyzin, and Smith, has shown how these obstacles may be overcome by allowing some auxiliary public information to be reliably sent from a server to the human user. Subsequent work of Boyen has shown how to extend these techniques, in the random oracle model, to enable unidirectional authentication from the user to the server without the assumption of a reliable communication channel. We show two efficient techniques enabling the use of biometric data to achieve mutual authentication or authenticated key exchange over a completely insecure (i.e., adversarially controlled) channel. In addition to achieving stronger security guarantees than the work of Boyen, we improve upon his solution in a number of other respects: we tolerate a broader class of errors and, in one case, improve upon the parameters of his solution and give a proof of security in the standard model. 1 Using Biometric Data for Secure Authentication Biometric data, as a potential source of high-entropy, secret information, have been suggested as a way to enable strong, cryptographically-secure authentication of human users without requiring them to remember or store traditional cryptographic keys. Before such data can be used in existing cryptographic protocols, however, two issues must be addressed: first, biometric data are not uniformly distributed and hence do not offer provable security guarantees if used  </paper_abstract>
<query_num> 4401 </query_num>
<text>   d to the system. Luckily, it is not difficult to verify that at least some existing protocols (e.g., [1, 17, 18, 11, 16]) satisfy a definition of this sort. 5 (Interestingly, the recent definition of =-=[7]-=- seems to imply the above properties.) Due to lack of space, the formal definition of security required for our application is deferred to the full version. 4.1 A Direct Construction With the above in   </text>
<query_num> 4402 </query_num>
<text>   data offer a potential source of high-entropy, secret information that can be used in cryptographic protocols provided two issues are addressed: (1) biometric data are not uniformly distributed; and =-=(2)-=- they are not exactly reproducible. Recent work, most notably that of Dodis, Reyzin, and Smith, has shown how these obstacles may be overcome by allowing some auxiliary public information to be reliab hus immediately abort in this case. A robust fuzzy extractor is defined similarly. We then show: (1) a construction of a robust sketch in the random oracle model, starting from any secure sketch; and =-=(2)-=- a conversion from any robust sketch to a robust fuzzy extractor; this conversion does not require random oracles. We conclude this section by showing the immediate application of robust fuzzy extract  Rec) is said to be well-formed if it satisfies the conditions of Definition 1 except for the following modifications: (1) Rec may now return either an element in M or the distinguished symbol ⊥; and =-=(2)-=- for all w ′ ∈ M and arbitrary pub ′ , if Rec(w ′ , pub ′ ) �=⊥ then d(w ′ , Rec(w ′ , pub ′ )) ≤ t. ♦ It is straightforward to transform any secure sketch (SS, Rec) into a well-formed secure sketch ( ′ , n, ε, t)-robust sketch over (M, d) if it is a well-formed (m, ⋆, t)-secure sketch and: (1) for all t-bounded distortion ensembles W with H∞(W0) ≥ m and all adversaries A we have Pr[Succ] ≤ ε; and =-=(2)-=- the average min-entropy of W0, conditioned on the entire view of A throughout the above game, is at least m ′′ . 2 ♦ A simpler definition would be to consider only random variables {W0, W1} and to ha i ) ≤ t. Applying Lemma 2 on {W0, Wi} (noticing that pub contains pub ∗ ), followed by Lemma 1 on {Wi, W ′ i }, we have γ ′′ pub,q1 ≥ min i (H∞(Wi | pub, q1)) − log Vol M t ≥ γ ′ pub,q1 − log VolMt . =-=(2)-=- We now consider the type 2 queries made by A. Clearly, the answers to these queries do not affect the conditional min-entropies of W ′ i (since these queries do not include pub ∗ ), so the best proba   </text>
<query_num> 4403 </query_num>
<text>   e correctness when the parties use a shared secret derived from biometric data. Much work has focused on addressing these problems in efforts to develop secure techniques for biometric authentication =-=[8, 15, 19, 14, 22, 21]-=-. Most recently, Dodis, Reyzin, and Smith [9] showed how to use biometric data to securely derive cryptographic keys which could then be used, in particular, for the purposes of authentication. Roughl   </text>
<query_num> 4404 </query_num>
<text>   iometric data. Much work has focused on addressing these problems in efforts to develop secure techniques for biometric authentication [8, 15, 19, 14, 22, 21]. Most recently, Dodis, Reyzin, and Smith =-=[9]-=- showed how to use biometric data to securely derive cryptographic keys which could then be used, in particular, for the purposes of authentication. Roughly speaking (see Section 2 for formal definiti tric features is roughly this order of magnitude (e.g., 173–250 bits for an iris scan [8, 13]). Organization. We review some basic definitions as well as the sketches/fuzzy extractors of Dodis et al. =-=[9]-=- in Section 2. In Section 3 we introduce the notion of robust sketches/fuzzy extractors which are resilient to modification of the public value, and can be used as a generic replacement for sketches/f  not need to specify any particular metric space in our work, as our results build in a generic way on earlier sketch and fuzzy extractor constructions over any such space (e.g., those constructed in =-=[9]-=- for a variety of metrics). A probability space (Ω, P ) is a finite set Ω and a function P : Ω → [0, 1] such that � ω∈Ω P (ω) = 1. A random variable W defined over the probability space (Ω, P ) and ta  a point x ∈ M we define Vol M t (x) def = |{x ′ ∈ M | d(x, x ′ ) ≤ t}| , Vol M t def = max x∈M {VolM t (x)}. The latter is the maximum number of points in any “ball” of radius t in (M, d). Following =-=[9]-=-, for a pair of random variables A and B, we define the minentropy H∞(A) of A, and the average min-entropy of A given B, as H∞(A) = − log(max a Pr[A = a]), H∞(A|B) ¯ def = − log(Expb←B[2 −H∞(A|B=b) ]) etween random variables A and B taking values in the same set D is defined as SD(A, B) def = 1 2 2.1 Secure Sketches and Fuzzy Extractors � d∈D |Pr[A = d] − Pr[B = d]|. We review the definitions from =-=[9]-=- using slightly different terminology. Recall from the introduction that a secure sketch provides a way to recover a shared secret w from any value w ′ which is a “close” approximation of w. More form   </text>
<query_num> 4405 </query_num>
<text>   nts; and (3) the adversary can specify such distributions adaptively at the time the client is added to the system. Luckily, it is not difficult to verify that at least some existing protocols (e.g., =-=[1, 17, 18, 11, 16]-=-) satisfy a definition of this sort. 5 (Interestingly, the recent definition of [7] seems to imply the above properties.) Due to lack of space, the formal definition of security required for our appli  as the partner ID of the server; and (3) the passwords w0 and w ′ are identical. Before discussing the security of this protocol, we need to introduce a 5 In fact, it is already stated explicitly in =-=[17, 11]-=- that the given protocols remain secure even under conditions 1 and 2, and it is not hard to see that they remain secure under condition 3 as well.sslight restriction of the notion of a t-bounded dist  extractors since the “effective key size” will be larger, as we discuss in the next paragraph). To obtain a solution in the standard model which is only slightly less efficient, the PAK protocols of =-=[17, 11, 16]-=- could be used. 7 Note that although these protocols were designed for use with “short” passwords, they can be easily modified to handle “large” passwords without much loss of efficiency; we discuss t  to remember or store these values. The difference is akin to the difference between PAK protocols in a “hybrid” PKI model (where clients store their server’s public key) and PAK protocols (including =-=[17, 11, 16]-=-) in which clients need only remember a short password.s4.2 Comparing Our Two Solutions It is somewhat difficult to compare the security offered by our two solutions (i.e., the one based on robust fuz   </text>
<query_num> 4406 </query_num>
<text>   nts; and (3) the adversary can specify such distributions adaptively at the time the client is added to the system. Luckily, it is not difficult to verify that at least some existing protocols (e.g., =-=[1, 17, 18, 11, 16]-=-) satisfy a definition of this sort. 5 (Interestingly, the recent definition of [7] seems to imply the above properties.) Due to lack of space, the formal definition of security required for our appli  extractors since the “effective key size” will be larger, as we discuss in the next paragraph). To obtain a solution in the standard model which is only slightly less efficient, the PAK protocols of =-=[17, 11, 16]-=- could be used. 7 Note that although these protocols were designed for use with “short” passwords, they can be easily modified to handle “large” passwords without much loss of efficiency; we discuss t  to remember or store these values. The difference is akin to the difference between PAK protocols in a “hybrid” PKI model (where clients store their server’s public key) and PAK protocols (including =-=[17, 11, 16]-=-) in which clients need only remember a short password.s4.2 Comparing Our Two Solutions It is somewhat difficult to compare the security offered by our two solutions (i.e., the one based on robust fuz   </text>
<query_num> 4407 </query_num>
<text>   o apply this idea to obtain a provably secure solution. In particular, we will require the PAK protocol to satisfy a slightly stronger definition of security than that usually considered for PAK (cf. =-=[1, 6, 12]-=-); informally, the PAK protocol should remain “secure” even when: (1) the adversary can dynamically add clients to the system, with (unique) identities chosen by the adversary; (2) the adversary can s   </text>
<query_num> 4408 </query_num>
<text>   on earlier sketch and fuzzy extractor constructions over any such space (e.g., those constructed in [9] for a variety of metrics). A probability space (Ω, P ) is a finite set Ω and a function P : Ω → =-=[0, 1]-=- such that � ω∈Ω P (ω) = 1. A random variable W defined over the probability space (Ω, P ) and taking values in a set M is a function W : Ω → M. If (Ω, P ) is a probability space over which two random ecause d(w0, wi) ≤ t, the user will recover ˆ R = R and thus both user and server will end up using the same key R in the underlying protocol Π. The security of Π ′ with respect to the definitions of =-=[3, 1]-=-, which consider an active adversary who may control all messages sent between the user and the server, follows from the following observations: – If the adversary forwards pub ′ �= pub to at most n d  eavesdropping on multiple executions of the protocol. With the problem laid out in this way, it becomes clear that one possibility is to use a password-only authenticated key exchange (PAK) protocol =-=[4, 1, 6]-=- as the underlying “equality test”. Although the above intuition is appealing, we remark that a number of subtleties arise when trying to apply this idea to obtain a provably secure solution. In parti nts; and (3) the adversary can specify such distributions adaptively at the time the client is added to the system. Luckily, it is not difficult to verify that at least some existing protocols (e.g., =-=[1, 17, 18, 11, 16]-=-) satisfy a definition of this sort. 5 (Interestingly, the recent definition of [7] seems to imply the above properties.) Due to lack of space, the formal definition of security required for our appli pecific instantiations. As noted earlier, a number of PAK protocols satisfying the required definition of security are known. If one is content to work in the random oracle model then the protocol of =-=[1]-=- may be used (note that this still represents an improvement over the solution based on robust fuzzy extractors since the “effective key size” will be larger, as we discuss in the next paragraph). To   </text>
<query_num> 4409 </query_num>
<text>   purpose of using biometric data in the first place: namely, to avoid the need for the user to store any additional cryptographic information — even if that information need not be kept secret. Boyen =-=[5]-=-, inter alia, partially addresses potential adversarial modification of pub (although his work focuses primarily on the orthogonal issue of re-using biometric data with multiple servers, which we do n our results). Our second construction is specific to the settings of remote authentication and key exchange, where it offers some improvements to the generic solution. Compared with the work of Boyen =-=[5]-=-, which was mostly concerned with the re-usability of biometrics, our constructions enjoy the following key advantages: – Both of our solutions tolerate a stronger class of errors. In particular, Boye deling Error in Biometric Applications As error correction is a key motivation for our work, it is necessary to develop some formal model of the types of errors that may occur. In prior work by Boyen =-=[5]-=-, the error in various biometric readings was assumed to be under adversarial control, with the restriction that the adversary could only specify data-independent errors (e.g., constant shifts, permut   </text>
<query_num> 4410 </query_num>
<text>   ybertrust grantsas is, say, as a key for a pseudorandom function. While the problem of nonuniformity can be addressed using a hash function, viewed either as a random oracle [2] or a strong extractor =-=[20]-=-, a second and more difficult problem is that biometric data are not exactly reproducible, as two biometric scans of the same feature are rarely identical. Thus, traditional protocols will not even gu (w ′ , pub) = R. ♦ As shown in [9, Lemma 3.1], it is easy to construct a fuzzy extractor over a metric space (M, d) given any secure sketch defined over the same space, by applying a strong extractor =-=[20]-=- using a random “key” which is then included as part of pub. Starting with an (m, m ′ , t)-secure sketch and with an appropriate choice of extractor, this transformation yields an (m, m ′ − 2 log( 1 δ   </text>
</top>
<top>
<paper_num> 45 </paper_num>
<paper_title>   Pitfalls in the fluid modeling of RTT variations in window-based congestion control.  </paper_title>
<paper_abstract>   Abstract—Deterministic delay differential equation models, where the packet traffic is modeled as a fluid, are widely used to study congestion control algorithms in the Internet. In this paper, we point out some pitfalls in such fluid modeling of window flow control algorithms. Specifically, we argue that the modeling assumptions used to capture the variability in the RTT (due to queue length fluctuations) may play a critical role in our ability to design stable algorithms. We study two scenarios to illustrate the dramatic impact of RTT modeling. We first consider TCP-Reno with RED, and show that assuming that the RTT is a constant (when it is actually time-varying) leads to conservative parameter choices, i.e., the system continues to be stable even with variable RTT. On the other hand, for the recently proposed Stabilized Vegas, we show the following result: while the network can be stabilized under the constant RTT assumption, there is no choice of parameters that would stabilize the system when the RTT variations are taken into account! Interestingly, such problems do not arise if the congestion-control mechanisms at the end-users are rate-based.  </paper_abstract>
<query_num> 4501 </query_num>
<text>   0 7000 8000 Fig. 13. RED for the variable RTT case: window size. R0 = 65 ms, which agrees with the R0 we picked in the Matlab simulations. Simulation 2 uses RED implemented in a virtual queue (REDvq) =-=[29, 4, 30]-=-. We choose the virtual queue’s capacity in this simulation to be equal to the real queue’s capacity in the previous simulation (so that the equilibrium arrival rates at the queues in both systems are   </text>
<query_num> 4502 </query_num>
<text>   0 7000 8000 Fig. 13. RED for the variable RTT case: window size. R0 = 65 ms, which agrees with the R0 we picked in the Matlab simulations. Simulation 2 uses RED implemented in a virtual queue (REDvq) =-=[29, 4, 30]-=-. We choose the virtual queue’s capacity in this simulation to be equal to the real queue’s capacity in the previous simulation (so that the equilibrium arrival rates at the queues in both systems are  since its remains close to zero all the time. So this REDvq simulation corresponds to the above scenario under the constant RTT modeling. Note that this result is not contradictory to the results in =-=[30]-=- since the model in [30] is rate-based and not windowbased. Remark 2. In the REDvq simulation, we note that the RTT is a constant (the propagation delay) plus the queueing b(t) and a(t) 400 350 300 25   </text>
<query_num> 4503 </query_num>
<text>   602-00-2-0542 and AFOSR URI F49620-01-1-0365. Shao Liu, Tamer Bas¸ar and R. Srikant closed-loop system [12, 13, 14]. Since it is hard to analyze the global stability of nonlinear delayed systems (see =-=[15, 16, 17, 18]-=- for some recent results), a common approach is to linearize the nonlinear system [19, 20, 21, 22, 23] and perform local analysis via the multivariable Nyquist Criterion [24, 25]. In the process of th   </text>
<query_num> 4504 </query_num>
<text>   Droptail, most of which work by providing congestion feedback based on the queue lengths at the routers. These mechanisms, called Active Queue Management (AQM) schemes, include RED [1], REM [2], AVQ =-=[3, 4]-=-, BLUE [5], E-RED [6, 7]. In addition, many new source congestion control algorithms have also been proposed, including TCP Vegas [8], Stabilized Vegas [9], HighSpeed TCP [10] and Scalable TCP [11]. A 0 7000 8000 Fig. 13. RED for the variable RTT case: window size. R0 = 65 ms, which agrees with the R0 we picked in the Matlab simulations. Simulation 2 uses RED implemented in a virtual queue (REDvq) =-=[29, 4, 30]-=-. We choose the virtual queue’s capacity in this simulation to be equal to the real queue’s capacity in the previous simulation (so that the equilibrium arrival rates at the queues in both systems are   </text>
<query_num> 4505 </query_num>
<text>   M) schemes, include RED [1], REM [2], AVQ [3, 4], BLUE [5], E-RED [6, 7]. In addition, many new source congestion control algorithms have also been proposed, including TCP Vegas [8], Stabilized Vegas =-=[9]-=-, HighSpeed TCP [10] and Scalable TCP [11]. A widely used approach to study the stability of these algorithms is to choose a fluid model of the congestioncontrolled network, and work with the nonlinea able RTT case is larger than that for the constant RTT. We perform Matlab and ns-2 simulations to support this conclusion. We then study delay-based source algorithms, in particular, Stabilized Vegas =-=[9]-=-. A modification of TCP-Vegas, Stabilized Vegas uses the queueing delay as the congestion measure, and adapts the source window size and thus the source rate based on both the current value and rate o t a constant, which then makes RTT also time varying. We show in this paper that if all delays are taken as time-varying, then “Stabilized Vegas” cannot be stabilized using the technique presented in =-=[9]-=-. In other words, for a general topology network, it is not possible to design the parameters of the algorithm to stabilize the network. II. NETWORK MODELING We have a set of users/routes, R, and a se in addition to the queue length being zero, the fluctuations in the queue length will also be zero, and thus the RTT can be modeled as a constant. V. RTT MODELING IN TCP VEGAS AND STABILIZED VEGAS In =-=[9]-=-, the authors have shown that TCP-Vegas is not stable for multi-link networks, and to overcome this shortcoming they have proposed a modified version, called Stabilized Vegas, which adjusts the source algorithm is unstable under the variable RTT modeling. From this analysis, it seems unlikely if not impossible for a delay-based window-flow algorithm to be stable and scalable using the technique in =-=[9]-=-. A. The analysis of TCP Vegas and Stabilized Vegas in [9] The fluid model of TCP-Vegas is characterized by the following equations ˙Wr = 1 Tr sgn(1 − xr(t)qr(t) ) , (39) αrdr ˙pl = [ 1 (yl − cl)] + p   </text>
<query_num> 4506 </query_num>
<text>   e Queue Management (AQM) schemes, include RED [1], REM [2], AVQ [3, 4], BLUE [5], E-RED [6, 7]. In addition, many new source congestion control algorithms have also been proposed, including TCP Vegas =-=[8]-=-, Stabilized Vegas [9], HighSpeed TCP [10] and Scalable TCP [11]. A widely used approach to study the stability of these algorithms is to choose a fluid model of the congestioncontrolled network, and   </text>
<query_num> 4507 </query_num>
<text>   e been proposed to Droptail, most of which work by providing congestion feedback based on the queue lengths at the routers. These mechanisms, called Active Queue Management (AQM) schemes, include RED =-=[1]-=-, REM [2], AVQ [3, 4], BLUE [5], E-RED [6, 7]. In addition, many new source congestion control algorithms have also been proposed, including TCP Vegas [8], Stabilized Vegas [9], HighSpeed TCP [10] and   </text>
<query_num> 4508 </query_num>
<text>   st of which work by providing congestion feedback based on the queue lengths at the routers. These mechanisms, called Active Queue Management (AQM) schemes, include RED [1], REM [2], AVQ [3, 4], BLUE =-=[5]-=-, E-RED [6, 7]. In addition, many new source congestion control algorithms have also been proposed, including TCP Vegas [8], Stabilized Vegas [9], HighSpeed TCP [10] and Scalable TCP [11]. A widely us   </text>
<query_num> 4509 </query_num>
<text>   stem [12, 13, 14]. Since it is hard to analyze the global stability of nonlinear delayed systems (see [15, 16, 17, 18] for some recent results), a common approach is to linearize the nonlinear system =-=[19, 20, 21, 22, 23]-=- and perform local analysis via the multivariable Nyquist Criterion [24, 25]. In the process of this linearization, a variety of modeling assumptions are made explicitly or implicitly, whose consequen   </text>
<query_num> 4510 </query_num>
<text>   stem [12, 13, 14]. Since it is hard to analyze the global stability of nonlinear delayed systems (see [15, 16, 17, 18] for some recent results), a common approach is to linearize the nonlinear system =-=[19, 20, 21, 22, 23]-=- and perform local analysis via the multivariable Nyquist Criterion [24, 25]. In the process of this linearization, a variety of modeling assumptions are made explicitly or implicitly, whose consequen  8 10 12 x 10 5 −6 Fig. 3. RED algorithm: the window size difference between case 3 and case 1. IV. RTT MODELING FOR RED Several results exist on the stability analysis of RED, such as those in [20], =-=[21]-=-, [28] and [27]. The first three discuss the single link identical sources case and the last one studies the general network topology case. Their common conclusion is that RED becomes unstable as RTT   </text>
<query_num> 4511 </query_num>
<text>   stem [12, 13, 14]. Since it is hard to analyze the global stability of nonlinear delayed systems (see [15, 16, 17, 18] for some recent results), a common approach is to linearize the nonlinear system =-=[19, 20, 21, 22, 23]-=- and perform local analysis via the multivariable Nyquist Criterion [24, 25]. In the process of this linearization, a variety of modeling assumptions are made explicitly or implicitly, whose consequen following another variable, such as xr(t − τ f lr (t)) as constants and to model the independent variable delays (like xr(t) = Wr(t)/Tr(t)) to be variables. This method is widely used (e.g., [27] and =-=[20]-=-), but without much justification. We provide below some evidence supporting this simplification. In the fluid model (1)-(10), delays appear as arguments in (1), (2), (4) and (6). The linearization of  2 4 6 8 10 12 x 10 5 −6 Fig. 3. RED algorithm: the window size difference between case 3 and case 1. IV. RTT MODELING FOR RED Several results exist on the stability analysis of RED, such as those in =-=[20]-=-, [21], [28] and [27]. The first three discuss the single link identical sources case and the last one studies the general network topology case. Their common conclusion is that RED becomes unstable a   </text>
<query_num> 4512 </query_num>
<text>   work by providing congestion feedback based on the queue lengths at the routers. These mechanisms, called Active Queue Management (AQM) schemes, include RED [1], REM [2], AVQ [3, 4], BLUE [5], E-RED =-=[6, 7]-=-. In addition, many new source congestion control algorithms have also been proposed, including TCP Vegas [8], Stabilized Vegas [9], HighSpeed TCP [10] and Scalable TCP [11]. A widely used approach to   </text>
</top>
<top>
<paper_num> 46 </paper_num>
<paper_title>   Achieving range-free localization beyond connectivity.  </paper_title>
<paper_abstract>   Wireless sensor networks have been proposed for many location-dependent applications. In such applications, the requirement of low system cost prohibitsmany range-based methods for sensor node localization; on the other hand, range-free localization depending only on connectivity may underutilize the proximity information embedded in neighborhood sensing. In response to the above limitations, this paper presents a range-free approach to capturing a relative distancebetween1-hopneighboringnodesfromtheirneighborhood orderings that serve as unique high-dimensional location signatures for nodes in the network. With little overhead, the proposed design can be conveniently applied as a transparent supporting layer for many state-of-the-art connectivity-based localization solutions to achieve better positioningaccuracy. Weimplementedourdesignwiththree well-knownlocalizationalgorithmsandtesteditintwotypes ofoutdoortest-bedexperiments: an850-foot-longlinearnetwork with 54 MICAz motes, and a regular2D networkcovering an area of 10000 square feet with 49 motes. Results show that our design helps eliminate estimation ambiguity with sub-hop resolution, and reduces localization errors by as much as 35%. In addition, extensive simulations reveal an interestingfeature of robustnessfor our design underunevenly distributed radio propagation path loss, and confirm itseffectivenessforlarge-scalenetworks. Categories andSubject Descriptors  </paper_abstract>
<query_num> 4601 </query_num>
<text>   Based on whether ranging is conducted at the resourceconstrained sensor nodes, most of the previous work about node localization can be categorized into one of following two classes: (i) range-based =-=[9, 10, 11, 14, 12, 15, 16, 17, 18, 19, 20, 21, 22, 24]-=-, and (ii) range-free localization [25, 26, 27,28, 30,29, 31,32, 34,35, 36,37, 38]. Range-based solutions try to estimate absolute distances or angles among randomly deployed sensor nodes and then app   </text>
<query_num> 4602 </query_num>
<text>   a promising tool for many location-dependent applications [1, 2], e.g., battlefield surveillance [3], environment datacollection[4],eventorhumanlocalization[5,6]. Inaddition,someoftheroutingprotocols=-=[7,8]-=-andnetworkmanagement mechanisms proposed for such networks are built ontheassumptionthatgeographicparametersofeachsensor node are available. Althoughsensor node localization plays an important role in   </text>
<query_num> 4603 </query_num>
<text>   basically two types of methods: (i) range-based localization and (ii) range-free localization. Range-based localization could achievegoodaccuracybut costly for requiring eitherper-noderanginghardware=-=[10,12,14,16,22]-=-orcareful system calibrationand environmentprofiling[9, 11, 40], andthusisnotappropriateforlarge-scaleoutdoorsensornetworks. Range-freeapproacheslocalize nodesbased on simple sensing, such as wireless  Based on whether ranging is conducted at the resourceconstrained sensor nodes, most of the previous work about node localization can be categorized into one of following two classes: (i) range-based =-=[9, 10, 11, 14, 12, 15, 16, 17, 18, 19, 20, 21, 22, 24]-=-, and (ii) range-free localization [25, 26, 27,28, 30,29, 31,32, 34,35, 36,37, 38]. Range-based solutions try to estimate absolute distances or angles among randomly deployed sensor nodes and then app istance or angles among nodes and anchors (also called beacons or reference nodes with pre-known location information). Those methods can be accurate but costly by adding per-node additional hardware =-=[10, 12, 14]-=-, requiring intensive tuning [17] or not suitable for large-scale systems due to their limited effective range [10]. Although some researchhastriedtoutilizeRSS(ReceiveSignalStrength)with noise filteri   </text>
<query_num> 4604 </query_num>
<text>   basically two types of methods: (i) range-based localization and (ii) range-free localization. Range-based localization could achievegoodaccuracybut costly for requiring eitherper-noderanginghardware=-=[10,12,14,16,22]-=-orcareful system calibrationand environmentprofiling[9, 11, 40], andthusisnotappropriateforlarge-scaleoutdoorsensornetworks. Range-freeapproacheslocalize nodesbased on simple sensing, such as wireless  Based on whether ranging is conducted at the resourceconstrained sensor nodes, most of the previous work about node localization can be categorized into one of following two classes: (i) range-based =-=[9, 10, 11, 14, 12, 15, 16, 17, 18, 19, 20, 21, 22, 24]-=-, and (ii) range-free localization [25, 26, 27,28, 30,29, 31,32, 34,35, 36,37, 38]. Range-based solutions try to estimate absolute distances or angles among randomly deployed sensor nodes and then app ulation or multilateration for location calculation. Many range-based methods use techniques such as Time of Arrival (ToA) [13], Time Difference of Arrival (TDoA) (e.g., Cricket [10], AHLos [11], TPS =-=[12]-=-) and Angle of Arrival (AOA) (e.g., APS [22], SpinLoc [18]) to measure distance or angles among nodes and anchors (also called beacons or reference nodes with pre-known location information). Those me   </text>
<query_num> 4605 </query_num>
<text>   e previous work about node localization can be categorized into one of following two classes: (i) range-based [9, 10, 11, 14, 12, 15, 16, 17, 18, 19, 20, 21, 22, 24], and (ii) range-free localization =-=[25, 26, 27,28, 30,29, 31,32, 34,35, 36,37, 38]-=-. Range-based solutions try to estimate absolute distances or angles among randomly deployed sensor nodes and then apply triangulation or multilateration for location calculation. Many range-based met   </text>
<query_num> 4606 </query_num>
<text>   e previous work about node localization can be categorized into one of following two classes: (i) range-based [9, 10, 11, 14, 12, 15, 16, 17, 18, 19, 20, 21, 22, 24], and (ii) range-free localization =-=[25, 26, 27,28, 30,29, 31,32, 34,35, 36,37, 38]-=-. Range-based solutions try to estimate absolute distances or angles among randomly deployed sensor nodes and then apply triangulation or multilateration for location calculation. Many range-based met k localization. In those systems, only a small numberof anchorsare necessary for constructingthe global coordinates,whichsignificantlyreducesthesystemcost. Recent work helps solve problems of “holes” =-=[31, 34]-=- and “complex shapes” [29], contributing to connectivity-based solutions in practical irregular node deploymentwith obstacles. However, we found that localization by means of connectivity alone does n al strengthwasset asthereceiversensitivitythreshold. In the simulation, we modeled the area of interest as a square map without holes where radio can not reach. More complicatedmapscanbeusedwithworks =-=[29,31,34]-=-. Unlessotherwisementioned,Table4liststhedefaultsimulation configurations for the following sections. All the statistics reportedwereaveragedover50runsforhighconfidence. Table 4. DefaultConfigurations   </text>
<query_num> 4607 </query_num>
<text>   em calibrationand environmentprofiling[9, 11, 40], andthusisnotappropriateforlarge-scaleoutdoorsensornetworks. Range-freeapproacheslocalize nodesbased on simple sensing, such as wireless connectivity =-=[26, 27, 29, 32, 33]-=-,anchorproximity[25,28,30],orlocalizationeventsdetection [36, 37]. Amongthese, connectivity-basedsolutions featurealowoverallsystemcost,however,bysacrificinglocalizationaccuracy. Our work is motivated he-art connectivity-based localization algorithms,providingalow-costbuteffectivesolutionforsystem accuracy. We augmented three range-free localization algorithms, i.e., MDS-MAP [26], DV-Hop [27], RPA =-=[33]-=-, with our design, and evaluated the effectiveness of the proposed design in two types of outdoor test-bed systems: an 850-footlong linear network with 54 MICAz motes, and a regular 2dimensional netwo e a good accuracy, however, a high anchor density is required, which is impractical for large-scale systems. Concurrently, wireless connectivity-based protocols such as DV-Hop [27], MDS-MAP [26], RPA =-=[33]-=-, Amorphous [32] and so on, proposed using local neighborhood sensing to build hop-based virtual distances for large-scale sensor network localization. In those systems, only a small numberof anchorsa hops between two nodes to evaluate the physical distance between them. The following threeschemesaretypicalexamples: • MDS-MAP[26]byY. Shang,W. Ruml,et al. • DV-Hop[27] byD.NiculescuandB. Nath. • RPA =-=[33]-=- byC. Savarese,J. M.Rabaey,et al. Here we provide brief descriptions of these algorithms and detailscanbefoundin[26, 27,33]. MDS-MAP [26] first forms a distance matrix A in which thevalueattheithrowan j,respectively. Foreachsensornodeui,itestimates its distance to anchor vi with HopSize ×Hops(ui,vi). Finally, each node’s location is computed with least-square multilaterationonavailableanchors. RPA =-=[33]-=-, which is proposed independently from DVHop, uses a similar mechanism of hop-based distance estimation called Hop-TERRAIN for its first step. Besides, it introduces an iterative refinement step for p   </text>
<query_num> 4608 </query_num>
<text>   i) range-free localization. Range-based localization could achievegoodaccuracybut costly for requiring eitherper-noderanginghardware[10,12,14,16,22]orcareful system calibrationand environmentprofiling=-=[9, 11, 40]-=-, andthusisnotappropriateforlarge-scaleoutdoorsensornetworks. Range-freeapproacheslocalize nodesbased on simple sensing, such as wireless connectivity [26, 27, 29, 32, 33],anchorproximity[25,28,30],or  Based on whether ranging is conducted at the resourceconstrained sensor nodes, most of the previous work about node localization can be categorized into one of following two classes: (i) range-based =-=[9, 10, 11, 14, 12, 15, 16, 17, 18, 19, 20, 21, 22, 24]-=-, and (ii) range-free localization [25, 26, 27,28, 30,29, 31,32, 34,35, 36,37, 38]. Range-based solutions try to estimate absolute distances or angles among randomly deployed sensor nodes and then app   </text>
<query_num> 4609 </query_num>
<text>   large-scaleoutdoorsensornetworks. Range-freeapproacheslocalize nodesbased on simple sensing, such as wireless connectivity [26, 27, 29, 32, 33],anchorproximity[25,28,30],orlocalizationeventsdetection =-=[36, 37]-=-. Amongthese, connectivity-basedsolutions featurealowoverallsystemcost,however,bysacrificinglocalizationaccuracy. Our work is motivated by the finding that localization by means of mere connectivity m e previous work about node localization can be categorized into one of following two classes: (i) range-based [9, 10, 11, 14, 12, 15, 16, 17, 18, 19, 20, 21, 22, 24], and (ii) range-free localization =-=[25, 26, 27,28, 30,29, 31,32, 34,35, 36,37, 38]-=-. Range-based solutions try to estimate absolute distances or angles among randomly deployed sensor nodes and then apply triangulation or multilateration for location calculation. Many range-based met   </text>
<query_num> 4610 </query_num>
<text>   ng[9, 11, 40], andthusisnotappropriateforlarge-scaleoutdoorsensornetworks. Range-freeapproacheslocalize nodesbased on simple sensing, such as wireless connectivity [26, 27, 29, 32, 33],anchorproximity=-=[25,28,30]-=-,orlocalizationeventsdetection [36, 37]. Amongthese, connectivity-basedsolutions featurealowoverallsystemcost,however,bysacrificinglocalizationaccuracy. Our work is motivated by the finding that local e previous work about node localization can be categorized into one of following two classes: (i) range-based [9, 10, 11, 14, 12, 15, 16, 17, 18, 19, 20, 21, 22, 24], and (ii) range-free localization =-=[25, 26, 27,28, 30,29, 31,32, 34,35, 36,37, 38]-=-. Range-based solutions try to estimate absolute distances or angles among randomly deployed sensor nodes and then apply triangulation or multilateration for location calculation. Many range-based met lied many smart ideas for pursuinga low-costdesign. Earlyrange-freesolutionsmade use of the proximity information to anchor nodes. Typical examples are Centroid [25], APIT [28], Concave [35] and Self =-=[30]-=-, in which the high cost of anchors is supposed to beamortizedwithalargenumberoflow-costordinarysenor nodes. To achieve a good accuracy, however, a high anchor density is required, which is impractica   </text>
<query_num> 4611 </query_num>
<text>   t considered a good choice for physical distance estimation in many scenarios because of unknown radio path loss factors, multi-path effects, hardware discrepancies, antenna orientation, and so forth =-=[40, 41, 42]-=-, it does provide some useful distancerelated information beyond indicating connectivity among neighboringnodes. Our experimentalstudy confirmsthat in outdoor open-air scenarios, the radio signal stre   </text>
<query_num> 4612 </query_num>
<text>   t considered a good choice for physical distance estimation in many scenarios because of unknown radio path loss factors, multi-path effects, hardware discrepancies, antenna orientation, and so forth =-=[40, 41, 42]-=-, it does provide some useful distancerelated information beyond indicating connectivity among neighboringnodes. Our experimentalstudy confirmsthat in outdoor open-air scenarios, the radio signal stre alStrength)with noise filtering for distance estimation or for wireless fingerprint matching (e.g., Radar [9], wMDS [43], SpotOn [45], Indoor GPS [46], Sequence [48], Ranking [49]), empirical studies =-=[39, 40, 41]-=- have concluded that unless careful calibrationandenvironmentprofilingcanbeaccomplished,RSS isnotagoodchoiceforaccurateranging. Range-free methods have applied many smart ideas for pursuinga low-costd   </text>
</top>
<top>
<paper_num> 47 </paper_num>
<paper_title>   Cutoff rate optimal binary inputs with imperfect CSI.  </paper_title>
<paper_abstract>   Abstract — We use the cutoff rate to study the optimal binary input distributions for the Rayleigh flat-fading channel with imperfect receiver channel state information (CSI). First, we evaluate the cutoff rate and analyze the optimal binary input as a function of the CSI quality and receiver SNR. Next, we study the limiting distributions – BPSK and On-Off Keying (OOK) – and derive an analytic design rule that allows adaptive switching between these two as the receiver CSI changes. We establish the virtues of a modulation scheme that employs only these limiting distributions, rather than the full spectrum of binary inputs. Finally, we use our results to design an adaptive modulation scheme for Pilot Symbol Assisted Modulation systems. We show that switching between just BPSK and equiprobable-OOK is nearly optimal for moderate to large SNR, and that switching between BPSK and generalized-OOK is nearly optimal for all SNR. Index Terms — Adaptive modulation, cutoff rate, correlated  </paper_abstract>
<query_num> 4701 </query_num>
<text>   6], and studies have been conducted for perfect receiver CSI in [21] (independent fading), [23] and [27] (temporally correlated fading), and for no CSI multiple-input multipleoutput (MIMO) systems in =-=[18]-=-. The cutoff rate is a lower bound on capacity that also provides a bound on the random coding exponent (thereby characterizing the entire rate vs. performance curve) via Pe ≤ 2−N(Ro−R) ,whereRis the   </text>
<query_num> 4702 </query_num>
<text>   ANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 5, NO. 10, OCTOBER 2006 and discussion in Section V. In this paper, we do not consider the design of higher order inputs (which has been done in [2], [19], =-=[10]-=-, and [34] for the capacity) or optimal inputs when the channel is peakconstrained (e.g., see [30], [31]). Here, we use the cutoff rate to (i) study the behavior of the optimal binary inputs when only   </text>
<query_num> 4703 </query_num>
<text>   certain encoding-decoding structures can achieve rates greater than Ro (e.g., turbo coding with iterative decoding), the cutoff rate remains a metric of interest for these systems, as well as others =-=[6]-=-. For example, in sequential decoding, the cutoff rate specifies the largest rate for which decoding complexity remains finite [4]. The cutoff rate often leads to a tractable analysis that often would   </text>
<query_num> 4704 </query_num>
<text>   ches to PSAM have been studied from the perspectives of frequency and timing offset estimation [22], [14], bit-error rate (BER) [8], [11], [7], [12], and the channel capacity or its bounds [17], [3], =-=[24]-=-, [29]. A. System Model We generalize the Rayleigh fading channel of (1) to include temporal correlation. The observation equation is yk = √ Ehksk + nk, where hk ∼ CN � 0,σ2 � h now exhibits temporal   </text>
<query_num> 4705 </query_num>
<text>   ct to each other. However, the marginal statistics of the channel estimate and estimation error are preserved, i.e., �hmT +ℓ ∼ CN � 0, �σ 2 � ℓ and �hmT +ℓ ∼ CN � 0,σ2 h − �σ2 � ℓ . B. Cutoff Rate In =-=[28]-=-, we have previously studied the cutoff rate of a PSAM communications system. Under the assumption of perfect interleaving, and by modification of the proof in [28] to the case of generalized binary i  based on switching between just two inputs6 , captures the optimality of scheme C3 over a wide range of SNR, while requiring a fraction of the complexity. Simulation. We consider two estimators (see =-=[28]-=-,[1]): The causal (1, 0) estimator N = {m}, forwhich ω (1,0) ℓ = � � 2 Rh(ℓ) � � κ 1+κ , and the non-causal (1, 1) estimator N = {m, m+1},forwhich ω (1,1) (κ ℓ = 2 � + κ) |Rh(ℓ)| 2 � 2 + |Rh(T−ℓ)| (κ   </text>
<query_num> 4706 </query_num>
<text>   general, there is no guarantee that PSAM-based approaches are optimal, and PSAM has been shown to be suboptimal when the channel coherence time is small and/or the SNR small from various perspectives =-=[17]-=-, [25], [2], [11]. Nevertheless, the technique is of great practical significance. In addition to providing implementable receiver structures, PSAM facilitates accurate timing and synchronization. PSA ized approaches to PSAM have been studied from the perspectives of frequency and timing offset estimation [22], [14], bit-error rate (BER) [8], [11], [7], [12], and the channel capacity or its bounds =-=[17]-=-, [3], [24], [29]. A. System Model We generalize the Rayleigh fading channel of (1) to include temporal correlation. The observation equation is yk = √ Ehksk + nk, where hk ∼ CN � 0,σ2 � h now exhibit   </text>
<query_num> 4707 </query_num>
<text>   l, there is no guarantee that PSAM-based approaches are optimal, and PSAM has been shown to be suboptimal when the channel coherence time is small and/or the SNR small from various perspectives [17], =-=[25]-=-, [2], [11]. Nevertheless, the technique is of great practical significance. In addition to providing implementable receiver structures, PSAM facilitates accurate timing and synchronization. PSAM has   </text>
<query_num> 4708 </query_num>
<text>   many commercial and Military standards, and optimized approaches to PSAM have been studied from the perspectives of frequency and timing offset estimation [22], [14], bit-error rate (BER) [8], [11], =-=[7]-=-, [12], and the channel capacity or its bounds [17], [3], [24], [29]. A. System Model We generalize the Rayleigh fading channel of (1) to include temporal correlation. The observation equation is yk = utoff rate of this system by RBPSK. This scheme provides a lower C. Adaptive Modulation Scheme Adaptive transmission techniques for fading channels have been well studied in the literature (e.g., see =-=[7]-=-, [16], [9], and the references therein). Typically, a subset of the key transmission parameters – power, rate, modulation shape and size, and bandwidth – is adapted based on some instantaneous measur   </text>
<query_num> 4709 </query_num>
<text>   metrics. The widespread analysis of binary inputs follows from their tractability and optimality, or near optimality, at low SNR under varying amounts of receiver channel state information [2], [34], =-=[19]-=-. We consider reliable rates (i.e., those for which the probability of decoding error can be made arbitrarily small) for communications over a discrete-time Rayleigh flat-fading channel. We assume tha EEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 5, NO. 10, OCTOBER 2006 and discussion in Section V. In this paper, we do not consider the design of higher order inputs (which has been done in [2], =-=[19]-=-, [10], and [34] for the capacity) or optimal inputs when the channel is peakconstrained (e.g., see [30], [31]). Here, we use the cutoff rate to (i) study the behavior of the optimal binary inputs whe   </text>
<query_num> 4710 </query_num>
<text>   n no CSI is available (and as long as the CSI remains below the ⋆ ω threshold). We have restricted our attention to binary inputs. However, even at low SNR, binary inputs are not second-order optimal =-=[33]-=-, and a study of optimal higher order inputs for imperfect CSI may be of interest. APPENDIX A. DERIVATION OF THE CUTOFF RATE, EQUATION (3) Define S � {A, −B} and omit the subscript k for brevity. Star   </text>
<query_num> 4711 </query_num>
<text>   o PSAM have been studied from the perspectives of frequency and timing offset estimation [22], [14], bit-error rate (BER) [8], [11], [7], [12], and the channel capacity or its bounds [17], [3], [24], =-=[29]-=-. A. System Model We generalize the Rayleigh fading channel of (1) to include temporal correlation. The observation equation is yk = √ Ehksk + nk, where hk ∼ CN � 0,σ2 � h now exhibits temporal correl   </text>
</top>
<top>
<paper_num> 48 </paper_num>
<paper_title>   ParseTalk about Textual Ellipsis  </paper_title>
<paper_abstract>   A hybrid methodology for the resolution of text-level ellipsis is presented in this paper. It incorporates conceptual proximity criteria applied to ontologically well-engineered domain knowledge bases and an approach to centering based on functional topic/comment patterns. We state text grammatical predicates for ellipsis and then turn to the procedural aspects of their evaluation within the framework of an actor-based implementation of a lexically distributed parser.  Keywords: text understanding, text parsing, text ellipsis, conceptual distance metric, topic/comment, centering approach 1 Introduction  The work reported in this paper is part of a large-scale text understanding system for knowledge acquisition from German expository texts. Text phenomena are a particularly challenging issue for the design of appropriate parsers, since lacking recognition facilities either result in referentially incohesive or, even worse, invalid text knowledge representations. In a previous paper (Str...  </paper_abstract>
<query_num> 4801 </query_num>
<text>   5 Grammatical Predicates for Textual Ellipsis We here build on the ParseTalk model of dependency grammar, a fully lexicalized grammar theory which employs default inheritance for lexical hierarchies (=-=Broker et al., 1994; Hahn et al., 1994-=-)). The grammar formalism is based on dependency relations between lexical heads and modifiers at the sentence level. The dependency specifications allow a tight integration (not a   </text>
<query_num> 4802 </query_num>
<text>   PUNDIT system (=-=Palmer et al. 1986-=-) which provides a rough implementation solution for a particular domain. This work shares a lot of similarities with our approach (e.g., the use of focus mechanisms (=-=Grosz and Sidner, 1986-=-)). But we consider our proposal superior, since it provides a more general, implementation-independent treatment at the grammar level. The approach reported in this paper also extends our own previou   </text>
<query_num> 4803 </query_num>
<text>   del (=-=Agha and Hewitt, 1987-=-) provides the background for the procedural interpretation of lexicalized grammar specifications, as those given in the previous section, in terms of so-called word actors (=-=Schacht et al., 1994-=-). Word actors combine object-oriented features with concurrency and thus provide a methodologically clean platform for inherently parallel, lexically distributed computations. The model assumes word   </text>
<query_num> 4804 </query_num>
<text>   oach to text ellipsis resolution we propose integrates language-independent conceptual (distance measure) and languagedependent functional (topic/comment) constraints based on the centering approach (=-=Grosz et al., 1995-=-). We explicitly exclude two terminologically related problems from our study. First, we restrict the consideration of ellipses to their textual form, i.e., one that extends over formal sentence bound omment patterns which originate from the (dependency) structure of the preceding sentence. The organizational framework for this type of information is provided by the well-known centeringsmechanism (=-=Grosz et al., 1995-=-). Accordingly, we distinguish each utterance&amp;apos;s backward-looking center (C b (U n )) and its forward-looking-centers (C f (U n )). The ranking imposed on the elements of the C f reflects the assumptio   </text>
<query_num> 4805 </query_num>
<text>   ssue for the design of appropriate parsers, since lacking recognition facilities either result in referentially incohesive or, even worse, invalid text knowledge representations. In a previous paper (=-=Strube and Hahn, 1995-=-), we have already dealt with text-level anaphora (e.g., &amp;quot;Jack owns a car. It cost him $35,000.&amp;quot;), the resolution of which contributes to the construction of (referentially) valid text knowledge bases roper elliptical antecedents. The particular advantage of our approach lies in the integrated treatment of text-level ellipsis within a single coherent grammar format. The anaphora resolution module (=-=Strube and Hahn, 1995-=-) and the ellipsis handler have both been implemented in Smalltalk as part of a comprehensive text parser for German. Besides the information technology domain, experiments with our parser have also b   </text>
<query_num> 4806 </query_num>
<text>   tes for Textual Ellipsis We here build on the ParseTalk model of dependency grammar, a fully lexicalized grammar theory which employs default inheritance for lexical hierarchies (=-=Broker et al., 1994; Hahn et al., 1994-=-)). The grammar formalism is based on dependency relations between lexical heads and modifiers at the sentence level. The dependency specifications allow a tight integration (not a mixture!) of lingui   </text>
</top>
<top>
<paper_num> 49 </paper_num>
<paper_title>   Directed explicit-state model checking in the validation of communication protocols.  </paper_title>
<paper_abstract>   The success of model checking is largely based on its ability to e#ciently locate errors in software designs. If an error is found, a model checker produces a trail that shows how the error state can be reached, which greatly facilitates debugging. However, while current model checkers find error states e#ciently, the counterexamples are often unnecessarily lengthy, which hampers error explanation. This is due to the use of &amp;quot;naive&amp;quot; search algorithms in the state space exploration.  </paper_abstract>
<query_num> 4901 </query_num>
<text>   accepting strongly connected components will be exploited during state space exploration. Precursor Work. Much of the content of this paper is a revision of work that was first published in [13] and =-=[12]-=-. The former paper considers safety property analysis for simple protocols. The latter paper extends this work by providing an approach to validating LTL-specified liveness properties and experimentin   </text>
<query_num> 4902 </query_num>
<text>   actical problems. We are not primarily interested in optimal solutions, which is why we can tolerate non-admissible heuristic estimates when optimistic estimates are not available. In concurrent work =-=[14]-=- we describe an approach to shorten existing error trails using refined state distance metrics as heuristic estimates. This approach has already been implemented in HSF-SPIN. For selected benchmark an   </text>
<query_num> 4903 </query_num>
<text>   asures in [25] are merely stochastic and will not yield optimal solutions, the heuristics we propose are in most cases lower bound estimators and hence allow us to find optimal solutions. Recent work =-=[18]-=- applies heuristic search to the verification of java programs. It is proposing heuristics that increase coverage of the program while disregarding a targeted search for error states. This approach do   </text>
<query_num> 4904 </query_num>
<text>   bined with heuristic search. Our model-checker HSF-SPIN extends the SPIN framework with various heuristic search algorithms to support directed model checking, e.g. A* [19] and iterative deepening A* =-=[24]-=-. Experimental results show that in many cases the number of expanded nodes and the length of the counter-examples are significantly reduced. HSF-SPIN has been applied to the detection of deadlocks, i * has one severe drawback. Once the space resources for storing all expanded and generated nodes are exhausted, no further progress can be made. Therefore, the iterative deepening variant of A*, IDA* =-=[24]-=- for short, counterbalances time for space. It traverses the tree expansion of the problem graph instead of the problem graph itself with a memory requirement that grows linear with the depth of the s   </text>
<query_num> 4905 </query_num>
<text>   efore, in the SPIN validation tool LTL formulae representing a desired property are first negated, and then translated into an equivalent Büchi automaton. In the terminology of the SPIN model checker =-=[21]-=- and its Promela input language this automaton is called a never claim, and we will adopt this terminology throughout this paper. As an example we consider the commonly used response property which st y violation is discovered, the first stack will contain the path into an accepting state, while the second stack will illustrate the cycle through the accepting state. 2.3 The Model Checker SPIN SPIN =-=[21]-=- is a model checking tool implementing the above discussed approach to automata-based model checking. Its input language Promela permits the definition of concurrent processes, called proctypes in Pro   </text>
<query_num> 4906 </query_num>
<text>   irst search procedure they propose incorporates symbolic information based on the Hamming distance of two states. This approach has been improved in [32], where a BDDbased version of the A* algorithm =-=[15]-=- for the µcke model checker [3] is presented. The algorithm outperforms symbolic breadth-first search exploration for two scalable hardware circuits. The heuristic is determined in a static analysis p   </text>
<query_num> 4907 </query_num>
<text>   l set called HSF-SPIN. We provide experimental results from the protocol validation domain using HSF-SPIN. Keywords: Model Checking, Directed Search, Protocol Validation 1 Introduction Model Checking =-=[6]-=- is a formal analysis technique that has been developed to automatically validate 1 functional 1 Within the scope of this paper we use the word “validation” to denote the experimental approach to esta  8. We discuss related work in Section 9 and conclude in Section 10. 2 Automata-based Model Checking In this Section we review the automata theoretic framework for explicit state model checking (c.f. =-=[6]-=-), describe the validation algorithms in use, and present a practical model checker, the SPIN tool set. 2.1 Automata-theoretic Framework Since we model reactive systems with infinite behaviors, the ap  represents the state-space of a Büchi automaton and reports non-emptiness if and only if there exists at least one accepting cycle in the graph. A correctness proof of this algorithm can be found in =-=[6]-=-. Our improvement to the nested depth-first search algorithm is depicted in Figure 15. To prove the correctness of the algorithm we start with some definitions. Definition 1. A Büchi automaton is a fi in fact accepting. The algorithm closes accepting cycles in the first and in the second search. When the algorithm closes cycles in the second search, it acts like the original algorithm. As shown in =-=[6]-=-, cycles closed in the second search are accepting, since the second search is started from accepting states only. On the other hand cycles closed in the first search correspond only to cycles present  which contradicts our assumption. IIb) Suppose now that an accepting cycle exists in a partial-accepting component and that the second search fails to find it. In this case a similar reasoning as in =-=[6]-=- can be done to show that this cannot be happen. Let s be the first accepting state belonging to a partial-accepting component from which the second search starts but fails to find a cycle even though   </text>
<query_num> 4908 </query_num>
<text>   oach of directed stochastic model checking is to derive a stochastic model for search prediction that takes correlations of propositions into consideration in order to direct the search. Further work =-=[26]-=- investigates the combination of partial order reduction techniques with the directed model checking approach of HSF-SPIN. Both theoretically and empirically we show that A* and IDA* can be combined w   </text>
<query_num> 4909 </query_num>
<text>   on and assertion violation based on enabledness of transitions and message exchanges are proposed. The identification of three phases in the verification process is at the heart of work documented in =-=[7]-=-. In exploratory mode the system designer tries to find a first error, in fault finding mode s/he aims at meaningful counterexamples, while in the maintenance mode one does not expect errors at all. F   </text>
<query_num> 4910 </query_num>
<text>   rding a targeted search for error states. This approach does not guarantee optimal counterexamples and accomplishes faster error finding through improved code coverage. The same holds for recent work =-=[31]-=- that proposes the application of genetic algorithms for finding errors in very large state spaces. Genetic algorithms require fitness functions which are a variant of heuristic evaluation functions.   </text>
<query_num> 4911 </query_num>
<text>   s determined in a static analysis prior to the search taking the actual circuit layout and the failure formula into account. The approach to symbolic guided search in CTL model checking documented in =-=[4]-=- applies ‘hints’ to avoid sections of the search space that are difficult to represent for BDDs. This permits splitting the fix-point iteration process used in symbolic exploration into two parts yiel   </text>
<query_num> 4912 </query_num>
<text>   se incorporates symbolic information based on the Hamming distance of two states. This approach has been improved in [32], where a BDDbased version of the A* algorithm [15] for the µcke model checker =-=[3]-=- is presented. The algorithm outperforms symbolic breadth-first search exploration for two scalable hardware circuits. The heuristic is determined in a static analysis prior to the search taking the a   </text>
<query_num> 4913 </query_num>
<text>   state model checking, which is our focus. The need for heuristics is apparent in conformant planning, where the symbolic representation compensates partial knowledge of the current state. The work of =-=[2]-=- trades information gain for exploration time with an estimate preferring belief states with low cardinality. Timed automata call for a finite partitioning of the state space through a symbolic repres   </text>
</top>
<top>
<paper_num> 50 </paper_num>
<paper_title>   The computational complexity of scenario-based agent verification and design.  </paper_title>
<paper_abstract>   We first advocate that the AUML (Agent Unified Modelling Language) notation, even in its new version, is not precise enough to adequately describe protocols. This problem was long identified by Harel and we propose to follow his solution: extend sequence diagrams with a “prechart”, i.e. single out the initiation sequence of the protocol. This new notation keeps readability and intuition, but is also technically adequate and is given a formal semantics. It actually is a form of simple temporal logics, equipped with a game-based semantics, which is appropriate for the modeling of agent-based systems. We then go on to study its complexity. Unsurprisingly, the version with protocol roles is undecidable. The main interesting problem is to synthesize agents that follow the protocol described. Surprisingly, it is undecidable even if we remove roles, alternatives, loops, asynchronous communication, conditions, constraints, negations (already removed in AUML). The complexity of checking whether a society of agents obeys a protocol given in this trivial notation is also surprisingly high: it is in PSPACE-complete, like temporal logic, while we show that this simple language is strongly less expressive than temporal logic. Notations in-between have the expected increase in expressiveness, but no increase in complexity. This justifies the use of a language including alternatives, asynchronous communication and conditions, since it increases expressiveness with no cost in complexity.  </paper_abstract>
<query_num> 5001 </query_num>
<text>   ) with those syntactic constructs. Hence, one can distinguish between provisional and mandatory behavior. Actually, Live Sequence Charts provide engineers with a graphical front-end to Temporal Logic =-=[7,8]-=-. However, this language remains (i) graphical and (ii) scenario-based. In [9], we have shown that LSCs can be smoothly equipped with a game-based semantics, hence making it usable for agent systems s exponentially more succinct than DBA and ACTL det . It is possible to translate LSCs to LTL with only a polynomial blow-up. This improves on previous translations that involved an exponential blow-up =-=[8,7]-=-. Another polynomial translation had already been proposed by Kugler et al. [37]. Yet, their translation applies only to LSCs in which no event appears twice. Proposition 7 (From LSCs to LTL) Any LSC  is the size of the specification. The first player (protagonist) has a winning strategy on this game graph if, and only if, the specification is consistent. This generalizes the approach presented in =-=[7]-=-. ✷ PROOF (Hardness) We encode an alternating PSPACE Turing machine into an LSC, as we did before (see Th. 22). The result will follow from the fact that APSPACE=EXPTIME [50]. The only difference is t   </text>
<query_num> 5002 </query_num>
<text>   Closed Distributed Model Checking) is stated as follows: “Given an LSC L, a list of strategies (fa)a∈Ag, represented by (Aa)a∈Ag, decide whether Out(fAg) |= L. Unfortunately, as usual in verification =-=[43]-=-, distribution makes model checking more complex. Now, the problem becomes PSPACE-complete instead of coNP-complete. Remark that we present, in CDMC, a degenerated problem, for only one scenario is us  (Undecidable) Table 1 summarizes our complexity results. There are two axes along which complexity increases. The distributed version of the problems is always harder than the centralized one, as in =-=[43]-=-, while synthesis is also more complex than model checking, for it adds alternation to the problem [50]. The most interesting part is to investigate what causes such a high complexity. We identify two   </text>
<query_num> 5003 </query_num>
<text>   and is intended to be an alternative to Statecharts [15], which are used in Electronic Institutions [16]. Walton proposes a translation of MAP to Promela, the input language of the SPIN model-checker =-=[17]-=-, which allows one to check MAP models against LTL formulae. Their work is more pragmatic than ours, but could be coupled with our approach. Here, we propose to use a graphical, user-friendly, languag   </text>
<query_num> 5004 </query_num>
<text>   ction (w|Σ ′) is the operation that removes from w all symbols that are not in Σ ′ . 4sWe assume here, for simplicity, that communication is instantaneous. (In contrast, some undecidability proofs of =-=[25]-=- require the more complex FIFO communication). From agents’ behavior emerge sequences of events, which we can observe. Hence, we identify behaviors and sequences of events. 2.1 Live Sequence Charts Li   </text>
<query_num> 5005 </query_num>
<text>   e protocols. MAP could be such an implementation language (although designed as a specification language). Another possibility would be to use agent-oriented programming languages, such as AgentSpeak =-=[18]-=-, 3APL [19], ConGoLog [20], for instance. There is also some tool support for the verification of AgentSpeak programs [21]. First, Agent Speak programs are made finite, then they are translated to Pro   </text>
<query_num> 5006 </query_num>
<text>   efs, desires and intentions, and require information about these facts as well. MABLE has been extended to support the verification that MABLE programs comply with a protocol description given in ACL =-=[24]-=-. These extensions entail an unsurprising high complexity. The goal of our paper is to show that even a very basic language will have a high complexity, we have thus purposedly excluded such features.   </text>
<query_num> 5007 </query_num>
<text>   gents from Sys have to be implemented. Sys implementation will be deployed among Env agents that provide thus the model-time context of the specification. Agents act according to plans, or strategies =-=[13,40]-=-. Remember that we abstract away from agent’s actions and focus on coordination instead. Thus, our abstract view of agent a is a strategy f : Σ ∗ → 2 Σs a. A strategy tells the agent that actions f(w)   </text>
<query_num> 5008 </query_num>
<text>   initiation sequence. However, these different meanings are implicit: there are no syntactic constructs carrying this information. For this reason, Damm and Harel have introduced Live Sequence Charts =-=[6]-=-. This language extends exactly MSCs (and Interaction Diagrams) with those syntactic constructs. Hence, one can distinguish between provisional and mandatory behavior. Actually, Live Sequence Charts p omplex FIFO communication). From agents’ behavior emerge sequences of events, which we can observe. Hence, we identify behaviors and sequences of events. 2.1 Live Sequence Charts Live Sequence Charts =-=[6]-=- are based on Message Sequence Charts (MSCs) [2]. They present the various interactions of agents. Every agent owns a “life-line”, labeled by its name, e.g. “ui”, “cm”, “client1” in Fig. 4. Interactio ever a meeting is called, all members are called for a vote by the chair” or if it is a possible execution that has been singled out. Fig. 1. Interaction Diagram (UN Vote Procedure) LSCs clarify this =-=[6]-=-. They add syntactic constructs to MSCs to state explicitly whether the diagram is a mere example (existential scenarios) or constrains all behaviors of the future system (universal scenarios). The fo   </text>
<query_num> 5009 </query_num>
<text>   is introduced. When the task description is PSPACE-complete, verification is PSPACE-complete as well. Walton presents a lightweight language for describing agent dialogues, named Multi-Agent Protocol =-=[14]-=-. This language is based on the theory of Speech Act and is intended to be an alternative to Statecharts [15], which are used in Electronic Institutions [16]. Walton proposes a translation of MAP to P   </text>
<query_num> 5010 </query_num>
<text>   lar languages and a very restricted sub-class indeed. Live Sequence Charts are strictly less expressive than Deterministic Büchi Automata (DBA) [34] and ACTL det , the common fragment of LTL and ACTL =-=[35]-=-, as we showed in [36]. In section 4, we will prove that LSCs are exponentially more succinct than DBA and ACTL det . It is possible to translate LSCs to LTL with only a polynomial blow-up. This impro ually, it is not even possible to translate LSCs to NBA recognizing either the language of the specification or its complement without this blow-up. It follows from this fact and from the theorems in =-=[35]-=- that turning LSCs to equivalent ACTL det formulae also involves an exponential blow-up. Indeed, for every ACTL det formula, there is a nondeterministic Büchi automaton recognizing their complement, w   </text>
<query_num> 5011 </query_num>
<text>   lity would be to use agent-oriented programming languages, such as AgentSpeak [18], 3APL [19], ConGoLog [20], for instance. There is also some tool support for the verification of AgentSpeak programs =-=[21]-=-. First, Agent Speak programs are made finite, then they are translated to Promela. Bordini et al. also present a logic based on BDI (Beliefs-Desires-Intentions) for specifying the requirements that A   </text>
<query_num> 5012 </query_num>
<text>   presents a lightweight language for describing agent dialogues, named Multi-Agent Protocol [14]. This language is based on the theory of Speech Act and is intended to be an alternative to Statecharts =-=[15]-=-, which are used in Electronic Institutions [16]. Walton proposes a translation of MAP to Promela, the input language of the SPIN model-checker [17], which allows one to check MAP models against LTL f   </text>
<query_num> 5013 </query_num>
<text>   restricted thanks to a “restricts” clause. This provides the scenario with a scope (alphabet). Fig. 2. Symbolic LSC (UN Proposal Scenario) Harel and Marelly have extended LSC with symbolic instances =-=[27]-=-. This construct allows one to talk about the roles played by agents in protocols. The basic idea is to introduce first-order variables, that are placeholders for agents. These variables may be quanti havior of unbounded families of agents. We introduce roles in our approach. In logical terms, Symbolic LSCs are to LSCs what first-order logic is to propositional logic. We follow as much as possible =-=[27]-=-, even though their solution has been tuned for animation, and its formalization might not sound as clean as it could be. Role is a set of roles. A population is a partial function, with finite domain op, θ ∪ {x ↦→ a}, γ |= Q. A Symbolic uLSC is a pair �(P, M) such that (1) P is a Σ ′ (V )-LPO. Thus, all variables in V are free in P . We do not allow quantifiers in the prechart, as is also done by =-=[27]-=-. (2) M is a Σ ′′ (V ′ )-LPO, with Σ ′′ (V ′ ) ⊇ Σ ′ (V ), in which the sole free variables are V . Symbolic LSCs are interpreted against populations and infinite words γ ∈ Σ ω . An interpretation sat   </text>
<query_num> 5014 </query_num>
<text>   s heuristics [11], so as to alleviate the work of agent programmers in well-known cases, leaving them the more creative parts. The work of Wooldridge and colleagues is related to what we present here =-=[12,13]-=-. They study the computational complexity of agent verification and agent design, with respect to task descriptions. A task description is represented as a subset of all runs, i.e. a language, which a gents from Sys have to be implemented. Sys implementation will be deployed among Env agents that provide thus the model-time context of the specification. Agents act according to plans, or strategies =-=[13,40]-=-. Remember that we abstract away from agent’s actions and focus on coordination instead. Thus, our abstract view of agent a is a strategy f : Σ ∗ → 2 Σs a. A strategy tells the agent that actions f(w)   </text>
<query_num> 5015 </query_num>
<text>   such an implementation language (although designed as a specification language). Another possibility would be to use agent-oriented programming languages, such as AgentSpeak [18], 3APL [19], ConGoLog =-=[20]-=-, for instance. There is also some tool support for the verification of AgentSpeak programs [21]. First, Agent Speak programs are made finite, then they are translated to Promela. Bordini et al. also   </text>
<query_num> 5016 </query_num>
<text>   such as the weather synchronization logic of NASA’s Center TRACON Automation System (CTAS) [28], a radio-based train system [29], virtual wrappers for PCI bus [30] and some part of the C elegans worm =-=[31]-=-. Examples displayed in Fig. 4, 5, 6 and 7 are based on the CTAS system. This system aims at synchronizing various clients that make use of weather data reports. When new data is available, a certain   </text>
<query_num> 5017 </query_num>
<text>   uage. Wooldridge et al. present another language for agent programming, called MABLE, which is based on classical imperative languages, enriched with features from agent-oriented programming paradigm =-=[22]-=-. Essentially, it is possible to use a belief-desire-intention logic instead of classical boolean expressions. if-then-else constructs are modified into if-then-else-unsure constructs, to cope with th   </text>
</top>
<top>
<paper_num> 51 </paper_num>
<paper_title>   Acknowledgements.  </paper_title>
<paper_abstract>   ii First and foremost, I’d like to thank my advisor John Preskill, for immensely interest-ing discussions and for the patience of a Buddhist in awaiting the completion of this manuscript. In addition, I’d like to thank Michael Nielsen for especially useful discussions on many topics. I’d also like to thank Dorit Aharonov and Amalavoyal Chari for helpful discussions regarding graph isomorphism; and Charlene Ahn, John Cortese, Jim Harring-ton, Alexei Kitaev, Andrew Landahl, Barry Simon, and Clint White for their comments and discussions regarding causality and causal operations. Also deserving of recognition are all those who, by encouraging comments or pointed questions, helped to convince me to finally complete this thesis. These are too numerous to list here, but special mention should be made of my parents, who mostly restricted themselves to encouraging comments, and of Matt Matuszewski, who felt free to ask pointed questions. Finally, I would like to thank the management of Toyon Research Corporation for their patience and flexibility in allowing me the time to complete and defend my dissertation. iii  </paper_abstract>
<query_num> 5101 </query_num>
<text>   N = exp(log N). The number-field sieve [72], the classical algorithm whose asymptotic complexity is the lowest known, requires time O � exp(c(log N) 1/3 (log log N) 2/3 ) � . Two years later, Grover =-=[55, 56]-=- demonstrated another result, less spectacular in its speed increase but far more general, for finding roots of arbitrary black-box functions in less time (measured by the number of function evaluatio cult to implement.) (v) If a search doesn’t seem to be reducible to a period-finding problem, all is not lost. Given an arbitrary function f : {0, . . . , N − 1} → {0, 1}, the Grover search algorithm =-=[55]-=- (the presentation here follows Preskill [88, §6.4] in a formalism originally due to Brassard and Høyer [24]) allows determination of a solution i to f(i) = 1, if one exists, in O( √ N) evaluations of   </text>
<query_num> 5102 </query_num>
<text>   as slightly tightened these bounds but has not sufficed to answer the question of whether BQP contains problems which are classically difficult to solve. Bennett et al. [17], Simon [93], and Vazirani =-=[98]-=- have proved several oracle-relative results, giving both upper and lower (relative) bounds. These indicate that BQP, while showing occasional surprising strength, cannot provide superpolynomial reduc   </text>
<query_num> 5103 </query_num>
<text>   d operation sets, has slightly tightened these bounds but has not sufficed to answer the question of whether BQP contains problems which are classically difficult to solve. Bennett et al. [17], Simon =-=[93]-=-, and Vazirani [98] have proved several oracle-relative results, giving both upper and lower (relative) bounds. These indicate that BQP, while showing occasional surprising strength, cannot provide su   </text>
<query_num> 5104 </query_num>
<text>   efine the Pauli group as P ⊗n d . (This group is commonly used in analyses of quantum error-correcting codes on d-dimensional Hilbert spaces. There is a large body of work on this subject; see, e.g., =-=[53, 64, 90]-=- for more details.) There is also extensive literature describing the theory of stabilizer codes; for example, see Gottesman’s readable papers on binary [52] and nonbinary [53] stabilizer codes. 18 Th   </text>
<query_num> 5105 </query_num>
<text>   ion that quantum computers might be exponentially more powerful than classical ones, at least for some problems. The most famous result in the theory of quantum algorithms occurred in 1994, when Shor =-=[91, 92]-=- proved that a quantum computer could factor integers in “polynomial” time, 6 in contrast to the classical case in which no polynomial-time algorithm is known. (The best classical factoring algorithms  is whether the strong form of the Church-Turing thesis is true for quantum as well as classical computers. At this time there are some indications that it may not be true; Shor’s factoring algorithm =-=[91, 92]-=-, the DeutschJozsa algorithm [37], and Feynman’s early idea of simulating quantum systems [46, 47] all provide better-than-polynomial improvements in time complexity over the best classical algorithms  the list of factors multiplying to N, together with proofs that each is prime, provide a succinct certificate for N.) Factoring is known to have a polynomial-time quantum algorithm (Shor’s algorithm =-=[92]-=-); whether the same is true for GI is currently an open question. GI is one of a number of related isomorphism problems, many of which are easily reducible to each other [67]. GA, the problem of deter fices to determine r with high probability. A detailed consideration of PeriodFinding, including the distribution of the measured denominators r ′ , is somewhat involved; see Shor’s pioneering papers =-=[91, 92]-=- for the full details. 10 ) The QFFT is quite fast: it can be implemented in time O � (log N) 2� , compared with O(N log N) for an implementation of the classical FFT. 11 The reason the QFFT is so fas   </text>
<query_num> 5106 </query_num>
<text>   ion that quantum computers might be exponentially more powerful than classical ones, at least for some problems. The most famous result in the theory of quantum algorithms occurred in 1994, when Shor =-=[91, 92]-=- proved that a quantum computer could factor integers in “polynomial” time, 6 in contrast to the classical case in which no polynomial-time algorithm is known. (The best classical factoring algorithms  is whether the strong form of the Church-Turing thesis is true for quantum as well as classical computers. At this time there are some indications that it may not be true; Shor’s factoring algorithm =-=[91, 92]-=-, the DeutschJozsa algorithm [37], and Feynman’s early idea of simulating quantum systems [46, 47] all provide better-than-polynomial improvements in time complexity over the best classical algorithms fices to determine r with high probability. A detailed consideration of PeriodFinding, including the distribution of the measured denominators r ′ , is somewhat involved; see Shor’s pioneering papers =-=[91, 92]-=- for the full details. 10 ) The QFFT is quite fast: it can be implemented in time O � (log N) 2� , compared with O(N log N) for an implementation of the classical FFT. 11 The reason the QFFT is so fas   </text>
<query_num> 5107 </query_num>
<text>   it is a useful notation.s21 HA H � H HB H ❣ H ≡ H A H B Figure 1.1: An interesting quantum circuit equivalence. above and below: BPP ⊆ BQP ⊆ P #P . 27 Other work, such as the paper by Adleman et al. =-=[2]-=- considering quantum computers with restricted operation sets, has slightly tightened these bounds but has not sufficed to answer the question of whether BQP contains problems which are classically di   </text>
<query_num> 5108 </query_num>
<text>   ity, but in some cases the classical algorithms do not seem to reduce easily to the Grover black-box search form. Similarly, some harder problems (not in NP) do not seem amenable to Grover reductions =-=[84, 44]-=-. 10 A complete physical model should thus use a complete unified field theory as basis for the computational model; this is beyond the scope of this paper and unlikely to be relevant to the functioni   </text>
<query_num> 5109 </query_num>
<text>   lian group A, a finite set X acted on by A, and x ∈ X, find the stabilizer Ax = {g ∈ A : g(x) = x} of x in A, e.g., by giving a basis for Ax), which is also efficiently solvable on a quantum computer =-=[61, 62]-=-. Unfortunately, though, the nonabelian stabilizer problem, to which GI easily reduces, has no known polynomial-time algorithm. (An algorithm for GI based on this reduction, developed by Ettinger and   </text>
<query_num> 5110 </query_num>
<text>   m with � iterations. The initial projectors on the states (2.5) takes O � |〈α |ω〉| −1 � = O � 1 √N state may also be some state other than |α〉, as analyzed by Brassard et al. [25] and by Biron et al. =-=[23, 22]-=-; clever choice of initial state, if partial information is known about the solution set, can reduce the time required to find a solution. There are a number of possible generalizations to the form of 13 ; the rotation angles will generally be incommensurate, leading to complicated trajectories for generic input states. 14 Another generalization, studied by Brassard et al. [25] and by Biron et al. =-=[23, 22]-=-, is the removal of the requirement that the initial state be a uniform superposition over the solution space. As a prelude to the discussion of GI, I would like to discuss one other primitive which d   </text>
<query_num> 5111 </query_num>
<text>   of a classical computer is a subject of some debate. A few interesting discussions can be found in the manifesto of the Quantum Computation Collective [30] and in articles by Steane [95] and by Jozsa =-=[59]-=- and Ekert and Jozsa [42], but many other opinions are there for the asking.sof a cyclic group Zn) by 32 |x〉 QFFT −→ 1 N−1 √ N i=0 � e 2πixy/N |y〉 , (2.4) can be used to sample the eigenvalues e iφn o   </text>
<query_num> 5112 </query_num>
<text>   of all stabilizer operators. ¯ H is just the subspace of H with all eigenvalues +1. The theory of abelian stabilizer codes has proved useful in the development of quantum error-correcting codes (see =-=[52, 90, 64, 65, 94]-=- for several examples) and is well developed (Preskill [88, Ch. 7] and Nielsen and Chuang [83, Ch. 10] have two introductory treatments). Various useful operations, including stabilizer measurements ( efine the Pauli group as P ⊗n d . (This group is commonly used in analyses of quantum error-correcting codes on d-dimensional Hilbert spaces. There is a large body of work on this subject; see, e.g., =-=[53, 64, 90]-=- for more details.) There is also extensive literature describing the theory of stabilizer codes; for example, see Gottesman’s readable papers on binary [52] and nonbinary [53] stabilizer codes. 18 Th   </text>
<query_num> 5113 </query_num>
<text>   t isomorphic. Classically the same problem exists: Succinct, verifiable classical certificates are known for GI but not for GNI (the problem of proving that two graphs are not isomorphic). 23 Watrous =-=[101]-=- has independently used many of the ideas above to find a verifiable certificate for GNM (Group Non-Membership), the problem of proving that a particular element g of some finite group G is not in the   </text>
<query_num> 5114 </query_num>
<text>   t. Given an arbitrary function f : {0, . . . , N − 1} → {0, 1}, the Grover search algorithm [55] (the presentation here follows Preskill [88, §6.4] in a formalism originally due to Brassard and Høyer =-=[24]-=-) allows determination of a solution i to f(i) = 1, if one exists, in O( √ N) evaluations of f, a quadratic improvement over the O(N) evaluations of f required classically. The Grover search can be th   </text>
<query_num> 5115 </query_num>
<text>   th restricted operation sets, has slightly tightened these bounds but has not sufficed to answer the question of whether BQP contains problems which are classically difficult to solve. Bennett et al. =-=[17]-=-, Simon [93], and Vazirani [98] have proved several oracle-relative results, giving both upper and lower (relative) bounds. These indicate that BQP, while showing occasional surprising strength, canno   </text>
<query_num> 5116 </query_num>
<text>   whether two graphs are isomorphic to one another. This problem is of interest in classical complexity theory because it, like Factoring, is believed to be a relatively easy problem in NP (see, e.g., =-=[66]-=-) which is, however, not known to lie in BPP. We present several potentially-fruitful quantum approaches to a solution of the problem.s2.1.1 The problem 24 We begin with definitions of a few terms. An   </text>
</top>
<top>
<paper_num> 52 </paper_num>
<paper_title>   High-performance bilingual text alignment using statistical and dictionary information.  </paper_title>
<paper_abstract>   This paper describes an accurate and  robust text alignment system for structurally  different languages. Among  structurally different languages such as  Japanese and English, there is a limitation  on the amount of word correspondences  that can be statistically acquired. The  proposed method makes use of two kinds  of word correspondences in aligning bilingual  texts. One is a bilingual dictionary of  general use. The other is the word correspondences  that are statistically acquired  in the alignment process. Our method  gradually determines sentence pairs (anchors)  that correspond to each other by relaxing  parameters. The method, by combining  two kinds of word correspondences,  achieves adequate word correspondences  for complete alignment. As a result, texts  of various length and of various genres  in structurally different languages can be  aligned with high precision. Experimental  results show our system outperforms  conventional methods for various kinds of  Japanese-English texts.  </paper_abstract>
<query_num> 5201 </query_num>
<text>   alignment. Thus, the problem cannot be addressed as long as the method relies only on statistics. Other methods in the lexicon-based approach embed lexical knowledge into stochastic models (=-=Wu, 1994;Chen, 1993-=-), but these methods were tested using rigid translations. To tackle the problem, we describe in this paper a text alignment system that uses both statistics and bilingual dictionaries at the same tim   </text>
<query_num> 5202 </query_num>
<text>   apanese-English texts. 1 Introduction Corpus-based approaches based on bilingual texts are promising for various applications(i.e., lexical knowledge extraction (=-=Kupiec, 1993; Matsumoto et al., 1993; Smadja et al., 1996; Dagan and Church, 1994; Kumano and Hirakawa, 1994; Haruno et al., 1996-=-), machine translation (=-=Brown and others, 1993; Sato and Nagao, 1990; Kaji et al., 1992-=-) and information retrieval (=-=Sato, 1992-=-)) r noisy corpora and do not require any information source, aligned sentences are necessary for higher level applications such as well-grained translation template acquisition (=-=Matsumoto et as., 1993; Smadja et al., 1996; Haruno et al., 1996-=-) and example-based translation (=-=Sato and Nagao, 1990-=-). Our method performs accurate alignment for such use by combining the detailed word correspondences: statistically acquired   </text>
<query_num> 5203 </query_num>
<text>   asily implemented with regular expressions. 2We use in this phase the JUMAN morphological analyzing system (=-=Kurohashi et al., 1994-=-) for tagging Japanese texts and Brill&amp;apos;s transformation-based tagger (=-=Brill, 1992; Brill, 1994-=-) for tagging English texts (=-=JUMAN: ftp://ftp.aist-nara.ac.jp/pub/nlp/tools/juman/ Brih ftp://ftp.cs.jhu.edu/pub/brill-=-). We would like to thank all people concerned for providing us with   </text>
<query_num> 5204 </query_num>
<text>   ds have been proposed to align bilingual corpora. One of the major approaches is based on the statistics of simple features such as sentence length in words (=-=Brown and others, 1991-=-) or in characters (=-=Gale and Church, 1993-=-). These techniques are widely used because they can be imple131 mented in an efficient and simple way through dynamic programing. However, their main targets are rigid translations that are almost li se the data. 135 categories of matches by manual alignment and indicate the difficulty of the task. Our evaluation focuses on much smaller texts than those used in other study(=-=Brown and others, 1993;Gale and Church, 1993; Wu, 1994; Fung, 1995; Kay and Roscheisen, 1993-=-) because our main targets are well-separated articles. However, our method will work on larger and noisy sets too, by using word anchors rather than us   </text>
<query_num> 5205 </query_num>
<text>   e for JapaneseEnglish computer manuals both containing lots of the same alphabetic technical terms. However, the method cannot be applied to general translations in structurally different languages. (=-=Kay and Roscheisen, 1993-=-) proposed a relaxation method to iteratively align bilingual texts using the word correspondences acquired during the alignment process. Although the method works well among European languages, the m ty on the occurrence distribution and t-score represents the confidence of the similarity. These two parameters permit more effective relaxation than the single parameter used in conventional methods(=-=Kay and Roscheisen, 1993-=-). Our basic data structure is the alignable sentence matrix (ASM) and the anchor matrix (AM). ASM represents possible sentence correspondences and consists of ones and zeros. A one in ASM indicates t f the two anchors with as many as O(~/~) (L is the number of sentences existing between two anchors) sentences in the other text because the maximum deviation can be stochastically modeled as O(~rL) (=-=Kay and Roscheisen, 1993-=-). The initial ASM has little effect on the alignment performance so long as it contains all correct sentence correspondences. 3.2.2 Constructing AM This step constructs an AM when given an ASM and a  ual alignment and indicate the difficulty of the task. Our evaluation focuses on much smaller texts than those used in other study(=-=Brown and others, 1993; Gale and Church, 1993; Wu, 1994; Fung, 1995;Kay and Roscheisen, 1993-=-) because our main targets are well-separated articles. However, our method will work on larger and noisy sets too, by using word anchors rather than using sentence boundaries as segment boundaries. I se a CD-ROM version of a JapaneseEnglish dictionary containing 40 thousands entries. Statistics repeats the iteration by using statistical corresponding words only. This is identical to Kay&amp;apos;s method (=-=Kay and Roscheisen, 1993-=-) except for the statistics used. Dictionary performs the iteration of the algorithm by using corresponding words of the bilingual dictionary. This delineates the coverage of the dictionary. The param   </text>
<query_num> 5206 </query_num>
<text>   een proposed to align bilingual corpora. One of the major approaches is based on the statistics of simple features such as sentence length in words (=-=Brown and others, 1991-=-) or in characters (=-=Gale and Church, 1993-=-). These techniques are widely used because they can be imple131 mented in an efficient and simple way through dynamic programing. However, their main targets are rigid translations that are almost li  simple-feature based methods to Japanese-English translations. One alternative alignment method is the lexiconbased approach that makes use of the wordcorrespondence knowledge of the two languages. (=-=Church, 1993-=-) employed n-grams shared by two languages. His method is also effective for JapaneseEnglish computer manuals both containing lots of the same alphabetic technical terms. However, the method cannot be ta. 135 categories of matches by manual alignment and indicate the difficulty of the task. Our evaluation focuses on much smaller texts than those used in other study(=-=Brown and others, 1993; Gale and Church, 1993; Wu, 1994; Fung, 1995; Kay and Roscheisen, 1993-=-) because our main targets are well-separated articles. However, our method will work on larger and noisy sets too, by using word anchors rather than us  Our method performs accurate alignment for such use by combining the detailed word correspondences: statistically acquired word correspondences and those from a bilingual dictionary of general use. (=-=Church, 1993-=-) proposed char_align that makes use of n-grams shared by two languages. This kind of matching techniques will be helpful in our dictionary-based approach in the following situation: Entries of a bili   </text>
<query_num> 5207 </query_num>
<text>   knowledge extraction (=-=Kupiec, 1993; Matsumoto et al., 1993; Smadja et al., 1996; Dagan and Church, 1994; Kumano and Hirakawa, 1994; Haruno et al., 1996-=-), machine translation (=-=Brown and others, 1993;Sato and Nagao, 1990; Kaji et al., 1992-=-) and information retrieval (=-=Sato, 1992-=-)). Most of these works assume voluminous aligned corpora. Many methods have been proposed to align bilingual corpora. One of the major approa ces are necessary for higher level applications such as well-grained translation template acquisition (=-=Matsumoto et as., 1993; Smadja et al., 1996; Haruno et al., 1996-=-) and example-based translation (=-=Sato and Nagao, 1990-=-). Our method performs accurate alignment for such use by combining the detailed word correspondences: statistically acquired word correspondences and those from a bilingual dictionary of general use.   </text>
<query_num> 5208 </query_num>
<text>   l Jsentence 2 Esentence2 Jsentence 3 Esentence3 • • , ° • • , • ° • ° , ° ° , ° , , , • • • , PM Jsentence Esentence N Figure 3: Possible Sentence Correspondences We introduce the contingency matrix (=-=Fung and Church, 1994-=-) to evaluate the similarity of word occurrences. Consider the contingency matrix shown Table 1, between Japanese word wjp n and English word Weng. The contingency matrix shows: (a) the number of pis  ntence alignment on the sentence-sentence basis. Our iterative approach decides sentence alignment level by level by counting the word correspondences between a Japanese and an English sentence. 137 (=-=Fung and Church, 1994; Fung, 1995-=-) proposed methods to find Chinese-English word correspondences without aligning parallel texts. Their motivation is that structurally different languages such as Chinese-English and Japan   </text>
<query_num> 5209 </query_num>
<text>   nted with regular expressions. 2We use in this phase the JUMAN morphological analyzing system (=-=Kurohashi et al., 1994-=-) for tagging Japanese texts and Brill&amp;apos;s transformation-based tagger (=-=Brill, 1992; Brill, 1994-=-) for tagging English texts (=-=JUMAN: ftp://ftp.aist-nara.ac.jp/pub/nlp/tools/juman/ Brih ftp://ftp.cs.jhu.edu/pub/brill-=-). We would like to thank all people concerned for providing us with the tools.sTh   </text>
<query_num> 5210 </query_num>
<text>   ntional methods for various kinds of Japanese-English texts. 1 Introduction Corpus-based approaches based on bilingual texts are promising for various applications(i.e., lexical knowledge extraction (=-=Kupiec, 1993; Matsumoto et al., 1993; Smadja et al., 1996; Dagan and Church, 1994; Kumano and Hirakawa, 1994; Haruno et al., 1996-=-), machine translation (=-=Brown and others, 1993; Sato and Nagao, 1990; Kaji et al., -=-  </text>
<query_num> 5211 </query_num>
<text>   of corresponding words in a hand-crafted bilingual dictionary. Although some results were promising, the method&amp;apos;s performance strongly depended on the domain of the texts and the dictionary entries. (=-=Utsuro et al., 1994-=-) introduced a statistical postprocessing step to tackle the problem. He first applied Sato&amp;apos;s method and extracted statistical word correspondences from the result of the first path. Sato&amp;apos;s method was   </text>
<query_num> 5212 </query_num>
<text>   tches by manual alignment and indicate the difficulty of the task. Our evaluation focuses on much smaller texts than those used in other study(=-=Brown and others, 1993; Gale and Church, 1993; Wu, 1994;Fung, 1995; Kay and Roscheisen, 1993-=-) because our main targets are well-separated articles. However, our method will work on larger and noisy sets too, by using word anchors rather than using sentence boundarie  sentence-sentence basis. Our iterative approach decides sentence alignment level by level by counting the word correspondences between a Japanese and an English sentence. 137 (=-=Fung and Church, 1994;Fung, 1995-=-) proposed methods to find Chinese-English word correspondences without aligning parallel texts. Their motivation is that structurally different languages such as Chinese-English and Japanese-English   </text>
</top>
<top>
<paper_num> 53 </paper_num>
<paper_title>   Leap Before You Look: Information Gathering in the PUCCINI Planner.  </paper_title>
<paper_abstract>   Most of the work in planning with incomplete information  takes a &amp;quot;look before you leap&amp;quot; perspective: Actions  must be guaranteed to have their intended effects  before they can be executed. We argue that this approach  is impossible to follow in many real-world domains.  The agent may not have enough information to  ensure that an action will have a given effect in advance  of executing it. This paper describes puccini, a partialorder  planner used to control the Internet Softbot (=-=Etzioni  &amp; Weld 1994-=-). puccini takes a different approach  to coping with incomplete information: &amp;quot;Leap before  you look!&amp;quot; puccini doesn&amp;apos;t require actions to be known  to have the desired effects before execution. However,  it still maintains soundness, by requiring the effects to  be verified eventually. We discuss how this is achieved  using a simple generalization of causal links.  Introduction  A boy&amp;apos;s appetite grows very fast, and in a few moments  the queer, empty feeling had become hunger, and the  hu...  </paper_abstract>
<query_num> 5301 </query_num>
<text>   Combining these features with observe effects yields expressive sensor models, such as those shown in the next section. puccini overview puccini is a partial-order planner in the same family as snlp (=-=McAllester &amp; Rosenblitt 1991-=-) and ucpop (=-=Penberthy &amp; Weld 1992-=-). It builds plans incrementally by starting with an empty plan and a goal agenda of goals that need to be achieved. When goals are achieved, they are removed from th   </text>
<query_num> 5302 </query_num>
<text>   action language (=-=Etzioni et al. 1992-=-). We believe that our use of verification links is unique. However, it should be possible for a Partially Observable Markov Decision Process (POMDP) (=-=Koenig 1992;Dean et al. 1995-=-), or the planner C-buridan (=-=Draper, Hanks, &amp; Weld 1994-=-), to produce plans similar to those produced by puccini using verification links. However, they accomplish this by following a generate-and-test   </text>
<query_num> 5303 </query_num>
<text>   ay not have enough information to ensure that an action will have a given effect in advance of executing it. This paper describes puccini, a partialorder planner used to control the Internet Softbot (=-=Etzioni &amp; Weld 1994-=-). puccini takes a different approach to coping with incomplete information: &amp;quot;Leap before you look!&amp;quot; puccini doesn&amp;apos;t require actions to be known to have the desired effects before execution. However,  tware environments, such as the Unix operating system or World Wide Web, in which the agent has massively incomplete (but correct) information about the world. One such agent is the Internet Softbot (=-=Etzioni &amp; Weld 1994-=-). Internet resources and Unix commands are represented as planner actions, and a planner, called puccini, 2 is used to find some combination of these actions that together will achieve the user&amp;apos;s goa   </text>
<query_num> 5304 </query_num>
<text>   erifying the preconditions afterward. Finally, we evaluate the cost of this added flexibility. Back in the sadl puccini goals and actions are described using the languagessadl, 3 which builds on uwl (=-=Etzioni et al. 1992-=-) and adl (=-=Penberthy 1993-=-). Like uwl, sadl is designed to represent sensing actions and information goals. To distinguish sensory effects from causal effects and goals of information from traditional  ccini, socrates utilized the Softbot domain as its testbed and interleaved planning with execution. However, socrates utilized knowledge preconditions and supported a less expressive action language (=-=Etzioni et al. 1992-=-). We believe that our use of verification links is unique. However, it should be possible for a Partially Observable Markov Decision Process (POMDP) (=-=Koenig 1992; Dean et al. 1995-=-), or the planner C-   </text>
<query_num> 5305 </query_num>
<text>   erward. Finally, we evaluate the cost of this added flexibility. Back in the sadl puccini goals and actions are described using the languagessadl, 3 which builds on uwl (=-=Etzioni et al. 1992-=-) and adl (=-=Penberthy 1993-=-). Like uwl, sadl is designed to represent sensing actions and information goals. To distinguish sensory effects from causal effects and goals of information from traditional goals of satisfaction, sa   </text>
<query_num> 5306 </query_num>
<text>   is organized as follows. 2 puccini stands for Planning with Universal quantification, Conditional effects, Causal links, and INcomplete Information. puccini is a partial-order planner based on ucpop (=-=Penberthy &amp; Weld 1992-=-). An earlier version of puccini was called xii. First, we introduce the fundamentals of the sadl language and the puccini planner. Then, in the next section, we briefly discuss the problem with knowl ects yields expressive sensor models, such as those shown in the next section. puccini overview puccini is a partial-order planner in the same family as snlp (=-=McAllester &amp; Rosenblitt 1991-=-) and ucpop (=-=Penberthy &amp; Weld 1992-=-). It builds plans incrementally by starting with an empty plan and a goal agenda of goals that need to be achieved. When goals are achieved, they are removed from the goal agenda. When actions are ad We have shown that this mechanism can be implemented without impairing tractability. Related Work puccini is an extension of xii (=-=Golden, Etzioni, &amp; Weld 1994-=-), which is based on the ucpop algorithm (=-=Penberthy &amp; Weld 1992-=-). puccini builds on xii by supporting a more expressive language, sadl (=-=Golden &amp; Weld 1996-=-), and handling verification links. xii builds on ucpop by dealing with information goals and effects, interl   </text>
<query_num> 5307 </query_num>
<text>   ost is daunting. Some planners represent uncertain outcomes using conditional effects, and can execute actions for their uncertain effects (=-=Kushmerick, Hanks, &amp; Weld 1995; Draper, Hanks, &amp; Weld 1994; Pryor &amp; Collins 1996; Goldman &amp; Boddy 1994-=-). For example, Cassandra (=-=Pryor &amp; Collins 1996-=-) represents uncertain outcomes as conditional effects with &amp;quot;:unknown&amp;quot; preconditions, and is capable of using these actions for the   </text>
<query_num> 5308 </query_num>
<text>   planners represent uncertain outcomes using conditional effects, and can execute actions for their uncertain effects (=-=Kushmerick, Hanks, &amp; Weld 1995; Draper, Hanks, &amp; Weld 1994; Pryor &amp; Collins 1996;Goldman &amp; Boddy 1994-=-). For example, Cassandra (=-=Pryor &amp; Collins 1996-=-) represents uncertain outcomes as conditional effects with &amp;quot;:unknown&amp;quot; preconditions, and is capable of using these actions for their uncertain effects.   </text>
<query_num> 5309 </query_num>
<text>   x), T) means &amp;quot;Ensure that there&amp;apos;s at least one file in directory tex,&amp;quot; and satisfy(in.dir (myfile, tex), tv) means &amp;quot;Find out whether or not myfile is in tex.&amp;quot; The initially annotation, introduced in (=-=Golden &amp; Weld 1996-=-), is similar to satisfy, but it refers to the time when the goal is given to the agent, not to the time when the goal is achieved. initially(P , tv) means that by the time the agent has finished exec  . However, changing P after determining its initial value is fine. By combining initially with satisfy we can express &amp;quot;tidiness&amp;quot; goals: modify P at will, but restore its initial value by plan&amp;apos;s end (=-=Golden &amp; Weld 1996; Weld &amp; Etzioni 1994-=-). Furthermore, we can express goals such as &amp;quot;Find the the file currently named paper.tex, and rename it to kr.tex,&amp;quot; which is beyond the expressive power of most planners (Golden  es &amp;quot;flaws&amp;quot; in the plan (open goals, threats and unexecuted actions) until the plan is complete. The lines of the algorithm relevant to this paper are shown in bold. needs to know the combination. In (=-=Golden &amp; Weld 1996-=-), we argued that the practice of specifying knowledge preconditions for actions was too restrictive and should be abandoned. Moore (=-=Moore 1985-=-) identified two kinds of knowledge preconditions an agen  g) to B. ffl Add hhContext ) pi, S p i to G Figure 7: Procedure Addlink. The lines of the algorithm relevant to this paper are shown in bold. these schemas. Many of these struggles are discussed in (=-=Golden &amp; Weld 1996; Etzioni, Golden, &amp; Weld 1997-=-). One of the greatest representational gains came from the elimination of knowledge preconditions and the introduction of verification links. For example, when knowledge ork puccini is an extension of xii (=-=Golden, Etzioni, &amp; Weld 1994-=-), which is based on the ucpop algorithm (=-=Penberthy &amp; Weld 1992-=-). puccini builds on xii by supporting a more expressive language, sadl (=-=Golden &amp; Weld 1996-=-), and handling verification links. xii builds on ucpop by dealing with information goals and effects, interleaving planning with execution and reasoning with Local Closed World knowledge (lcw) (=-=Etzio-=-   </text>
</top>
<top>
<paper_num> 54 </paper_num>
<paper_title>   Unification and polymorphism in region inference.  </paper_title>
<paper_abstract>   Region Inference is a technique for inferring lifetimes of values in strict, higher-order programming languages such as Standard ML. The purpose of this paper is to show how ideas from Milner&amp;apos;s polymorphic type discipline can serve as a basis for region inference, even in the presence of a limited form of polymorphic recursion.  </paper_abstract>
<query_num> 5401 </query_num>
<text>   27 are used for. 3 Related Work The basic ideas of the region inference scheme are described in [23]. An extended version, including a proof that the region inference rules are sound, may be found in =-=[24]-=-. Other analysis which have been combined with region inference are described in [1, 2]. The emphasis of this paper is on using unification to constrain region variables and arrow effects. (Arrow effe s equivalent to semi-unification[5, 8], which is undecidable[7]. However, it can be proved that every well-typed source expression can be region-annotated in accordance with the region inference rules=-=[24]-=-. Moreover, in the absence of region-polymorphic recursion, Milner&amp;apos;s notion of principal type schemes extends to region inference[22]. In the presence of polymorphic recursion, however, it is not know e chosen disjoint from B. We emphasise that consistency of type schemes is introduced for algorithmic reasons only. Soundness of the region inference rules has been proved without assuming consistency=-=[23, 24]-=-. Also, we have not found the restriction to consistent type schemes (and the loss in polymorphism it entails) to be serious in practice. Definition 12 A type environment TE is consistent in B = (Q; \   </text>
<query_num> 5402 </query_num>
<text>   it relies on the results about unification proved in the present paper. A different approach to region inference is to use constraints. Constraints have been used in previous work on ML type inference=-=[5]-=-, subtyping [12, 4] and effect systems[19, 18, 14, 13]. We are currently exploring using constraints for region inference. The relative merits of the two approaches to region inference (syntax-directe ent of recursive functions in ML as far as type polymorphism is concerned[10]. The reason for this limitation is that type inference for (type-) polymorphic recursion is equivalent to semi-unification=-=[5, 8]-=-, which is undecidable[7]. However, it can be proved that every well-typed source expression can be region-annotated in accordance with the region inference rules[24]. Moreover, in the absence of regi   </text>
<query_num> 5403 </query_num>
<text>   n proved in the present paper. A different approach to region inference is to use constraints. Constraints have been used in previous work on ML type inference[5], subtyping [12, 4] and effect systems=-=[19, 18, 14, 13]-=-. We are currently exploring using constraints for region inference. The relative merits of the two approaches to region inference (syntax-directed region inference versus constraint generation and co   </text>
<query_num> 5404 </query_num>
<text>   n proved in the present paper. A different approach to region inference is to use constraints. Constraints have been used in previous work on ML type inference[5], subtyping [12, 4] and effect systems=-=[19, 18, 14, 13]-=-. We are currently exploring using constraints for region inference. The relative merits of the two approaches to region inference (syntax-directed region inference versus constraint generation and co ffect Variables and Arrow Effects In the type scheme for hanoi we use ffl and &amp;quot;where&amp;quot; as meta-notation to make the type scheme easier to read. A crucial next step, invented in work on effect inference=-=[18]-=-, is to make effect variables part of the language of types, on a par with type variables and region variables. Moreover, we represent &amp;quot;where ffl = fae 6 ; ae 7 ; ae 8 ; ae 9 ; ae 11 g&amp;quot; by a formal ob :(&amp;apos; 1 [ &amp;apos; 2 ) In the picture, S is a substitution. (Substitutions will be defined in Section 7.) Another way of thinking of ffl 1 :&amp;apos; 1 and ffl 2 :&amp;apos; 2 is as two constraints ffl 1 &amp;apos; &amp;apos; 1 and ffl 2 &amp;apos; &amp;apos; 2 =-=[18]-=-; then the &amp;quot;unification&amp;quot; of the two arrow effect corresponds to replacing ffl 2 by ffl 1 everywhere and replacing the two constraints by a single constraint ffl 1 &amp;apos; (&amp;apos; 1 [ &amp;apos; 2 ). Although the region i   </text>
<query_num> 5405 </query_num>
<text>   nference specialises memory management to the particular program that is being compiled. The relevance of region inference to this Festschrift is that Milner&amp;apos;s work on type inference and type checking=-=[10]-=- has provided many of the technical insights which underlie region inference. (Other origins of region inference are listed in Section 3.) In particular, the idea of using unification in type checking s and effects. In order to satisfy the premises of the inference rule, the region inference algorithm needs to be able to unify types which contain region and effect information. Hindley[6] and Milner=-=[10]-=- discovered that Robinson&amp;apos;s unification algorithm can be used for unifying types. We assume that type inference has already been carried out by the time region inference is performed. Thus the problem lymorphic recursion in types: the type scheme for f in e 1 is onlysoe, not oe. Thus the region rules are faithful to the treatment of recursive functions in ML as far as type polymorphism is concerned=-=[10]-=-. The reason for this limitation is that type inference for (type-) polymorphic recursion is equivalent to semi-unification[5, 8], which is undecidable[7]. However, it can be proved that every well-ty   </text>
<query_num> 5406 </query_num>
<text>   ort for reasoning about the lifetimes of storage cells, and thus it is in general very difficult for programmers to reason about how much memory their programs will use. Region-based memory management=-=[23, 1, 2]-=- is yet a form of automatic management of dynamic allocation. Conceptually, the store consists of a stack of regions. A region can be thought of as a heap which can grow dynamically depending on how m scribed in [23]. An extended version, including a proof that the region inference rules are sound, may be found in [24]. Other analysis which have been combined with region inference are described in =-=[1, 2]-=-. The emphasis of this paper is on using unification to constrain region variables and arrow effects. (Arrow effects decorate function arrows in the region type system with information of how the func   </text>
<query_num> 5407 </query_num>
<text>   ort for reasoning about the lifetimes of storage cells, and thus it is in general very difficult for programmers to reason about how much memory their programs will use. Region-based memory management=-=[23, 1, 2]-=- is yet a form of automatic management of dynamic allocation. Conceptually, the store consists of a stack of regions. A region can be thought of as a heap which can grow dynamically depending on how m scribed in [23]. An extended version, including a proof that the region inference rules are sound, may be found in [24]. Other analysis which have been combined with region inference are described in =-=[1, 2]-=-. The emphasis of this paper is on using unification to constrain region variables and arrow effects. (Arrow effects decorate function arrows in the region type system with information of how the func above effect would be written fget(ae 6 ); put(ae 7 ); put(ae 8 ); put(ae 9 ); get(ae 11 )g. This distinction is useful for other analyses and optimisations which can be combined with region inference=-=[2]-=-; however, for the purpose of this paper, the distinction between put and get is not important and is therefore omitted. 5.2 Effect Variables and Arrow Effects In the type scheme for hanoi we use ffl   </text>
<query_num> 5408 </query_num>
<text>   storage cells. The term &amp;quot;garbage collection&amp;quot; is traditionally used for a range of heap memory management techniques, including reference counting, copying collection and generational collection (see =-=[25]-=- for an excellent overview). 1 Common to all of these techniques is that there is a strict separation between the program which allocates memory, called the mutator, and the part of the runtime system   </text>
</top>
<top>
<paper_num> 55 </paper_num>
<paper_title>   Organizing shared enterprise workspaces using component-based cooperative hypermedia.  </paper_title>
<paper_abstract>   Cooperative work in Extended Enterprises needs a flexible shared workspace for team members to access and manipulate shared information objects in a well-coordinated working process. Current shared workspace systems do not adequately support the evolving character of shared workspaces as needed by Extended Enterprises, i.e. the dynamic cooperation processes, various kinds of shared information contents and the set of cooperative tools. In this paper, the usage scenarios and requirements developed in a European Extended Enterprise project are used to derive the requirements for shared enterprise workspaces. Our approach utilizes component-based cooperative hypermedia to organize shared enterprise workspaces that contain team and process structures, information contents and their corresponding tools. The approach extends classical hypertext models to shared hypermedia objects as well as dynamic bindings between these and the Groupware Components working on them. To demonstrate the approach, a prototype system and a prototypical usage scenario are presented.  </paper_abstract>
<query_num> 5501 </query_num>
<text>   ained notifications giving immediate feedback about the activities of other users, whereas asynchronous cooperation can happen at different times with usually no notifications of other users&amp;apos; actions =-=[13]-=-. Dynamic Networked Organizations Dynamic networked organizations are a new form of organizing business activities in our evolving global economy. Virtual Enterprises (VEs) and Extended Enterprises (E   </text>
<query_num> 5502 </query_num>
<text>   and manipulated by heterogeneous tools. It also provides an open protocol to integrate new tools into its infrastructure. It supports asynchronous cooperative authoring with heterogeneous tools. CAOS =-=[21]-=- is a component-based open hypermedia system, which focuses on spatial parsing and cooperation support. Posties is a WebDAV application for cooperative work [6]. It supports asynchronous cooperative a   </text>
<query_num> 5503 </query_num>
<text>   ation a conferencing system is used. Software components for extensibility and tailorability of the cooperation environment are not used in this approach. Web-integration is not mentioned. HyperDisco =-=[31]-=- supports shared workspaces containing hypermedia objects and external application files that can be accessed and manipulated by heterogeneous tools. It also provides an open protocol to integrate new   </text>
<query_num> 5504 </query_num>
<text>   can’t be used synchronously by multiple users. Up to now, synchronous cooperation with immediate interactive feedback has been added to classical workflow systems only for conference purposes like in =-=[30]-=-, but not in order to work synchronously on shared objects. Web-based workflow systems often provide users only with a “ToDo” list and task descriptions, e.g. [1], or document lists with some document   </text>
<query_num> 5505 </query_num>
<text>   d to be supported to address R3, R5 and R6. The first two aspects require generic cooperation support and emergent workflow modeling. Here, solutions have been presented, e.g. [11] or in CHIPS ([28], =-=[29]-=-). The third and fourth aspects, though, pose new problems for hypermedia approaches. While Dexter-based and open hypermedia protocol-based systems have been designed to accommodate integration of hyp ation services, hypermedia objects and Groupware Components is achieved. Cooperative Modeling and Execution of Emergent Work Processes Our model is based on experiences from our earlier work on CHIPS =-=[28, 29]-=-. Extensions to this work are motivated by the requirements of EXTERNAL. We model the work processes in the shared enterprise workspace using shared hypermedia with process support [28] (see Figure 2) d data flow semantics into a special link class called Flow, such as which nodes contained in a source task composite should flow into a destination task composite after completion of the source task =-=[29]-=-. Hence, a process is represented as a set of hypermedia tasks connected by flows. Such a hypermedia-based process representation can contain associated materials because the hypermedia structure is n t model while our approach provides more tight and flexible tool integration needed by Groupware systems that support synchronous cooperation. Cooperative Hypermedia systems like SEPIA [23] and CHIPS =-=[29]-=- provide flexible ways for structuring shared workspaces. Additionally, asynchronous as well as synchronous cooperative work is supported. Awareness is supported on hypermedia nodes and on some specia   </text>
<query_num> 5506 </query_num>
<text>   dia system, which focuses on spatial parsing and cooperation support. Posties is a WebDAV application for cooperative work [6]. It supports asynchronous cooperative authoring over the Web. Manufaktur =-=[15]-=- uses open hypermedia technology to integrate documents in a cooperative virtual environment with a 3D user interface. Open hypermedia systems, e.g., HyperDisco, focus on linking and integrating heter   </text>
<query_num> 5507 </query_num>
<text>   e a document or a task generates some documents. Active Queries for Workflow Support. Workflow technology is intended to support work by enacting explicitly modeled and represented business processes =-=[4]-=-. Based on the process related semantics (incorporated in our hypermedia model) the modeled cooperative processes can be executed, e.g. setting a task’s state to completed (by using the cooperative hy   </text>
<query_num> 5508 </query_num>
<text>   h concepts – shared hypermedia and software components – and extends classical hypertext models like the Dexter hypertext reference model [12] and the core data model of the open hypermedia framework =-=[20]-=- to shared hypermedia objects. In addition, it provides dynamic bindings between shared hypermedia objects and Groupware Components that manipulate the hypermedia objects. In this way, a tighter integ es and links and therefore allow nested structures. In the Dexter model [12], nodes, links and composites are unified to the term “Components”. In the core data model of the open hypermedia framework =-=[20]-=-, a wider range of these different hypermedia types, including e.g. anchors, is unified to the term “HMObject”. We also � � � � � � � � � � � �suse the term “HMObject” since we use the term “Component does not explicitly deal with the sharing and cooperative manipulation aspects of hypertexts. Our approach goes beyond the Dexter model as well as the core data model of the open hypermedia framework =-=[20]-=- by including shared hypermedia objects as well as bindings between shared hypermedia objects and Groupware Components. In the core data model of the open hypermedia framework a collaboration service   </text>
<query_num> 5509 </query_num>
<text>   of who is doing what. All EXTERNAL usage scenarios require supporting workspace awareness. This is meant to be an up-to-the-moment understanding of another person’s interaction in a shared workspace =-=[10]-=-. Workspace awareness improves the usability of cooperation: R4. Workspace awareness should be supported to provide users with information about the current state of their cooperative work. Additional   </text>
<query_num> 5510 </query_num>
<text>   our Groupware Components are). Finally, Web front ends to these systems provide only limited functionality. Systems using the metaphor of rooms for grouping users and group-aware tools like TeamRooms =-=[22]-=- or CBE [14] are extensible: New tools can be added to rooms. However, for structuring the information space only tool and URL lists [14] are available. Persistent documents that are manipulated by to  do not work on a shared data model, so that the tailorability is limited. Communication is supported via corresponding tools, such as a chat tool. Regarding awareness, user presence and telepointers =-=[22]-=- are provided. CBE is Webbased and therefore allows adding of URLs into rooms. Semantically rich structures in the sense of process support are provided by classical workflow systems. However, classic   </text>
<query_num> 5511 </query_num>
<text>   se the terms &amp;quot;Groupware Component&amp;quot; or &amp;quot;component&amp;quot; to refer to mobile cooperative software tools. It is important to bear in mind the difference to other uses of the term &amp;quot;component&amp;quot;, e.g. in [12] and =-=[18]-=-. manipulating, navigating and searching through the hypermedia workspace structure as well as for editing the actual contents. The component-based architecture provides easy deployment and extensibil   </text>
<query_num> 5512 </query_num>
<text>   upware Component model while our approach provides more tight and flexible tool integration needed by Groupware systems that support synchronous cooperation. Cooperative Hypermedia systems like SEPIA =-=[23]-=- and CHIPS [29] provide flexible ways for structuring shared workspaces. Additionally, asynchronous as well as synchronous cooperative work is supported. Awareness is supported on hypermedia nodes and   </text>
</top>
<top>
<paper_num> 56 </paper_num>
<paper_title>   Rule interestingness analysis using OLAP operations.  </paper_title>
<paper_abstract>   The problem of interestingness of discovered rules has been investigated by many researchers. The issue is that data mining algorithms often generate too many rules, which make it very hard for the user to find the interesting ones. Over the years many techniques have been proposed. However, few have made it to real-life applications. Since August 2004, we have been working on a major application for Motorola. The objective is to find causes of cellular phone call failures from a large amount of usage log data. Class association rules have been shown to be suitable for this type of diagnostic data mining application. We were also able to put several existing interestingness methods to the test, which revealed some major shortcomings. One of the main problems is that most existing methods treat rules individually. However, we discovered that users seldom regard a single rule to be interesting by itself. A rule is only interesting in the context of some other rules. Furthermore, in many cases, each individual rule may not be interesting, but a group of them together can represent an important piece of knowledge. This led us to discover a deficiency of the current rule mining paradigm. Using non-zero minimum support and non-zero minimum confidence eliminates a large amount of context information, which makes rule analysis difficult. This paper proposes a novel approach to deal with all of these issues, which casts rule analysis as OLAP operations and general impression mining. This approach enables the user to explore the knowledge space to find useful knowledge easily and systematically. It also provides a natural framework for visualization. As an evidence of its effectiveness, our system, called Opportunity Map, based on these ideas has been deployed, and it is in daily use in Motorola for finding actionable knowledge from its engineering and other types of data sets.  </paper_abstract>
<query_num> 5601 </query_num>
<text>   al rule query languages to enable the user to specify what rules that he/she needs and the system then retrieves the relevant rules. We tried this approach, but our users did not know what to ask. In =-=[29]-=-, a set of rule post-processing operators is defined to allow the user to filter unwanted rules, select rules of interest, group rules, etc. This is a good approach. However, it stops short of providi   </text>
<query_num> 5602 </query_num>
<text>   earlier. This method does not find generalized knowledge. Querying and filtering: In [6][21], some data mining query languages are proposed to select the right data to mine different types of rules. =-=[14]-=-[30][31] also report several rule query languages to enable the user to specify what rules that he/she needs and the system then retrieves the relevant rules. We tried this approach, but our users did   </text>
<query_num> 5603 </query_num>
<text>   ingness methods that can help the user find interesting knowledge. Unexpectedness: In this method, the user is asked to give some existing knowledge and the system then finds the unexpected rules [11]=-=[19]-=-[23][26][33]. This did not work well because our users were not sure what to expect. They wanted the system to find interesting knowledge for them. [4] studies neighborhood unexpectedness of rules. Th   </text>
<query_num> 5604 </query_num>
<text>   lier. This method does not find generalized knowledge. Querying and filtering: In [6][21], some data mining query languages are proposed to select the right data to mine different types of rules. [14]=-=[30]-=-[31] also report several rule query languages to enable the user to specify what rules that he/she needs and the system then retrieves the relevant rules. We tried this approach, but our users did not   </text>
<query_num> 5605 </query_num>
<text>   ly sufficient as we stated earlier. • We set non-zero minsup and minconf for rules with more conditions. This prevents combinatorial explosion. Multiple minimum supports give additional flexibilities =-=[18]-=-, which deal with skewed data and also enable the system not to generate unwanted rules [15]. These are useful in practice. All these can be done quite easily in a CAR miner such as CBA [17] by settin   </text>
<query_num> 5606 </query_num>
<text>   methods that can help the user find interesting knowledge. Unexpectedness: In this method, the user is asked to give some existing knowledge and the system then finds the unexpected rules [11][19][23]=-=[26]-=-[33]. This did not work well because our users were not sure what to expect. They wanted the system to find interesting knowledge for them. [4] studies neighborhood unexpectedness of rules. The neighb   </text>
<query_num> 5607 </query_num>
<text>   restingness methods that can help the user find interesting knowledge. Unexpectedness: In this method, the user is asked to give some existing knowledge and the system then finds the unexpected rules =-=[11]-=-[19][23][26][33]. This did not work well because our users were not sure what to expect. They wanted the system to find interesting knowledge for them. [4] studies neighborhood unexpectedness of rules   </text>
<query_num> 5608 </query_num>
<text>   s related to rule visualization [13]. [10] proposes interactive mosaic plots to visualize the contingency tables of association rules. In [5], classification rules are visualized using rule polygons. =-=[34]-=- visualizes the temporal behavior of rules. In [12], a postprocessing environment is proposed to browse and visualize association rules so that the user can divide a large rule set into smaller ones.   </text>
<query_num> 5609 </query_num>
<text>   the system then finds the unexpected rules [11][19][23][26][33]. This did not work well because our users were not sure what to expect. They wanted the system to find interesting knowledge for them. =-=[4]-=- studies neighborhood unexpectedness of rules. The neighborhood of a rule, which is similar to the concept of context, is a set of syntactically similar rules, i.e., involving similar items. This is n   </text>
<query_num> 5610 </query_num>
<text>   tically similar rules, i.e., involving similar items. This is not applicable to us. We need a different definition and also different rule mining. The idea of general impressions is first proposed in =-=[16]-=-. However, they need to be given by the user for finding unexpected rules. In this work, we mine general impressions from discovered rules. Rule ranking: Ranking rules according to some interestingnes   </text>
<query_num> 5611 </query_num>
<text>   y need to be given by the user for finding unexpected rules. In this work, we mine general impressions from discovered rules. Rule ranking: Ranking rules according to some interestingness measures [2]=-=[9]-=-[28]. Our experiences show that almost all top ranked rules represent some artifacts of the data rather than any useful patterns. Moreover, giving only individual rules without contexts to compare wit   </text>
</top>
<top>
<paper_num> 57 </paper_num>
<paper_title>   Simultaneous Feature Selection and Clustering Using Mixture Models.  </paper_title>
<paper_abstract>   Abstract—Clustering is a common unsupervised learning technique used to discover group structure in a set of data. While there exist many algorithms for clustering, the important issue of feature selection, that is, what attributes of the data should be used by the clustering algorithms, is rarely touched upon. Feature selection for clustering is difficult because, unlike in supervised learning, there are no class labels for the data and, thus, no obvious criteria to guide the search. Another important problem in clustering is the determination of the number of clusters, which clearly impacts and is influenced by the feature selection issue. In this paper, we propose the concept of feature saliency and introduce an expectation-maximization (EM) algorithm to estimate it, in the context of mixture-based clustering. Due to the introduction of a minimum message length model selection criterion, the saliency of irrelevant features is driven toward zero, which corresponds to performing feature selection. The criterion and algorithm are then extended to simultaneously estimate the feature saliencies and the number of clusters. Index Terms—Feature selection, clustering, unsupervised learning, mixture models, minimum message length, EM algorithm. 1  </paper_abstract>
<query_num> 5701 </query_num>
<text>   . It is, of course, impractical to exhaustively list the numerous uses of clustering techniques. Image segmentation, an important problem in computer vision, can be formulated as a clustering problem =-=[21]-=-, [28], [55]. Documents can be clustered [23] to generate topical hierarchies for information access [53] or retrieval [5]. Clustering is also used to perform market segmentation [2], [11] as well as   </text>
<query_num> 5702 </query_num>
<text>   assifiers, with each one using only one feature, and then apply boosting, which effectively performs featuresLAW ET AL.: SIMULTANEOUS FEATURE SELECTION AND CLUSTERING USING MIXTURE MODELS 3 selection =-=[59]-=-. It has also been proposed to approach feature selection using rough set theory [35]. All of the approaches mentioned above are concerned with feature selection in the presence of class labels. Compa   </text>
<query_num> 5703 </query_num>
<text>   course, impractical to exhaustively list the numerous uses of clustering techniques. Image segmentation, an important problem in computer vision, can be formulated as a clustering problem [21], [28], =-=[55]-=-. Documents can be clustered [23] to generate topical hierarchies for information access [53] or retrieval [5]. Clustering is also used to perform market segmentation [2], [11] as well as in biology,   </text>
<query_num> 5704 </query_num>
<text>   d not select good features for clustering, as features with large variance can be independent of the intrinsic grouping of the data (see example in Fig. 3). Most feature selection algorithms (such as =-=[9]-=-, [33], [47]) involve a combinatorial search through the space of all feature subsets. Usually, heuristic (nonexhaustive) methods have to be adopted, because the size of this space is 0162-8828/04/$20 ture subsets; for this task, different types of heuristics, such as sequential forward or backward searches, floating search, beam search, bidirectional search, and genetic search have been suggested =-=[9]-=-, [33], [47], [63]. It is also possible to construct a set of weak (in the boosting sense [20]) classifiers, with each one using only one feature, and then apply boosting, which effectively performs f   </text>
<query_num> 5705 </query_num>
<text>   e may have some knowledge of the class labels of different Gaussian components. This can happen when, say, we adopt a procedure to combine different Gaussian components to form a cluster (e.g., as in =-=[51]-=-), or in a semisupervised learning scenario, where we can use a small amount of labelled data to help us identify which Gaussian component belongs to which class. This additional information can sugge   </text>
<query_num> 5706 </query_num>
<text>   educe the complexity by adopting optimization techniques applicable for standard EM for Gaussian mixture, such as sampling the data, compressing the data [8], or using efficient data structures [45], =-=[54]-=-. For the postprocessing step in Section 3.4, each computation of the quantity J and its gradient and Hessian takes OðNDKÞ time. The number of iterations is difficult to predict, as it depends on the   </text>
<query_num> 5707 </query_num>
<text>   exture) consists of 4,000 19dimensional Gabor filter features from a collage of four Brodatz textures [27]. A data set (zer) of 47 Zernike moments extracted from images of handwriting numerals (as in =-=[26]-=-) are also used; there are 200 images for each digit, totaling 2,000 patterns. The data sets wine, wdbc, image, and zernike are from the UCI machine learning repository (http://www.ics.uci.edu/~mlearn   </text>
<query_num> 5708 </query_num>
<text>   filters and wrappers. The filter approaches evaluate the relevance of each feature (subset) using the data set alone, regardless of the subsequent learning algorithm. RELIEF [32] and its enhancement =-=[36]-=- are representatives of this class, where the basic idea is to assign feature weights based on the consistency of the feature value in the k nearest neighbors of every data point. Informationtheoretic   </text>
<query_num> 5709 </query_num>
<text>   ing is also used to perform market segmentation [2], [11] as well as in biology, e.g., to study genome data [3]. Many clustering algorithms have been proposed in different application scenarios [25], =-=[29]-=-. They can be divided roughly into two categories: hierarchical clustering, which creates a “tree” with branches merging at different levels, and partitional clustering, which divides the data into di   </text>
<query_num> 5710 </query_num>
<text>   irrelevant features are uncorrelated with the relevant features. Reference [14] describes the notion of “category utility” for feature selection in a conceptual clustering task. The CLIQUE algorithm =-=[1]-=- is popular in the data mining community and it finds hyperrectangular shaped clusters using a subset of attributes for a large database. The wrapper approach can also be adopted to select features fo   </text>
<query_num> 5711 </query_num>
<text>   iscriminant. Feature weighting for k-means clustering is also considered in [41], but the goal there is to find the best description of the clusters after they are identified. The method described in =-=[46]-=- can be classified as learning feature weights for conditional Gaussian networks. An EM algorithm based on Bayesian shrinking is proposed in [22] for unsupervised learning. 3 EM ALGORITHM FOR FEATURE   </text>
<query_num> 5712 </query_num>
<text>   kward searches, floating search, beam search, bidirectional search, and genetic search have been suggested [9], [33], [47], [63]. It is also possible to construct a set of weak (in the boosting sense =-=[20]-=-) classifiers, with each one using only one feature, and then apply boosting, which effectively performs featuresLAW ET AL.: SIMULTANEOUS FEATURE SELECTION AND CLUSTERING USING MIXTURE MODELS 3 select   </text>
<query_num> 5713 </query_num>
<text>   lassification methods may use each pixel as a feature [6], thus easily involving thousands of features. Feature selection has been widely studied in the context of supervised learning (see [7], [24], =-=[33]-=-, [34] and references therein), where the ultimate goal is to select features that can achieve the highest accuracy on unseen data. Feature selection has received comparatively very little attention i  select good features for clustering, as features with large variance can be independent of the intrinsic grouping of the data (see example in Fig. 3). Most feature selection algorithms (such as [9], =-=[33]-=-, [47]) involve a combinatorial search through the space of all feature subsets. Usually, heuristic (nonexhaustive) methods have to be adopted, because the size of this space is 0162-8828/04/$20.00 ß  ost of the literature on feature selection pertains to supervised learning (both classification [24] and regression [40]). Feature selection algorithms can be broadly divided into two categories [7], =-=[33]-=-: filters and wrappers. The filter approaches evaluate the relevance of each feature (subset) using the data set alone, regardless of the subsequent learning algorithm. RELIEF [32] and its enhancement  it is conditionally independent of the class labels given other features. The concept of Markov blanket is used to formalize this notion of irrelevancy in [34]. On the other hand, wrapper approaches =-=[33]-=- invoke the learning algorithm to evaluate the quality of each feature (subset). Specifically, a learning algorithm (e.g., a nearest neighbor classifier, a decision tree, a naive Bayes method) is run  classification accuracy. Wrappers are usually more computationally demanding, but they can be superior in accuracy when compared with filters, which ignore the properties of the learning task at hand =-=[33]-=-. Both approaches, filters and wrappers, usually involve combinatorial searches through the space of possible feature subsets; for this task, different types of heuristics, such as sequential forward   </text>
<query_num> 5714 </query_num>
<text>   mage classification methods may use each pixel as a feature [6], thus easily involving thousands of features. Feature selection has been widely studied in the context of supervised learning (see [7], =-=[24]-=-, [33], [34] and references therein), where the ultimate goal is to select features that can achieve the highest accuracy on unseen data. Feature selection has received comparatively very little atten n Section 5. Finally, we conclude in Section 6 and outline some future work directions. 2 RELATED WORK Most of the literature on feature selection pertains to supervised learning (both classification =-=[24]-=- and regression [40]). Feature selection algorithms can be broadly divided into two categories [7], [33]: filters and wrappers. The filter approaches evaluate the relevance of each feature (subset) us in Fig. 8a. We can conclude that, in this case, the algorithm successfully locates the true clusters and correctly assigns the feature saliencies. In the second experiment, we consider the Trunk data =-=[24]-=-, [57]: two 20-dimensional Gaussians Nðm1; IÞ and Nðm2; IÞ, where m1 ð1; 1ffiffi p ; ...; 2 1ffiffiffi p Þ, m2  m1. Data are obtained by 20 sampling 5,000 points from each of these two Gaussians. No   </text>
<query_num> 5715 </query_num>
<text>   may involve thousands of features [3], [62], and a Web page can be represented by thousands of different key-terms [58]. Appearance-based image classification methods may use each pixel as a feature =-=[6]-=-, thus easily involving thousands of features. Feature selection has been widely studied in the context of supervised learning (see [7], [24], [33], [34] and references therein), where the ultimate go   </text>
<query_num> 5716 </query_num>
<text>   n [43], weights are assigned to different groups of features for k-means clustering based on a score related to the Fisher discriminant. Feature weighting for k-means clustering is also considered in =-=[41]-=-, but the goal there is to find the best description of the clusters after they are identified. The method described in [46] can be classified as learning feature weights for conditional Gaussian netw   </text>
<query_num> 5717 </query_num>
<text>   ortant problem in computer vision, can be formulated as a clustering problem [21], [28], [55]. Documents can be clustered [23] to generate topical hierarchies for information access [53] or retrieval =-=[5]-=-. Clustering is also used to perform market segmentation [2], [11] as well as in biology, e.g., to study genome data [3]. Many clustering algorithms have been proposed in different application scenari   </text>
<query_num> 5718 </query_num>
<text>   roblem is made even more challenging when the number of clusters is unknown, since the optimal number of clusters and the optimal feature subset are interrelated, as illustrated in Fig. 2 (taken from =-=[16]-=-). Note that methods based on variance (such as principal components analysis) need not select good features for clustering, as features with large variance can be independent of the intrinsic groupin nsupervised learning; it is the case for methods that measure feature similarity to detect redundant features, using, e.g., mutual information [53] or a maximum information compression index [42]. In =-=[16]-=-, [17], the normalized log-likelihood and cluster separability are used to evaluate the quality of clusters obtained with different feature subsets. Different feature subsets and numbers of clusters,   </text>
<query_num> 5719 </query_num>
<text>   s with large numbers of features, e.g., classification problems in molecular biology may involve thousands of features [3], [62], and a Web page can be represented by thousands of different key-terms =-=[58]-=-. Appearance-based image classification methods may use each pixel as a feature [6], thus easily involving thousands of features. Feature selection has been widely studied in the context of supervised ned with different feature subsets. Different feature subsets and numbers of clusters, for multinomial model-based clustering, are evaluated using marginal likelihood and crossvalidated likelihood in =-=[58]-=-. The algorithm described in [52] uses automatic relevance determination priors to select features when there are two clusters. In [13], the clustering tendency of each feature is assessed by an entro ls, as well as in the emission densities of continuous hidden Markov models. Among different definitions of feature irrelevancy (proposed for supervised learning), we adopt the one suggested in [48], =-=[58]-=-, which is suitable for unsupervised learning: the lth feature is irrelevant if its distribution is independent of the class labels, i.e., if it follows a common density, denoted by qðylj lÞ. Let ð 1   </text>
<query_num> 5720 </query_num>
<text>   sed for unsupervised learning; it is the case for methods that measure feature similarity to detect redundant features, using, e.g., mutual information [53] or a maximum information compression index =-=[42]-=-. In [16], [17], the normalized log-likelihood and cluster separability are used to evaluate the quality of clusters obtained with different feature subsets. Different feature subsets and numbers of c   </text>
<query_num> 5721 </query_num>
<text>   terpretable models. Feature selection is particularly important for data sets with large numbers of features, e.g., classification problems in molecular biology may involve thousands of features [3], =-=[62]-=-, and a Web page can be represented by thousands of different key-terms [58]. Appearance-based image classification methods may use each pixel as a feature [6], thus easily involving thousands of feat   </text>
<query_num> 5722 </query_num>
<text>   ther reduce the complexity by adopting optimization techniques applicable for standard EM for Gaussian mixture, such as sampling the data, compressing the data [8], or using efficient data structures =-=[45]-=-, [54]. For the postprocessing step in Section 3.4, each computation of the quantity J and its gradient and Hessian takes OðNDKÞ time. The number of iterations is difficult to predict, as it depends o   </text>
<query_num> 5723 </query_num>
<text>   this task, different types of heuristics, such as sequential forward or backward searches, floating search, beam search, bidirectional search, and genetic search have been suggested [9], [33], [47], =-=[63]-=-. It is also possible to construct a set of weak (in the boosting sense [20]) classifiers, with each one using only one feature, and then apply boosting, which effectively performs featuresLAW ET AL.:   </text>
<query_num> 5724 </query_num>
<text>   tomatic relevance determination priors to select features when there are two clusters. In [13], the clustering tendency of each feature is assessed by an entropy index. A genetic algorithm is used in =-=[31]-=- for feature selection in k-means clustering. In [56], feature selection for symbolic data is addressed by assuming that irrelevant features are uncorrelated with the relevant features. Reference [14]   </text>
<query_num> 5725 </query_num>
<text>   uated using marginal likelihood and crossvalidated likelihood in [58]. The algorithm described in [52] uses automatic relevance determination priors to select features when there are two clusters. In =-=[13]-=-, the clustering tendency of each feature is assessed by an entropy index. A genetic algorithm is used in [31] for feature selection in k-means clustering. In [56], feature selection for symbolic data   </text>
<query_num> 5726 </query_num>
<text>   ures when there are two clusters. In [13], the clustering tendency of each feature is assessed by an entropy index. A genetic algorithm is used in [31] for feature selection in k-means clustering. In =-=[56]-=-, feature selection for symbolic data is addressed by assuming that irrelevant features are uncorrelated with the relevant features. Reference [14] describes the notion of “category utility” for featu   </text>
<query_num> 5727 </query_num>
<text>   vised learning; it is the case for methods that measure feature similarity to detect redundant features, using, e.g., mutual information [53] or a maximum information compression index [42]. In [16], =-=[17]-=-, the normalized log-likelihood and cluster separability are used to evaluate the quality of clusters obtained with different feature subsets. Different feature subsets and numbers of clusters, for mu   </text>
</top>
<top>
<paper_num> 58 </paper_num>
<paper_title>   Resettable Zero Knowledge in the Bare Public-Key Model under Standard Assumption.  </paper_title>
<paper_abstract>   In this paper we resolve an open problem regarding resettable zero knowledge  in the bare public-key (BPK for short) model: Does there exist constant  round resettable zero knowledge argument with concurrent soundness for    in BPK model without assuming sub-exponential hardness? We give a  positive answer to this question by presenting such a protocol for any language    in the bare public-key model assuming only collision-resistant  hash functions against polynomial-time adversaries.  </paper_abstract>
<query_num> 5801 </query_num>
<text>   any property of pk. Consequently the BPK model is considered as a very weak set-up assumption compared to previously models such as common reference model and PKI model. However, as Micali and Reyzin =-=[18]-=- pointed out, the notion of soundness in this model is more subtle. There are four distinct notions of soundness: one time, sequential, concurrent and resettable soundness, each of which implies the p   </text>
<query_num> 5802 </query_num>
<text>   design of some cryptographic protocols. In recent years, the research is moving towards extending the security to cope with some more malicious communication environment. In particular, Dwork et al. =-=[12]-=-introduced the concept of concurrent zero knowledge, and initiate the study of the effect of executing ZK proofs concurrently in some realistic and asynchronous networks like the Internet. Though the   </text>
<query_num> 5803 </query_num>
<text>   eaders should keep in mind that the way to construct the OR-proof is also applied to 4-round Σ-protocol. Interestingly, Σ-protocols can be composed to proving the OR of atomic statements, as shown in =-=[8, 7]-=-. Specifically, given two protocols Σ0,Σ1 for two relationships R0, R1, respectively, we can construct a ΣOR-protocol for the following 5srelationship efficiently: ROR = ((x0, x1), y) : (x0, y) ∈ R0or   </text>
<query_num> 5804 </query_num>
<text>   ed concurrent zero knowledge protocol though it is very inefficient. Motivated by the application in which the prover (such as the user of a smart card) may encounter resetting attack, Canetti et al. =-=[4]-=- introduced the notion of resettable zero knowledge (rZK for short). An rZK formalizes security in a scenario in which the verifier is allowed to reset the prover in the middle of proof to any previou ger than that of concurrent zero knowledge and therefore we can not construct a constant round black-box rZK protocol in the plain model for non-trivial languages. To get constant round rZK, the work =-=[4]-=- also introduced a very attracting model, the bare public-key model(BPK). In this model, Each verifier deposits a public key pk in a public file and stores the associated secret key sk before any inte ich implies the previous one. Moreover they also pointed out that there is NO black-box rZK satisfying resettable soundness for non-trivial language and the original rZK arguments in the BPK model of =-=[4]-=- does not seem to be concurrently sound. The 4-round(optimal) rZK arguments with concurrent soundness in the bare publickey model was proposed by Di Crescenzo et al. in [10] and also appeared in [24]. icious prover P ∗ , the probability that in an execution of concurrent attack, V ever outputs ”accept x” for x /∈ L is negligible in n. The notion of resettable zero-knowledge was first introduced in =-=[4]-=-. The notion gives a verifier the ability to rewind the prover to a previous state (after rewinding the prover uses the same random bits), and the malicious verifier can generate an arbitrary file F w  adopt the 3-round ΣOR-protocol just for the sake of simplicity. Proof. Completeness. Straightforward. Resettable (black-box) Zero Knowledge. The analysis is very similar to the analysis presented in =-=[4, 10]-=-. Here we omit the tedious proof and just provide some intuition. As usual, we can construct a simulator Sim that extracts all secret keys corresponding to those public keys registered by the maliciou   </text>
<query_num> 5805 </query_num>
<text>   is a Σ-protocol for the language of Hamiltonian Graphs [1], assuming that one-way permutation families exists; if the commitment scheme used by the protocol in [1] is implemented using the scheme in =-=[19]-=- from any pseudo-random generator family, then the assumption can be reduced to the existence of one-way function families, at the cost of adding one preliminary message from the verifier. Note that a licious) sender S ∗ cannot open this commitment to another value m ′ �= m except with negligible probability. Under the assumption of existence of any one-way function families (using the scheme from =-=[19]-=- and the result from [17]) or under number-theoretic assumptions (e.g., the scheme from [21]), we can construct a schemes in which the first phase consists of 2 messages. Assuming the existence of one ng property: for all powerful sender S ∗ (without running time restriction), it cannot open a valid commitment to two different values except with exponentially small probability. We refer readers to =-=[13, 19]-=- for the details for constructing statistically-binding commitments. 6sA perfect-hiding commitment scheme (with computational binding) is the one except with a stronger requirement on hiding property: r security analysis can be also applied to this variant. We also note that collision-resistant hash functions implies one-way functions which suffices to build statistically-binding commitment scheme =-=[19]-=-(therefore computationalbinding scheme), thus, if we proved our protocol is a rZK argument with concurrent soundness, then we get theorem 1. Here we adopt the 3-round ΣOR-protocol just for the sake of   </text>
<query_num> 5806 </query_num>
<text>   ledge with concurrent soundness was presented in [11] under standard assumption (without using ”complexity leveraging”). For the sake of simplification, we modify the flawed construction presented in =-=[26]-=- to get concurrent zero knowledge argument with concurrent soundness. Considering the following twophase argument in BPK model: Let n be the security parameter, and f be a one way function that maps {   </text>
<query_num> 5807 </query_num>
<text>   of round inefficiency. In the Common Reference String model, Damgaard [6] showed that 3-round concurrent zero-knowledge can be achieved efficiently. Surprisingly, using non-black-box technique, Barak =-=[1]-=- constructed a constant round non-black-box bounded concurrent zero knowledge protocol though it is very inefficient. Motivated by the application in which the prover (such as the user of a smart card nscript between the honest P , V on input x. Many known efficient protocols, such as those in [16] and [23], are Σ-protocols. Furthermore, there is a Σ-protocol for the language of Hamiltonian Graphs =-=[1]-=-, assuming that one-way permutation families exists; if the commitment scheme used by the protocol in [1] is implemented using the scheme in [19] from any pseudo-random generator family, then the assu  observation (in a different context) that enables the analysis of concurrent non-malleability of their commitment scheme. Now we recall the Barak’s constant round public-coin zero knowledge argument =-=[1]-=-, and show this protocol satisfies one-many simulatability, and then so does the resettably-sound zero knowledge argument transformed from it. Informally, Barak’s protocol for a N P language L consist  ∈ Hn × {o, 1} n × {o, 1} n is in Λ, if there exist a program Π and a string s ∈ {0, 1} poly(n) such that z = C(h(Π), s) and Π(z) = r within superpolynomial time (i.e., n ω(1) ). The Barak’s Protocol =-=[1]-=- Common input: an instance x ∈ L (|x| = n) 2 Barak also presented a constant round bounded concurrent ZK arguments, hence we can obtain a constant round resettably-sound bounded concurrent ZK argument routine of V ∗ , and it interacts with V ∗ internally). We note that the next message of the joint residual code of V ∗ and Sreal is only determined by the commitment message from Sj, so as showed in =-=[1]-=-, Sj works. On the other hand, the Sreal’s behavior is identical to the honest provers. Thus, the whole simulator S satisfies our requirement. When we transform a constant round public-coin zero knowl   </text>
<query_num> 5808 </query_num>
<text>   reveals nothing but the validity of the assertion, is put forward in the seminal paper of Goldwasser, Micali and Rackoff [15]. Since its introduction, especially after the generality demonstrated in =-=[14]-=-, ZK proofs have become a fundamental tools in design of some cryptographic protocols. In recent years, the research is moving towards extending the security to cope with some more malicious communica   </text>
<query_num> 5809 </query_num>
<text>   t open this commitment to another value m ′ �= m except with negligible probability. Under the assumption of existence of any one-way function families (using the scheme from [19] and the result from =-=[17]-=-) or under number-theoretic assumptions (e.g., the scheme from [21]), we can construct a schemes in which the first phase consists of 2 messages. Assuming the existence of one-way permutation families   </text>
<query_num> 5810 </query_num>
<text>   tely, they requires 1slogarithmic rounds for languages outside BPP in the plain model for the blackbox case [5] and therefore are of round inefficiency. In the Common Reference String model, Damgaard =-=[6]-=- showed that 3-round concurrent zero-knowledge can be achieved efficiently. Surprisingly, using non-black-box technique, Barak [1] constructed a constant round non-black-box bounded concurrent zero kn   </text>
<query_num> 5811 </query_num>
<text>   thms 1 , but their protocol enjoys only sequential soundness. The existence of constant round rZK arguments with concurrent soundness in BPK model under only polynomial-time hardness 1 using idea from=-=[3]-=-, this results also holds under standard assumptions that there exist hash functions that are collision-resistent against all polynomial-time adversaries. 2sassumption is an interesting problem. Our r   </text>
<query_num> 5812 </query_num>
<text>   we just need to simulate only one execution among all concurrent executions of the resettably-sound zero knowledge argument. We call this property one-many simulatability. We note that Pass and Rosen =-=[22]-=- made a similar observation (in a different context) that enables the analysis of concurrent non-malleability of their commitment scheme. Now we recall the Barak’s constant round public-coin zero know   </text>
</top>
<top>
<paper_num> 59 </paper_num>
<paper_title>   Blobworld: A System for Region-Based Image Indexing and Retrieval.  </paper_title>
<paper_abstract>   Blobworld is a system for image retrieval based on finding coherent image regions which roughly correspond to objects. The image is segmented into regions by fitting a mixture of Gaussians to the pixel distribution in a joint color-texture-position feature space. Each region (&amp;quot;blob&amp;quot;) is then associated with color and texture descriptors. Querying is based on the user specifying attributes of one or two regions of interest, rather than a description of the entire image. In order to make largescale retrieval feasible, we index the blob descriptions using a tree. Because indexing in the high-dimensional feature space is computationally prohibitive, we use a lower-rank approximation to the high-dimensional distance. Experiments show encouraging results for both querying and indexing.  </paper_abstract>
<query_num> 5901 </query_num>
<text>   , allow the computer to find images relevant to a query without looking at every image in the database. We investigated indexing the color feature vectors to speed up atomic queries. We used R*-trees =-=[1]-=-, index structures for data representable as points in N-dimensional space. R*-trees are not the state of the art for nearest-neighbor search in multiple dimensions; using a newer tree [22, 24] would   </text>
<query_num> 5902 </query_num>
<text>   cing performance, and (ii) indices over blobs can do better clustering than whole-image indices. As we tune and scale the system, we intend to examine new indexing schemes. We used the GiST framework =-=[11]-=- to experiment with the indices. R*-trees break the multi-dimensional data space into smaller and smaller rectangles. Each node contains a list of the minimumbounding rectangles (MBRs) of and pointers   </text>
<query_num> 5903 </query_num>
<text>   d to Visual Information Systems &amp;apos;99. 2 In previous work we described &amp;quot;Blobworld,&amp;quot; a new framework for image retrieval based on segmentation into regions and querying using properties of these regions =-=[2, 3]-=-. These regions generally correspond to objects or parts of objects. In this paper we present a complete online system for retrieval in a collection of 10,000 Corel images using this approach. The seg   </text>
<query_num> 5904 </query_num>
<text>   d to Visual Information Systems &amp;apos;99. 2 In previous work we described &amp;quot;Blobworld,&amp;quot; a new framework for image retrieval based on segmentation into regions and querying using properties of these regions =-=[2, 3]-=-. These regions generally correspond to objects or parts of objects. In this paper we present a complete online system for retrieval in a collection of 10,000 Corel images using this approach. The seg r and texture. A blob is described by its color distribution and mean texture descriptors. Figure 1 illustrates the stages in creating Blobworld. Details of the segmentation algorithm may be found in =-=[2]-=-. raw pixel features regions image region features feature pixel vectors image group extract features features combine describe regions features Fig. 1. The stages of Blobworld processing: From pixels s to avoid oversegmentation arising from local color variations due to texture. The three texture features are contrast, anisotropy, and polarity, extracted at a scale which is selected automatically =-=[2]-=-. The position features are simply the (x; y) position of the pixel; including the position generally decreases oversegmentation and leads to smoother regions. We model the distribution of pixels in t   </text>
<query_num> 5905 </query_num>
<text>   ieve images based on spatial and photometric relationships within and across simple image regions. Little or no segmentation is done; the regions are derived from low-resolution images. Jacobs et al. =-=[13]-=- use multiresolution wavelet decompositions to perform queries based on iconic matching. Ma and Manjunath [17] perform retrieval based on segmented image regions. Their segmentation is not fully autom   </text>
<query_num> 5906 </query_num>
<text>   o segmentation is done; the regions are derived from low-resolution images. Jacobs et al. [13] use multiresolution wavelet decompositions to perform queries based on iconic matching. Ma and Manjunath =-=[17]-=- perform retrieval based on segmented image regions. Their segmentation is not fully automatic, as it requires some parameter tuning and hand pruning of regions. Much research has gone into dimensiona   </text>
<query_num> 5907 </query_num>
<text>   one. We built indices over the color feature vectors of the blobs as well as over color feature vectors based on global histograms. We measured the recall of the indices using nearest-neighbor search =-=[12]-=- to retrieve and rank images against the top 40 images retrieved by a full Blobworld query or global histogram query over all the images. Figure 4 displays results for indices of various dimensions. 0   </text>
<query_num> 5908 </query_num>
<text>   results. 1.1 Related Work Many current image retrieval systems perform retrieval based primarily on lowlevel image features, including IBM&amp;apos;s Query by Image Content (QBIC) [6], Photobook [19], Virage =-=[9]-=-, VisualSEEk [23], Candid [15], and Chabot [18]. Lipson et al. [16] retrieve images based on spatial and photometric relationships within and across simple image regions. Little or no segmentation is   </text>
<query_num> 5909 </query_num>
<text>   retrieval systems perform retrieval based primarily on lowlevel image features, including IBM&amp;apos;s Query by Image Content (QBIC) [6], Photobook [19], Virage [9], VisualSEEk [23], Candid [15], and Chabot =-=[18]-=-. Lipson et al. [16] retrieve images based on spatial and photometric relationships within and across simple image regions. Little or no segmentation is done; the regions are derived from low-resoluti   </text>
<query_num> 5910 </query_num>
<text>   ussion of our results. 1.1 Related Work Many current image retrieval systems perform retrieval based primarily on lowlevel image features, including IBM&amp;apos;s Query by Image Content (QBIC) [6], Photobook =-=[19]-=-, Virage [9], VisualSEEk [23], Candid [15], and Chabot [18]. Lipson et al. [16] retrieve images based on spatial and photometric relationships within and across simple image regions. Little or no segm   </text>
</top>
<top>
<paper_num> 60 </paper_num>
<paper_title>   Towards a More Efficient Evolutionary Induction of Bayesian Networks.  </paper_title>
<paper_abstract>   Abstract. Bayesian networks (BNs) constitute a useful tool to model the joint distribution of a set of random variables of interest. This paper is concerned with the network induction problem. We propose a number of hybrid recombination operators for extracting BNs from data. These hybrid operators make use of phenotypic information in order to guide the processing of information during recombination. The performance of these new operators is analyzed with respect to that of their genotypic counterparts. It is shown that these hybrid operators provide notably improved and rather robust results. Some remarks on the future of the area are also laid out. 1  </paper_abstract>
<query_num> 6001 </query_num>
<text>   (a variety of methods can be used to learn the probabilities θ). This turns out to be NP -hard [6], and hence the use of heuristic algorithms is in order [11]. In this sense, evolutionary algorithms =-=[2]-=- (EAs) emerge as interesting candidatessfor this task. Here we concentrate on the use of EAs for BN induction. More precisely, we explore in detail the role of recombination for this purpose. The orga   </text>
<query_num> 6002 </query_num>
<text>   conduct the search. Elements from Saux are then fed to a suitable (decoder) algorithm to obtain the actual BNs they represent. Consider, for example, the search in the space of n−element permutations =-=[13]-=-; the construction heuristic K2 [7] is subsequently used to build the actual BN. This approach has the advantage of filtering out infeasible solutions while it also introduces problem-specific knowled   </text>
<query_num> 6003 </query_num>
<text>   g = �n i=1 (ri − 1)qi is the number of free θ parameters in the model. This choice penalizes complex (i.e., highly dense) DAGs, and is closely related to the asymptotic Bayesian Information Criterion =-=[10]-=- or BIC 1 . As to the π(θ|G), it is taken to be the product of independent (conjugate) Dirichlet distributions. In the case of no missing data and noninformative Dirichlet hyperparameters α, it is the   </text>
<query_num> 6004 </query_num>
<text>   in feasible DAGs (alternatively, closed operators in SDAG can be defined; we will return to this point below). On the other hand, a repair function can be used to remove cycles before evaluation. See =-=[14]-=- for a comparison of both approaches. As regards indirect approaches, these use an auxiliary space Saux to conduct the search. Elements from Saux are then fed to a suitable (decoder) algorithm to obta   </text>
<query_num> 6005 </query_num>
<text>   lude by briefly providing some preliminary insights on this matter. It turns out that equivalence classes can be compactly represented by (certain class of) partially directed acyclic graphs or PDAGs =-=[1, 5]-=-. PDAGs include directed as well as undirected arcs. Chickering [5] provides an algorithm that takes a given DAG G and outputs the PDAG ¯G that uniquely represents its equivalence class [G]. Since ¯G  1. To conclude, consider now potential mutation and crossover operators for some parent PDAGs ¯G and ¯H. A first issue refers to PDAG validity: not all PDAGs represent equivalence classes. Chickering =-=[5]-=- presents various operators designed to modify a given ¯G so that the resulting PDAG effectively represents a different equivalence class. For example, both directed and undirected arcs can be added o   </text>
<query_num> 6006 </query_num>
<text>   on problem is learning the structure or DAG G (a variety of methods can be used to learn the probabilities θ). This turns out to be NP -hard [6], and hence the use of heuristic algorithms is in order =-=[11]-=-. In this sense, evolutionary algorithms [2] (EAs) emerge as interesting candidatessfor this task. Here we concentrate on the use of EAs for BN induction. More precisely, we explore in detail the role g to P (D|G) above. This key result adds to the computational tractability of the approach and will be considered here too. The hyperparameters α can be interpreted in terms of equivalent sample size =-=[11]-=-. The noninformative choice αi = 1 is usually adopted ri following theoretical considerations related to likelihood equivalence [12]. 2.2 Evolutionary Induction of DAGs Typically, the EA approach for   </text>
<query_num> 6007 </query_num>
<text>   rparameters α can be interpreted in terms of equivalent sample size [11]. The noninformative choice αi = 1 is usually adopted ri following theoretical considerations related to likelihood equivalence =-=[12]-=-. 2.2 Evolutionary Induction of DAGs Typically, the EA approach for designing BNs evolves DAG structures which –when submitted for fitness calculation– are augmented with ˆ θ parameters and fed into a   </text>
<query_num> 6008 </query_num>
<text>   the proposed approach: the ALARM network, a 37-variable network for monitoring patients in the intensive care unit [3], and the INSURANCE network, a 27-variable BN for evaluating car insurance risks =-=[4]-=-. However, due to space constraints, we concentrate here on the former (similar qualitative results have been obtained with the latter). Training sets of N = 2, 000 examples were simulated from the AL   </text>
<query_num> 6009 </query_num>
<text>   utual information MI(Xj, Xi) criterion has often been the choice for measuring the merit of single alleles η1 ij [16]. However, this measure is known to have some limitations due to its myopic nature =-=[9]-=-. The updated MI measure, namely the Conditional Mutual Information measure [9] CMI(Xj, Xi || Πj \ {Xi}) = � P (Πj \ {Xi}) � P (Xj, Xi | Πj \ {Xi}) log P (Xj, Xi | Πj \ {Xi}) P (Xj | Πj \ {Xi})P (Xi |   </text>
<query_num> 6010 </query_num>
<text>   ux are then fed to a suitable (decoder) algorithm to obtain the actual BNs they represent. Consider, for example, the search in the space of n−element permutations [13]; the construction heuristic K2 =-=[7]-=- is subsequently used to build the actual BN. This approach has the advantage of filtering out infeasible solutions while it also introduces problem-specific knowledge. There are some drawbacks though   </text>
</top>
<top>
<paper_num> 61 </paper_num>
<paper_title>   Relaxed Atomic Broadcast: State-Machine Replication Using Bounded Memory.  </paper_title>
<paper_abstract>   Atomic broadcast is a useful abstraction for implementing fault-tolerant distributed applications such as statemachine replication. Although a number of algorithms solving atomic broadcast have been published, the problem of bounding the memory used by these algorithms has not been given the attention it deserves. It is indeed impossible to solve repeated atomic broadcast with bounded memory in a system (non synchronous or not equipped with a perfect failure detector) in which consensus is solvable with bounded memory. The intuition behind this impossibility is the inability to safely garbage-collect unacknowledged messages, since a sender process cannot tell whether the destination process has crashed or is just slow. The usual technique to cope with this problem is to introduce  </paper_abstract>
<query_num> 6101 </query_num>
<text>   -tolerant distributed services [3] using the state-machine approach [21]. A number of different implementations of atomic broadcast have been proposed in the literature for a variety of system models =-=[7]-=-. However, they rarely tackle the problem of bounding the use of memory. The fact that an algorithm needs a potentially unbounded amount of buffers is often considered as a minor (implementation) issu   </text>
<query_num> 6102 </query_num>
<text>   consensus algorithm: Algorithm 3 is the consensus algorithm we consider [6]. The algorithm requires f &amp;lt; n/3. We have chosen this algorithm because of its simplicity. The analysis of Paxos/LastVoting =-=[13, 6]-=-, which requires only f &amp;lt; n/2 could be used instead, but would require more space. Algorithm 3 works as follows. As soon as more than 2n/3 processes have xp = v, then decision v is locked, i.e., in an   </text>
<query_num> 6103 </query_num>
<text>   ction 8 concludes the paper. 2 System Model We consider a system with a finite set of processes Π = {p1, p2, . . . , pn} that communicate by message exchange. We assume a partially synchronous system =-=[9]-=-, where after some unknown time GST (Global Stabilization Time) the system (both processes and channels) becomes synchronous and channels reliable. 3 Before GST the system is asyn1 This is called outp  if more than 2n/3 values received are equal to ¯x then 13: DECIDE(¯x) 76.4.1 Algorithm Round-based model: We consider a consensus algorithm for a partially synchronous system (see Section 2). As in =-=[9]-=-, we consider an abstraction on top of the system model, namely a round model. Using this abstraction, rather than the raw system model, improves the clarity of the algorithms and simplifies the proof   </text>
<query_num> 6104 </query_num>
<text>   e consensus memory bounds. 6.5 Implementation of the Round-Based Model We describe now the implementation of the round-based model (see Algorithm 4), which is almost identical to the one appearing in =-=[12]-=- (we made small extensions to bound the memory needed). The interaction between Algorithm 4 and Algorithm 3 is by function call: in other words, the execution thread is within Algorithm 4, and this th  14-15 (time is measured by the execution of receive steps: 1 receive step = 1 time unit), or (ii) whenever a message of a round larger than rp is received, see lines 20-21. The reader is referred to =-=[12]-=- for a proof that this ensures Pgood after GST. When the inner loop ends, the function T r p is called with the set of messages received 5 To be consistent, line 4 of Algorithm 3 should be expressed a   </text>
<query_num> 6105 </query_num>
<text>   es not allow us to distinguish a slow process from a crashed process, the ability of atomic broadcast algorithms to bound their memory – without affecting correctness – becomes challenging. Ricciardi =-=[18]-=- proved that a primitive as basic as (repeated) reliable broadcast cannot be implemented in a system with message losses in which slow processes are indistinguishable from crashed processes. Trivially n-uniform version of 3.3. As shown by Ricciardi, repeated reliable broadcast cannot be implemented in a system with message losses in which slow processes are indistinguishable from crashed processes =-=[18]-=-. The intuition behind this impossibility result is the following. Consider a sender process p, and its output buffer to q that contains bad periods (system is asynchronous and channels are lossy). Th   </text>
<query_num> 6106 </query_num>
<text>   his is done, all unacknowledged messages to p3 can be discarded. However, the dynamic model is not so straightforward as the static one: protocol specifications and implementations have to be revised =-=[19]-=- and are more complex. Besides, a membership service is needed, and the application logic needs to become aware of view changes. Relaxing the Specification of Atomic Broadcast: The paper proposes anot   </text>
<query_num> 6107 </query_num>
<text>   his is long enough from a practical point of view (see also [10]). 6.3 Relaxed Atomic Broadcast 6.3.1 Algorithm Algorithm 2 implements relaxed atomic broadcast by reduction to a sequence of consensus =-=[4]-=-. However, contrary to [4], each consensus decides only on one single message (in order to bound memory) rather than on a batch of messages. Although a number of optimizations can be performed, we hav kp a sender cp is designated in a round-robin manner, with the goal to propose Rcvp[cp] as the initial value for consensus (line 42). This initial value could be optimized to be the whole Rcvp vector =-=[4]-=-, but the rotating sender approach makes it easier to present both our algorithm and its memory bounds. When consensus #kp decides, p waits for evidence that at least f + 1 other processes have also d   </text>
<query_num> 6108 </query_num>
<text>   ic group; approach (1) performs a state transfer whenever a slow process catches up. Solution (2) is more complex than solution (1). First, solution (2) needs to define a policy for process exclusion =-=[20]-=-. This is simply not needed in (1). Second, static group communication is simpler and easier to understand than dynamic group communication, from a specification as well as from an implementation poin   </text>
<query_num> 6109 </query_num>
<text>   ning the specification of atomic broadcast. Note that (one instance of) consensus has been shown to be solvable with bounded memory [10] in an asynchronous system with the ⋄S failure detector, and in =-=[8]-=- Delporte-Gallet et al. show that solving (repeated) reliable broadcast requires indeed a stronger failure detector than solving (one instance of) consensus. Group communication prototypes built in th   </text>
<query_num> 6110 </query_num>
<text>   te-machine replication. 1 Introduction and Related Work Atomic broadcast has been proposed as the key abstraction to implement fault-tolerant distributed services [3] using the state-machine approach =-=[21]-=-. A number of different implementations of atomic broadcast have been proposed in the literature for a variety of system models [7]. However, they rarely tackle the problem of bounding the use of memo o repeated atomic broadcast, since atomic broadcast is strictly stronger than reliable broadcast. 4 How to Deal with Finite Memory Consider atomic broadcast used to implement statemachine replication =-=[21]-=- in a system with three processes (n = 3). Process p1, which receives clients’ requests, issues abcasts. Assume that the adelivery of these messages requires the cooperation of p1 with only p2 or with   </text>
<query_num> 6111 </query_num>
<text>   time GST (Global Stabilization Time) the system (both processes and channels) becomes synchronous and channels reliable. 3 Before GST the system is asyn1 This is called output triggered suspicions in =-=[5]-=-. 2 The size of the application state is controlled (and bounded) by the application. This is different from the state required for the implementation of atomic broadcast, which cannot be controlled b  or is just slow (or connected through a slow link), then p cannot safely dispose of unacknowledged messages sent to q. However, if q has crashed, the set of unacknowledged messages will grow forever =-=[5]-=-. The impossibility of repeated reliable broadcast also applies to repeated atomic broadcast, since atomic broadcast is strictly stronger than reliable broadcast. 4 How to Deal with Finite Memory Cons   </text>
<query_num> 6112 </query_num>
<text>   w infinitely. We now present two approaches to deal with this problem. The Dynamic Model: The traditional solution to bound memory consists in switching to the dynamic system (or dynamic group) model =-=[2, 1, 15, 16, 11]-=-. 4 In such a model processes can be added/removed to/from the system (or group) on the fly. In a dynamic model, a view describes the set of processes that are currently part of the system (or group). bounded memory. Namely, (1) our novel relaxed atomic broadcast algorithm, which was described in detail in Sections 5 and 6, and (2) atomic broadcast in the dynamic model, i.e., relying on membership =-=[2]-=-. Both approaches rely on state transfer: approach (2) requires a state transfer whenever a new process is added to the dynamic group; approach (1) performs a state transfer whenever a slow process ca   </text>
</top>
<top>
<paper_num> 62 </paper_num>
<paper_title>   Fourier Domain Scoring: A Novel Document Ranking Method.  </paper_title>
<paper_abstract>   ranking method  </paper_abstract>
<query_num> 6201 </query_num>
<text>   distances. FDS could easily be implemented in the Google search engine to store the precalculated phases of each word. Requiring less space and less time for calculations at query time. Ebert et al. =-=[13]-=- describe a method of visualizing clusters in documents. This system takes sequences of characters from the text to create vectors for the user to explore. It would be very interesting to replace thes   </text>
<query_num> 6202 </query_num>
<text>   e document. Many information retrieval systems employ the use of a vector space model to classify text. Some examples of current work which use the vector space model are SMART by Buckley et al. [7], =-=[8]-=-, the Okapi Basic Search System by Robertson et al. [9], IRIS by Yang et al. [10] and INQUERY by Allan et al. [11]. The problem with these techniques is that any spatial information contained in the d   </text>
<query_num> 6203 </query_num>
<text>   engine is critical. By obtaining a clear understanding of a page, we are able to find more relevant documents and make it harder for people to cheat the system. Some variant of the vector space model =-=[6]-=- has been the dominant method used to analyse the text in information retrieval systems for many years. The concept behind the vector space model is to convert each document into a vector, so they can  improved considerably by adding weighting to the document vectors before the score is calculated [16]. The weighting which we will be using when performing the cosine measure (shown in Witten et al. =-=[6]-=-) is the typical TF×IDF rule: Scos (d,T ) = � 1 � (1 + loge fd,t) · loge 1 + WdWt t∈T ∩d N � ft where the Scos (d,T ) is the score given to the dth document using query terms T , the TF (term frequenc   </text>
<query_num> 6204 </query_num>
<text>   ining the spatial information found in the documents (since FDS 3-4-1 uses a variant of TF×IDF as a preweighting scheme). The BD-ACI-BCA method is compared because it is suggested by Zobel and Moffat =-=[18]-=- that this method is generally the best of the vector space similarity measures. The results show us that the methods 3.4.x and 4.3.2 are superior amongst the FDS methods for this document set. These   </text>
<query_num> 6205 </query_num>
<text>   ples of current work which use the vector space model are SMART by Buckley et al. [7], [8], the Okapi Basic Search System by Robertson et al. [9], IRIS by Yang et al. [10] and INQUERY by Allan et al. =-=[11]-=-. The problem with these techniques is that any spatial information contained in the documents is lost. Once the documents are converted into document vectors, the number of times each word appears is   </text>
<query_num> 6206 </query_num>
<text>   ps of FDS are: • Collect words into spatial bins • Create inverted index • Perform preweighting • Perform Fourier transform • Calculate document scores Some initial work on this topic can be found in =-=[15]-=-. We discuss each of these steps in brief as it is essential to understand how FDS works before we can examine any of its properties. A. Collect words into spatial bins Rather than mapping a document  h the cosine measure with TF×IDF weighting. VI. EXPERIMENTS Initial results have shown that FDS improves the accuracy of search results on small document sets (containing about 100 to 1000 documents) =-=[15]-=-. In these experiments we will show that FDS can effectively improve the results of large document sets. By doing so, we will also show that using FDS in a Web search engine would be a great enhanceme   </text>
<query_num> 6207 </query_num>
<text>   risen. There have been many methods for seeking out information on the Web. Many new techniques involve utilizing the HTML tags found on Web pages to obtain a higher understanding of the page content =-=[1]-=-, [2], [3], [4], [5]. There have also been many page creators who have abused searching methods to obtain higher rankings in search engine results. A few search engines have branched out to analyzing   </text>
<query_num> 6208 </query_num>
<text>   the Web, using a thesaurus,. . . ). Cases of systems which use the vector space similarity measures which could easily be replaced by FDS can be found everywhere. Some examples are as follows. Google =-=[12]-=- records the position of every word in the document to give a higher ranking to document which contain query terms closer together. To calculate the distances between every term would require a lot of   </text>
<query_num> 6209 </query_num>
<text>   ve been many methods for seeking out information on the Web. Many new techniques involve utilizing the HTML tags found on Web pages to obtain a higher understanding of the page content [1], [2], [3], =-=[4]-=-, [5]. There have also been many page creators who have abused searching methods to obtain higher rankings in search engine results. A few search engines have branched out to analyzing pages in the fo   </text>
</top>
<top>
<paper_num> 63 </paper_num>
<paper_title>   Broadcast on Clusters of SMPs with Optimal Concurrency.  </paper_title>
<paper_abstract>   In this paper, we present a hierarchical method for broadcast on clusters of symmetric multiprocessors (CSMPs) connected by switches with one-port model. We focus on the inter-switch broadcast that forms the core part of a broadcast on CSMPs. The proposed broadcast method is based on single-source shortest path minimum-cost spanning tree (SSS-MST). Two heuristic algorithms, from-up-to-down and fromdown-to-up, are proposed to achieve the maximum concurrency using the information of the underlying network topology and the costs of links. Performance evaluation is also conducted to show the superiority of the proposed methods. I.  </paper_abstract>
<query_num> 6301 </query_num>
<text>   ludes the broadcast in a single SMP node (Level 0), intra-switch broadcast (Level 1), inter-switch broadcast (Level 2). The proposed hierarchical broadcast covers the two important techniques used in =-=[MSP99]-=-: switch-ordered ring and link scheduling. The essence in the switch-ordered ring technique is to order the processors by switch to eliminate potential link contention. For each switch, only one local der processors (leader SMPs) can communicate with other SMPs (switches). The proposed method puts no restriction on the underlying topology for inter-switch broadcast except the switchordered ring in =-=[MSP99]-=-. The topology for interswitch broadcast may be arbitrary (e.g., ring and tree). In the broadcast hierarchy, Level 0 and Level 1 can be implemented easily in hardware in SMPs and switches respectively ension-ordered routing of unicast messages. The algorithm cannot guarantee the minimum time on either other regular topologies such as Torus and star graphs or irregular topologies. Jacunski et al in =-=[MSP99]-=- presents an all-to-all broadcast on switch-based clusters of workstations. The switch-order ring and link scheduling techniques are proposed to approach to the optimal properties such as the degree o   </text>
<query_num> 6302 </query_num>
<text>   sed to designs various broadcast algorithms in many different interconnects, such as hypercube [JH89], and 2D-mesh [BMT92], wormhole-routed networks [MXEN94], Myrinet [BPDS00], and arbitrary topology =-=[RM99]-=-. Distinguished from the existing tree-based broadcasting algorithms, the proposed broadcast algorithm aims at achieving the maximum step concurrency, minimum steps of message passing among switches,  s not contention-free for the usage of the links. Buntinas et al in [BPDS00] constructs optimal multicast trees by a simple top-down greedy algorithm under the postal model. Raghavan and manimaran in =-=[RM99]-=- proposes a re-arrangeable algorithm for the construction of delay-constrained dynamic multicast trees. The method uses the concept of quality factor to describe the usefulness of a portion of the mul   </text>
<query_num> 6303 </query_num>
<text>   treebased techniques for broadcast have been proposed to designs various broadcast algorithms in many different interconnects, such as hypercube =-=[JH89]-=-, and 2D-mesh =-=[BMT92]-=-, wormhole-routed networks =-=[MXEN94]-=-, Myrinet =-=[BPDS00]-=-, and arbitrary topology =-=[RM99]-=-. Distinguished from the existing tree-based broadcasting algorithms, the proposed broadcast algorithm aims at achieving the maximum step concurrency,  erarchical Broadcast and MPI Broadcast 0.14 0.12 0.10 0.08 0.06 0.04 4 proc_num vs syzh proc_num vs MPI 8 10 12 14 16 18 20 The Num ber of SMP Nodes 22 36 32 24 56 40 26 38 44 64 28sMcKinley et al in =-=[MXEN94]-=- implemented multicast communication in wormhole-routed direct interconnects in the absence of hardware multicast support. The method exploits the properties of the switching technology on ndimensiona   </text>
<query_num> 6304 </query_num>
<text>   ues for broadcast have been proposed to designs various broadcast algorithms in many different interconnects, such as hypercube =-=[JH89]-=-, and 2D-mesh =-=[BMT92]-=-, wormhole-routed networks =-=[MXEN94]-=-, Myrinet =-=[BPDS00]-=-, and arbitrary topology =-=[RM99]-=-. Distinguished from the existing tree-based broadcasting algorithms, the proposed broadcast algorithm aims at achieving the maximum step concurrency, minimum steps of m s to a set of switches on the CSMP, and E corresponds to a set of links between any pair of switches. Each link may consist of a couple of channels that may simultaneously transfer different messages =-=[BPDS00]-=-. The three-level hierarchy of a broadcast in CSMPs includes the broadcast in a single SMP node (Level 0), intra-switch broadcast (Level 1), inter-switch broadcast (Level 2). The proposed hierarchical ch as the degree of pipelining of communication components, the minimal transmission latencies and minimal the node. The algorithm is not contention-free for the usage of the links. Buntinas et al in =-=[BPDS00]-=- constructs optimal multicast trees by a simple top-down greedy algorithm under the postal model. Raghavan and manimaran in [RM99] proposes a re-arrangeable algorithm for the construction of delay-con   </text>
</top>
<top>
<paper_num> 64 </paper_num>
<paper_title>   A tangible interface for organizing information using a grid.  </paper_title>
<paper_abstract>   The task of organizing information is typically performed either by physically manipulating note cards or sticky notes or by arranging icons on a computer with a graphical user interface. We present a new tangible interface platform for manipulating discrete pieces of abstract information, which attempts to combine the benefits of each of these two alternatives into a single system. We developed interaction techniques and an example application for organizing conference papers. We assessed the effectiveness of our system by experimentally comparing it to both graphical and paper interfaces. The results suggest that our tangible interface can provide a more effective means of organizing, grouping, and manipulating data than either physical operations or graphical computer interaction alone. notecards on a desk or, by collaboratively arranging sticky notes on a board. Such arrangement often begins in a freeform way, by accreting small groups of related items, and, later, develops into a larger structure or framework. Tasks like this have thus far been surprisingly resistant to computer support, perhaps because notecards or sticky notes allow manipulation that is more natural and fluid and particularly, free form without a predefined framework. Even when the information to be organized already exists in electronic form and the final output must be produced in digital form, many people find it advantageous to copy the information onto pieces of paper, manipulate them manually, and then re-enter the resulting organization into the computer.  </paper_abstract>
<query_num> 6401 </query_num>
<text>   FACES Tangible user interfaces are a growing area of user interface research that use physical forms to represent data. They were so named by Ishii and Ullmer[3], with roots in seminal work of Wellner=-=[9]-=-, and Fitzmaurice[1], and continue a growing range of more recent work[2, 7, 8, 10, 11, 14, 15]. Instead of a generic screen, mouse, and keyboard capable of representing all types of data, TUI uses sp  receiving input from the board over a serial port and sending its output to the video projector. Our system combines projection with physical manipulation in ways similar to the early work of Wellner=-=[9]-=- and Fitzmaurice[1] and more recent work such as DataTiles[11], metaDESK[14], and Sensetable[10]. Unlike several previous devices that used computer vision, the Senseboard uses RFID sensing, which pro   </text>
<query_num> 6402 </query_num>
<text>   devices that used computer vision, the Senseboard uses RFID sensing, which provides greater reliability and speed. Other related projects include LiveBoard[6], LegoWall (described in [2]), and Outpost=-=[5]-=-. Our device is also unusual in that it supports discrete, semistructured interaction, since the pucks can only be placed directly in grid cells. It is best used for tasks where completely free-form i   </text>
<query_num> 6403 </query_num>
<text>   interfaces are a growing area of user interface research that use physical forms to represent data. They were so named by Ishii and Ullmer[3], with roots in seminal work of Wellner[9], and Fitzmaurice=-=[1]-=-, and continue a growing range of more recent work[2, 7, 8, 10, 11, 14, 15]. Instead of a generic screen, mouse, and keyboard capable of representing all types of data, TUI uses specific physical form om the board over a serial port and sending its output to the video projector. Our system combines projection with physical manipulation in ways similar to the early work of Wellner[9] and Fitzmaurice=-=[1]-=- and more recent work such as DataTiles[11], metaDESK[14], and Sensetable[10]. Unlike several previous devices that used computer vision, the Senseboard uses RFID sensing, which provides greater relia cting using a mouse could be viewed as the opposite extreme, in that there is only one physical “puck,” and it can be attached to only one data object at a time, usually for very brief periods. Bricks=-=[1]-=- provides an intermediate point. We also exploit physical representation, with specially shaped physical objects to depict commands, and use purely digital representation where appropriate, such as co   </text>
<query_num> 6404 </query_num>
<text>   outweighs the former. TANGIBLE USER INTERFACES Tangible user interfaces are a growing area of user interface research that use physical forms to represent data. They were so named by Ishii and Ullmer=-=[3]-=-, with roots in seminal work of Wellner[9], and Fitzmaurice[1], and continue a growing range of more recent work[2, 7, 8, 10, 11, 14, 15]. Instead of a generic screen, mouse, and keyboard capable of r   </text>
<query_num> 6405 </query_num>
<text>   rch that use physical forms to represent data. They were so named by Ishii and Ullmer[3], with roots in seminal work of Wellner[9], and Fitzmaurice[1], and continue a growing range of more recent work=-=[2, 7, 8, 10, 11, 14, 15]-=-. Instead of a generic screen, mouse, and keyboard capable of representing all types of data, TUI uses specific physical forms to represent and manipulate the pieces of data in the system. TUI often u   </text>
<query_num> 6406 </query_num>
<text>   rch that use physical forms to represent data. They were so named by Ishii and Ullmer[3], with roots in seminal work of Wellner[9], and Fitzmaurice[1], and continue a growing range of more recent work=-=[2, 7, 8, 10, 11, 14, 15]-=-. Instead of a generic screen, mouse, and keyboard capable of representing all types of data, TUI uses specific physical forms to represent and manipulate the pieces of data in the system. TUI often u r. Our system combines projection with physical manipulation in ways similar to the early work of Wellner[9] and Fitzmaurice[1] and more recent work such as DataTiles[11], metaDESK[14], and Sensetable=-=[10]-=-. Unlike several previous devices that used computer vision, the Senseboard uses RFID sensing, which provides greater reliability and speed. Other related projects include LiveBoard[6], LegoWall (desc   </text>
<query_num> 6407 </query_num>
<text>   rch that use physical forms to represent data. They were so named by Ishii and Ullmer[3], with roots in seminal work of Wellner[9], and Fitzmaurice[1], and continue a growing range of more recent work=-=[2, 7, 8, 10, 11, 14, 15]-=-. Instead of a generic screen, mouse, and keyboard capable of representing all types of data, TUI uses specific physical forms to represent and manipulate the pieces of data in the system. TUI often u several previous devices that used computer vision, the Senseboard uses RFID sensing, which provides greater reliability and speed. Other related projects include LiveBoard[6], LegoWall (described in =-=[2]-=-), and Outpost[5]. Our device is also unusual in that it supports discrete, semistructured interaction, since the pucks can only be placed directly in grid cells. It is best used for tasks where compl   </text>
<query_num> 6408 </query_num>
<text>   rch that use physical forms to represent data. They were so named by Ishii and Ullmer[3], with roots in seminal work of Wellner[9], and Fitzmaurice[1], and continue a growing range of more recent work=-=[2, 7, 8, 10, 11, 14, 15]-=-. Instead of a generic screen, mouse, and keyboard capable of representing all types of data, TUI uses specific physical forms to represent and manipulate the pieces of data in the system. TUI often u ysical world to figure out how to operate them and what they mean. TUI also often involves the augmentation of existing physical objects by adding digital meaning to the objects and their manipulation=-=[11, 16]-=-. Some information and state is thus represented directly by the physical objects, while additional information is provided digitally, typically by video projection onto the physical objects (e.g., a  g its output to the video projector. Our system combines projection with physical manipulation in ways similar to the early work of Wellner[9] and Fitzmaurice[1] and more recent work such as DataTiles=-=[11]-=-, metaDESK[14], and Sensetable[10]. Unlike several previous devices that used computer vision, the Senseboard uses RFID sensing, which provides greater reliability and speed. Other related projects in   </text>
<query_num> 6409 </query_num>
<text>   ysical world to figure out how to operate them and what they mean. TUI also often involves the augmentation of existing physical objects by adding digital meaning to the objects and their manipulation=-=[11, 16]-=-. Some information and state is thus represented directly by the physical objects, while additional information is provided digitally, typically by video projection onto the physical objects (e.g., a   </text>
</top>
<top>
<paper_num> 65 </paper_num>
<paper_title>   BANANAS: an evolutionary framework for explicit and multipath routing in the internet.  </paper_title>
<paper_abstract>   Today the Internet offers a single path between end-systems even though it intrinsically has a large multiplicity of paths. This paper proposes an evolutionary architectural framework “BANANAS ” aimed at simplifying the introduction of multipath routing in the Internet. The framework starts with the observation that a path can be encoded as a short hash (“PathID”) of a sequence of globally known identifiers. The PathID therefore has global significance (unlike MPLS or ATM labels). This property allows multipath capable nodes to autonomously compute PathIDs in a partially upgraded network without requiring an explicit signaling protocol for path setup. We show that this framework allows the introduction of sophisticated explicit routing and multipath capabilities within the context of widely deployed connectionless routing protocols (e.g. OSPF, IS-IS, BGP) or overlay networks. We establish these characteristics through the development of PathID encoding and routecomputation schemes. The BANANAS framework also allows considerable flexibility in terms of architectural function placement and complexity management. To illustrate this feature, we develop an efficient variable-length hashing scheme that moves control-plane complexity and state overheads to network edges, allowing a very simple interior node design. All the schemes have been evaluated using both sizable SSFNet simulations and Linux/Zebra implementation evaluated on Utah’s Emulab testbed facility. 1.  </paper_abstract>
<query_num> 6501 </query_num>
<text>   MULATION RESULTS In this section, we illustrate the working of the proposed framework. We have implemented the BANANAS framework schemes in the Linux kernel: we use MIT’s Click Modular Router package =-=[34]-=- (data-plane) and GNU Zebra routing sofware version 0.92a [35] (control-plane). These implementations are tested on Utah’s Emulab testbed [36] to emulate sizable topologies running real implementation stbed with the Linux Zebra version 0.92a of OSPF (i.e. control-plane) upgraded with our BANANAS building blocks. The forwarding plane was implemented in Linux using MIT’s Click Modular Router package =-=[34]-=-. Note that this is a partially upgraded network: only nodes 1 and 2 (the dark colored nodes) are upgraded in this configuration. Figure 7 also indicates the IP addresses of various router interfaces   </text>
<query_num> 6502 </query_num>
<text>   ate complexity in comparison to MPLS label tables. Our computations can also be further optimized using incremental k-shortest path algorithms similar to those suggested for OSPF’s Dijkstra algorithm =-=[42, 43]-=-. In LIRA [11], Stoica et al briefly propose a forwarding scheme which they suggest could replace MPLS. A path is 8. SUMMARY AND CONCLUDING REMARKS The key contributions in this paper can be summarize   </text>
<query_num> 6503 </query_num>
<text>   ations in sizable topologies and Linux/Zebra implementation run on Utah’s Emulab emulation testbed facility. We are currently deploying the BANANAS framework on the worldwide PlanetLab infrastructure =-=[22]-=- as an public experimental wide-area network overlay service. We are 3 E.g. Packets from TCP connections would be mapped single “path” to avoid out-of-order packets also building a medium-sized multi-   </text>
<query_num> 6504 </query_num>
<text>   d last-mile multi-hop fixed-wireless networks. The answer to the second question is clearly not the lack of algorithms and protocols. There have been several proposals for multipath route-computation =-=[5, 6, 7, 8]-=-, Internet signaling architectures [9, 10, 11, 12, 13], novel overlay routing methods [14, 15] and transport-level approaches for multihomed hosts [16, 17]. The fact that these developments have not t multipath routing is used (e.g. MPLS-based [40, 41]). Protocol extensions to support multipath routing (both in RIP and OSPF) have been studied by Narvaez et al [7], Chen et al [6] and Vutukury et al =-=[8]-=-. In [7], authors propose to find loop-free multipaths only by concatenating the shortest paths of their neighbors with their link to the neighbors. This approach essentially uses a depth first search   </text>
<query_num> 6505 </query_num>
<text>   d last-mile multi-hop fixed-wireless networks. The answer to the second question is clearly not the lack of algorithms and protocols. There have been several proposals for multipath route-computation =-=[5, 6, 7, 8]-=-, Internet signaling architectures [9, 10, 11, 12, 13], novel overlay routing methods [14, 15] and transport-level approaches for multihomed hosts [16, 17]. The fact that these developments have not t straints. The subset of available loop-free paths can be computed using a multipath computation algorithm available in literature, for example k-shortestpaths, all k-hop paths, k-disjoint paths (see =-=[5]-=- and references within), DFS with constrained depth ([7] uses a depthconstraint of 1-hop) etc. The only constraint is that the algorithm should also compute the shortest (default) path. These algorith i-PathID). The route computation algorithm (Dijkstra’s algorithm) at upgraded routers must be extended to compute multiple paths (e.g. DFS under partial upgrade constraints (DFS-PU), k-shortest paths =-=[5]-=- etc), and a validation algorithm (Algorithm 1). The upgraded nodes must compute the shortest path as the default path. Incoming packets with erroneous PathIDs are forwarded on the shortest paths and   </text>
<query_num> 6506 </query_num>
<text>   e lack of algorithms and protocols. There have been several proposals for multipath route-computation [5, 6, 7, 8], Internet signaling architectures [9, 10, 11, 12, 13], novel overlay routing methods =-=[14, 15]-=- and transport-level approaches for multihomed hosts [16, 17]. The fact that these developments have not triggered widespread deployment suggests that the core problem is an architectural one 1 . The  e expectation of multiple end-toend paths will trigger application innovation in new areas such as end-to-end bandwidth aggregation [17], end-to-end resilience and video transmission over multi-paths =-=[14, 15, 23]-=- and end-to-end multi-path based security strategies (e.g. protecting data integrity using multipaths). The rest of the paper is organized as follows. Section 2 introduces the abstract framework and c   </text>
<query_num> 6507 </query_num>
<text>   ed in the Linux kernel (MIT’s Click Router platform) and simulated in SSFNet. We present our simulation results in this section on a sizeable topology that corresponds to the old MCI topology of 1995 =-=[38]-=-. In this configuration, only nodes 4, 6, 7, 9, 10 are upgraded. The source node in this simulation is node 6. Observe that node 6 is the only node that computes the kshortest-paths (k = 5) for all de   </text>
<query_num> 6508 </query_num>
<text>   next-hops of equal cost paths. Lorenz et al [39] show that OSPF routing performance could be improved by O(N) if traffic-matrix aware explicit source-based multipath routing is used (e.g. MPLS-based =-=[40, 41]-=-). Protocol extensions to support multipath routing (both in RIP and OSPF) have been studied by Narvaez et al [7], Chen et al [6] and Vutukury et al [8]. In [7], authors propose to find loop-free mult in two broad ways. First, in the context of traffic engineering within a partially upgraded legacy network. An operator may want to emulate signaled capabilities in a connectionless network (e.g. see =-=[41, 39]-=-) or might desire fine-grained traffic management control hard to extract from parameter tweaking (e.g. see [30, 29, 31, 32]). The building blocks may be mixed and matched in a limited number of ways.   </text>
<query_num> 6509 </query_num>
<text>   ti-homed and are increasingly active in managing their inbound and outbound traffic [1, 31]. While BANANAS is not designed to address multitude of configuration, stability and load-balancing problems =-=[32, 29, 33]-=- of BGP, it does provide a set of building blocks to enable fine-grained BGP traffic engineering both within and across domains. In particular, BANANAS introduces two new capabilities: explicit exit f tor may want to emulate signaled capabilities in a connectionless network (e.g. see [41, 39]) or might desire fine-grained traffic management control hard to extract from parameter tweaking (e.g. see =-=[30, 29, 31, 32]-=-). The building blocks may be mixed and matched in a limited number of ways. For example, one could select a MD5+CRC32 encoding for BGP-4 (i.e. ePathIDs) and a index-based encoding for OSPF (i-PathID)   </text>
</top>
<top>
<paper_num> 66 </paper_num>
<paper_title>   RRT-Plan: A Randomized Algorithm for STRIPS Planning.  </paper_title>
<paper_abstract>   We propose a randomized STRIPS planning algorithm called RRT-Plan. This planner is inspired by the idea of Rapidly exploring Random Trees, a concept originally designed for use in continuous path planning problems. Issues that arise in the conversion of RRTs from continuous to discrete spaces are discussed, and several additional mechanisms are proposed to improve performance. We pose several problems that planners based on the relaxed plan length heuristic cannot solve but RRT-Plan can. Our experimental results indicate that RRT-Plan is competitive with the state of the art in STRIPS planning, and that the use of randomization does not significantly worsen plan quality. The success of RRT-Plan indicates that similar randomization techniques could be effective in more advanced planning domains.  </paper_abstract>
<query_num> 6601 </query_num>
<text>   Another standard STRIPS planning problem is Logistics, in which the agent must direct the shipment of packages by trucks and airplanes. STRIPS planning has been studied extensively in the literature (=-=Bylander 1994; Blum &amp; Furst 1995-=-). We now describe two very successful heuristic planners. HSP - Heuristic Search Planner (=-=Bonet &amp; Geffner 1999-=-) first introduced the Heuristic Search Planner (HSP). The algorithm f than without, therefore whenever the length of the optimal relaxed plan can be calculated, the heuristic estimate is admissible. Unfortunately, solving the relaxed planning problem is itself NP-hard (=-=Bylander 1994-=-), therefore in practice HSP uses an approximation of the relaxed costs. This is generally quite effective and HSP can successfully solve a large number of STRIPS planning problems. HSP approximates t   </text>
<query_num> 6602 </query_num>
<text>   al that the goal can be found using a short deterministic search procedure. A good example of the kinds of problems that can be solved easily by RRTs is that of a humanoid robot picking up an object (=-=Kuffner et al. 2003-=-). The robot has many degrees of freedom, and thus the planning problem has high dimensionality. The size of the search space is prohibitive for most traditional deterministic planners, but RRTs are a   </text>
<query_num> 6603 </query_num>
<text>   d STRIPS planning problem is Logistics, in which the agent must direct the shipment of packages by trucks and airplanes. STRIPS planning has been studied extensively in the literature (=-=Bylander 1994;Blum &amp; Furst 1995-=-). We now describe two very successful heuristic planners. HSP - Heuristic Search Planner (=-=Bonet &amp; Geffner 1999-=-) first introduced the Heuristic Search Planner (HSP). The algorithm formulates the plann uces several improvements. First, a more sophisticated technique is used to estimate the length of the relaxed plan. This technique uses the planning graph, from Blum and Furst’s Graphplan procedure (=-=Blum &amp; Furst 1995-=-), to get an improved heuristic estimate. This approximation is more precise because the planning graph is able to take into account positive interactions between goal atoms, which HSP’s technique ign   </text>
<query_num> 6604 </query_num>
<text>   problems are an active area of research (=-=Vidal 2004; Gerevini &amp; Serina 2002; Helmert 2004-=-). Several recent successful algorithms for STRIPS planning are in the class of heuristic planners, e.g. HSP (=-=Bonet &amp; Geffner 1999-=-) and FF (=-=Hoffmann &amp; Nebel 2001-=-), which define a state-specific heuristic evaluation function to guide search through the state space. Despite their success, heurstic planners, due to their determinis d airplanes. STRIPS planning has been studied extensively in the literature (=-=Bylander 1994; Blum &amp; Furst 1995-=-). We now describe two very successful heuristic planners. HSP - Heuristic Search Planner (=-=Bonet &amp; Geffner 1999-=-) first introduced the Heuristic Search Planner (HSP). The algorithm formulates the planning problem as search through a discrete state space, where a heuristic estimate of the distance (# of actions) al cannot be achieved more rapidly by discarding delete effects, the ��� function is admissible. Unfortunately this heuristic is often somewhat uninformative. Indeed, in their original paper on HSP, (=-=Bonet &amp; Geffner 1999-=-) argue that an approximation which is closer to a correct estimate of the true distance to the goal is more helpful than one which provides a strict lower bound. In any event, it often occurs that a  (s) and using the preconditions of those actions as a new set of goals, from which to randomly sample. This could even involve growing another RRT from the goal which expands in the regression space (=-=Bonet &amp; Geffner 1999-=-) of the problem. A large body of other open questions involve how to apply RRT style randomization to the more sophisticated planning problems (such as those which involve temporal constraints, resou   </text>
<query_num> 6605 </query_num>
<text>   re known to be particularly useful in robotics path planning domains where the state space is continuous, highly dimensional, and there is a relatively large goal area. With the notable exception of (=-=Morgan &amp; Branicky 2004-=-), RRTs have not been extended to discrete-space planners. These authors considered only the general case of discrete search, and did not leverage any of the ideas or techniques that have been develop   </text>
<query_num> 6606 </query_num>
<text>   ting state ������������������������¥�� is the same as � , but with the atoms in ��������¥�� added and those in ��������¥�� removed. The canonical example of a STRIPS planning problem is Blocks World (=-=McAllester &amp; Rosenblitt 1991-=-). This classic domain models the arrangement of a collection of building blocks. In each problem there is some number of blocks in an initial configuration, and the goal is to achieve a different spe   </text>
<query_num> 6607 </query_num>
<text>   ve natural goal orderings, meaning that some goal atoms must be achieved before others. The classic example of this is again Blocks World. Several methods exist to discover goal orderings (=-=Korf 1985; Knoblock 1990-=-); we use the heuristic technique of (=-=Koehler &amp; Hoffmann 2000-=-). By utilizing this information, we can improve the selection of random goal subsets by respecting the ordering relationships between goal   </text>
</top>
<top>
<paper_num> 67 </paper_num>
<paper_title>   Clara: Partially Evaluating Runtime Monitors at Compile Time - Tutorial Supplement.  </paper_title>
<paper_abstract>   Clara is a novel static-analysis framework for partially evaluating finite-state runtime monitors at compile time. Clara uses static typestate analyses to automatically convert any AspectJ monitoring aspect into a residual runtime monitor that only monitors events triggered by program locations that the analyses failed to prove safe. If the static analysis succeeds on all locations, this gives strong static guarantees. If not, the efficient residual runtime monitor is guaranteed to capture property violations at runtime. Researchers can use Clara with most runtime-monitoring tools that implement monitors as AspectJ aspects. In this tutorial supplement, we provide references to related reading material that will allow the reader to obtain in-depth knowledge about the context in which Clara can be applied and about the techniques that underlie the Clara framework.  </paper_abstract>
<query_num> 6701 </query_num>
<text>   , Vechev and Yahav present QVM, the “Quality Virtual Machine”, an extension of IBM’s J9 Java Virtual Machine that implements a set of techniques that aim at aiding programmers to debug their programs =-=[3]-=-. QVM comes equipped with support for virtual-machine-level monitoring of single-object typestate properties. Programmers can use a simple syntax to define typestate properties for any given Java clas   </text>
<query_num> 6702 </query_num>
<text>   ch as Extended Regular Expressions or LinearTemporal Logic, e.g. using JavaMOP [18] or tracematches [1]. The engineer then uses some specification compiler such as JavaMOP or the AspectBench Compiler =-=[4]-=- (abc) to automatically translate these finite-state-property definitions into AspectJ monitoring aspects. These aspects may already be annotated with appropriate Dependency State Machines: we extende how Clara relates to existing approaches to runtime monitoring and static typestate analysis. 3 Further reading on Clara and its analyses Clara started out as an extension to the AspectBench Compiler =-=[4]-=- that was specific to one single specification formalism for runtime monitors, called trace-matches [1]. At ECOOP 2007, we presented a set of three static analyses that attempt to statically optimize   </text>
<query_num> 6703 </query_num>
<text>   chitecture, explain its major design decisions and give pointers to further in-depth reading material. 2 Architecture of Clara Clara targets two audiences: researchers in (1) runtime verification and =-=(2)-=- static typestate analysis. Clara defines clear interfaces to allow the two communities to productively interact. Developers of runtime verification tools simply generate AspectJ aspects annotated wit  a subset of the original instrumentation points. It is non-trivial to select subsets of instrumentation points that (1) still have the potential of causing, in combination, a property violation, and =-=(2)-=- will not cause any false warnings at runtime. In our approach, we present an algorithm to select such subsets. We further present an algorithm that enables certain subsets only from time to time. Thi sis approach to be a proper framework, Clara, that can be easily extended by others. For the first time, Clara provides a uniform way to (1) specify runtime monitors as annotated AspectJ aspects, and =-=(2)-=- integrate novel static typestate analyses. During the process, we discovered that the flow-sensitive analysis presented in 2008 [15] was incorrect: in certain cases it could occur that the analysis y scheme that allows the monitor, at any event that binds a variable v to an object o, to quickly look up all state-machine instances for the binding v = o. As Avgustinov et al. show, reclaiming memory =-=(2)-=- and indexing of partial matches (3) are both necessary to achieve a low runtime overhead in the general case. In all the experiments that we conducted with tracematches in our work, these optimizatio   </text>
<query_num> 6704 </query_num>
<text>   d be expressed as a context-free language. However, most interesting program properties are in fact finite-state properties. It is unclear how much runtime overhead tracecuts induce. In previous work =-=[5]-=-, we tried to compare the relative efficiency of J-LO, tracematches, tracecuts and another tool called PQL (see below). As we reported there, there is an implementation of tracecuts, but it is immatur   </text>
<query_num> 6705 </query_num>
<text>   es Clara with enough domain-specific knowledge to analyze the woven program. The accompanying research paper [17] summarizes Clara’s predefined analyses; further details can be found in previous work =-=[10,11,14]-=- and the first author’s dissertation [9]. The result is an optimized instrumented program that updates the runtime monitor at fewer locations. Sometimes, Clara optimizes away all updates, which proves ability of being a “true warning”, otherwise a higher probability. The Clara framework contains these filtering and ranking heuristics as well. In 2009, in joint work with Feng Chen and Grigore Ro¸su =-=[11]-=-, the developers of JavaMOP [18], we generalized the analyses from ECOOP so that they were applicable to AspectJ aspects in general, and to monitors generated by JavaMOP in particular. The analyses pr e ECOOP paper, however include also the following improvements: – The Quick Check in [14] can only detect cases in which a monitor cannot reach a final state as a whole. The improved Quick Check from =-=[11]-=-, on the other hand, considers individual paths to final states. This can yield advantages in case of complicated specifications. – The Orphan Shadows Analysis in [11] is highly optimized. In [14], th ations of the control flow from this point. It was this crucial piece of information that the original analysis was missing. The first author’s dissertation [9] proves this analysis (and the analyses =-=[11]-=- from 2009) sound. 4 Runtime monitoring tools In the following we discuss a number of monitoring tools that influenced the design and implementation of Clara. We also discuss whether programmers could ages. Therefore, the designers of JavaMOP are careful when it comes to making assumptions about the specifications used with their framework. To make JavaMOP compatible with Clara, Feng Chen extended =-=[11]-=- the JavaMOP implementation so that it would perform some limited analysis of the specification, so that JavaMOP could annotate the generated monitors with dependency information that Clara can use to   </text>
<query_num> 6706 </query_num>
<text>   es Clara with enough domain-specific knowledge to analyze the woven program. The accompanying research paper [17] summarizes Clara’s predefined analyses; further details can be found in previous work =-=[10,11,14]-=- and the first author’s dissertation [9]. The result is an optimized instrumented program that updates the runtime monitor at fewer locations. Sometimes, Clara optimizes away all updates, which proves le specification formalism for runtime monitors, called trace-matches [1]. At ECOOP 2007, we presented a set of three static analyses that attempt to statically optimize tracematches at compile time =-=[14]-=-. The three analyses presented there are similar to the three analysis stages that Clara contains today, however they were all bound to tracematches; they did not generalize to any other monitoring to  particular. The analyses presented in this novel work are generalizations ofthe first two analysis stages from the ECOOP paper, however include also the following improvements: – The Quick Check in =-=[14]-=- can only detect cases in which a monitor cannot reach a final state as a whole. The improved Quick Check from [11], on the other hand, considers individual paths to final states. This can yield advan   </text>
<query_num> 6707 </query_num>
<text>   gure) finite-state properties of interest, in some finite-state formalism for runtime monitoring, such as Extended Regular Expressions or LinearTemporal Logic, e.g. using JavaMOP [18] or tracematches =-=[1]-=-. The engineer then uses some specification compiler such as JavaMOP or the AspectBench Compiler [4] (abc) to automatically translate these finite-state-property definitions into AspectJ monitoring as reading on Clara and its analyses Clara started out as an extension to the AspectBench Compiler [4] that was specific to one single specification formalism for runtime monitors, called trace-matches =-=[1]-=-. At ECOOP 2007, we presented a set of three static analyses that attempt to statically optimize tracematches at compile time [14]. The three analyses presented there are similar to the three analysis r programs. Nevertheless, one could annotate the J-LO-generated aspects with dependency information and then use Clara’s static analyses to remove some of this overhead. 4.3 Tracematches Allan et al. =-=[1]-=- are the creators of tracematches. Tracematches share with JLO the idea of generating a low-level AspectJ-based runtime monitor from a high-level specification that uses AspectJ pointcuts to denote ev   </text>
<query_num> 6708 </query_num>
<text>   ines (top right of figure) finite-state properties of interest, in some finite-state formalism for runtime monitoring, such as Extended Regular Expressions or LinearTemporal Logic, e.g. using JavaMOP =-=[18]-=- or tracematches [1]. The engineer then uses some specification compiler such as JavaMOP or the AspectBench Compiler [4] (abc) to automatically translate these finite-state-property definitions into A ”, otherwise a higher probability. The Clara framework contains these filtering and ranking heuristics as well. In 2009, in joint work with Feng Chen and Grigore Ro¸su [11], the developers of JavaMOP =-=[18]-=-, we generalized the analyses from ECOOP so that they were applicable to AspectJ aspects in general, and to monitors generated by JavaMOP in particular. The analyses presented in this novel work are g s to their executables, they did not feel it was appropriate for us to use their prototype for our experiments. 4.5 JavaMOP JavaMOP provides an extensible logic framework for specification formalisms =-=[18]-=-. Via logic plug-ins, one can easily add new logics into JavaMOP and then use these logics within specifications. As we already showed in this thesis, JavaMOP has several specification formalisms buil   </text>
<query_num> 6709 </query_num>
<text>   is now uses may-alias information that is flow-sensitive on the intra-procedural level (opposed to being flowinsensitive everywhere). We use a novel pointer abstraction, called Object Representatives =-=[16]-=-, to transparently combine the different sources of alias information. The original Active-shadows Analysis had no access to such information, it only used flow-insensitive may-alias information. – Wh   </text>
<query_num> 6710 </query_num>
<text>   test; each version only contains partial monitoring code. Collaborative Runtime Verification interacts smoothly with the static analyses. Finally, Clara includes a set of built-in ranking heuristics =-=[15]-=-. These heuristics rank all program points that Clara reports as “potential point of failure” according to a computed confidence value. This confidence value enables Clara to prioritize program points  results showed that this approach scales very well. A journal version of this work appeared in 2008 [13]. Clara contains an option to enable Collaborative Runtime Verification. In 2008, we presented =-=[15]-=- a replacement for the ineffective Active-shadows Analysis. This new analysis improves on the Active-shadows Analysis: – It uses intra-procedural must-alias information to allow for strong updates. In y to (1) specify runtime monitors as annotated AspectJ aspects, and (2) integrate novel static typestate analyses. During the process, we discovered that the flow-sensitive analysis presented in 2008 =-=[15]-=- was incorrect: in certain cases it could occur that the analysis yielded optimized runtime monitors that give false warnings at runtime. (see [10] for an example) Interestingly, in the meantime Naeem   </text>
</top>
<top>
<paper_num> 68 </paper_num>
<paper_title>   Application layer reachability monitoring for IP multicast.  </paper_title>
<paper_abstract>   Monitoring and management have become key requirements for the success of multicast deployment in the Internet. One of the most important monitoring tasks for multicast is to verify the availability of service in the network. This task is usually referred to as reachability monitoring. In this paper, we present an application layer multicast reachability monitoring system called sdr-monitor. Sdrmonitor  has emerged in response to the practical need of verifying service availability and detecting potential problems during the early years of native multicast deployment in the inter-domain. Sdr-monitor leverages an existing application and provides close to real-time reachability monitoring for the multicast infrastructure. Since its initial deployment in 1998, sdr-monitor has been serving the multicast community in detecting and correcting multicast reachability problems in the Internet. In addition, sdr-monitor pioneered a number of additional research projects in multicast monitoring and management. In this paper, we first present the architecture of the sdr-monitor system and its outputs. Then, by using a four-year reachability monitoring data set, we present a long term analysis of the reachability characteristics of the multicast infrastructure. Next, by using additional network layer information, we classify reachability problems. Finally, we evaluate sdr-monitor as a reachability monitoring system and identify a number of ways in which it could be improved.  </paper_abstract>
<query_num> 6801 </query_num>
<text>   e is that if periodic traffic over the course of an hour cannot be received, then criteria for connectivity are not being met. Other research efforts are underway that analyze network layer statistics=-=[30]-=-, [31]. The remaining analysis is based on characterizing a specific type of reachability problem. This analysis was conducted using the data set produced by Phase 2 processing. The specific event we   </text>
<query_num> 6802 </query_num>
<text>   ective, it motivated a number of additional research projects for performing related monitoring and management tasks for multicast including MRM[13], RMPMon[14], HPMM[15], Multicast Beacon[16], Mantra=-=[17]-=-, and MCPM[18]. The remainder of this paper is organized as follows. In the next section, we motivate the importance of multicast monitoring. In Section III, we present the sdr-monitor architecture, i   </text>
<query_num> 6803 </query_num>
<text>   hat if periodic traffic over the course of an hour cannot be received, then criteria for connectivity are not being met. Other research efforts are underway that analyze network layer statistics[30], =-=[31]-=-. The remaining analysis is based on characterizing a specific type of reachability problem. This analysis was conducted using the data set produced by Phase 2 processing. The specific event we are lo   </text>
<query_num> 6804 </query_num>
<text>   olved into a significant portion of Internet traffic[1]. As a result, there is a need to develop better mechanisms to support multimedia data delivery. New network-services, such as multicast delivery=-=[2]-=-, quality-of-service[3], [4], and in-the-network processing[5] have all been proposed as potential solutions. The focus of this paper is multicast. Multicast offers mechanisms to reach tens, thousands   </text>
<query_num> 6805 </query_num>
<text>   result, there is a need to develop better mechanisms to support multimedia data delivery. New network-services, such as multicast delivery[2], quality-of-service[3], [4], and in-the-network processing=-=[5]-=- have all been proposed as potential solutions. The focus of this paper is multicast. Multicast offers mechanisms to reach tens, thousands, even millions of receivers simultaneously in a scalable and   </text>
<query_num> 6806 </query_num>
<text>   set of protocols. First, we use a protocol to construct a multicast forwarding tree connecting sources and receivers in a multicast group. Currently, Protocol Independent MulticastSparse Mode (PIM-SM)=-=[21]-=- is the most widely used protocol for multicast tree construction in the Internet. In addition, in order to provide inter-domain multicast service, we use Multiprotocol Border Gateway Protocol (MBGP)[   </text>
<query_num> 6807 </query_num>
<text>   st has been on developing necessary protocols[7]; deploying them in the Internet[8]; and providing a number of additional services on top of the infrastructure including reliability[9], security[10], =-=[11]-=-, and congestion control[12]. On the other hand, in order to achieve global deployment, we need the ability to monitor and manage multicast infrastructure-wide. One of the most important monitoring ta   </text>
<query_num> 6808 </query_num>
<text>   the work in multicast has been on developing necessary protocols[7]; deploying them in the Internet[8]; and providing a number of additional services on top of the infrastructure including reliability=-=[9]-=-, security[10], [11], and congestion control[12]. On the other hand, in order to achieve global deployment, we need the ability to monitor and manage multicast infrastructure-wide. One of the most imp   </text>
<query_num> 6809 </query_num>
<text>   to reachability problems. A general characteristics of soft-state protocols is that sources periodically transmit refresh messages to one or more number of receivers over lossy communication channels=-=[25]-=-. On the other hand, receivers keep these refresh messages for a finite amount of time. If a receiver does not receive any refresh messages during a timeout period, it removes the state from its cache   </text>
<query_num> 6810 </query_num>
<text>   ulticast has been on developing necessary protocols[7]; deploying them in the Internet[8]; and providing a number of additional services on top of the infrastructure including reliability[9], security=-=[10]-=-, [11], and congestion control[12]. On the other hand, in order to achieve global deployment, we need the ability to monitor and manage multicast infrastructure-wide. One of the most important monitor   </text>
<query_num> 6811 </query_num>
<text>   was relatively straightforward. The MBone network topology was a virtual, flat network. Reachability, in most cases, was all or nothing. Cases of only partial connectivity existed but were not typical=-=[20]-=-. As the MBone has evolved into a native network service, and as the multicast topology has become hierarchical, reachability monitoring has become more complicated. The opportunity for reachability p o block multicast data coming from a certain domain or source, or (3) multicast tree construction and maintenance problems due to buggy implementation or mis-behaving protocol functionality in routers=-=[20]-=- (early dropping of forwarding state in routers, etc.). At this point, we use mtrace to divide reachability problems into these two groups. Our reasoning is that if mtrace returns a valid path between   </text>
</top>
<top>
<paper_num> 69 </paper_num>
<paper_title>   A Theory of Shape by Space Carving.  </paper_title>
<paper_abstract>   In this paper we consider the problem of computing the 3D shape of an unknown, arbitrarily-shaped scene from multiple color photographs taken at known but arbitrarily-distributed viewpoints. By studying the equivalence class of all 3D shapes that reproduce the input photographs, we prove the existence of a special member of this class, the maximal photo-consistent shape, that (1) can be computed from an arbitrary volume that contains the scene, and (2) subsumes all other members of this class. We then give a provably-correct algorithm, called Space Carving, for computing this shape and present experimental results from applying it to the reconstruction of geometrically-complex scenes from several photographs. The approach is specifically designed to (1) build 3D shapes that allow faithful reproduction of all input photographs, (2) resolve the complex interactions between occlusion, parallax, shading, and their effects on arbitrary collections of photographs of a scene, and (3) follow a “least commitment ” approach to 3D shape recovery. 1 1  </paper_abstract>
<query_num> 6901 </query_num>
<text>   ave many attractive properties, existing algorithms [28–30, 33] are not general i.e., they rely on the presence of specific image features such as edges and hence generate only sparse reconstructions =-=[28]-=-, or they place strong constraints on the input viewpoints relative to the scene [29, 30]. Our implementation of the Space Carving Algorithm also uses plane sweeps, but unlike all previous methods the   </text>
<query_num> 6902 </query_num>
<text>   ce Carving Algorithm that iteratively “carves” out the scene from an initial set of voxels. This implementation can be seen as a generalization of silhouette-based techniques like volume intersection =-=[21, 44, 56, 57]-=- to the case of gray-scale and full-color images, and extends voxel coloring [29] and plenoptic decomposition [30] to the case of arbitrary 5sFig. 2: Viewing geometry. camera geometries. 3 Section 5 c s volume must necessarily rely on additional assumptions about the scene’s shape or radiance. While visual hull reconstruction has often been used as a method for recovering 3D shape from photographs =-=[21, 57, 58]-=-, the picture constraints captured by the visual hull only exploit information from the background pixels in these photographs. Unfortunately, these constraints become useless when photographs contain   </text>
<query_num> 6903 </query_num>
<text>   changes in visibility, due to the complete rotation of the object in front of the camera. The low-texture and occlusion properties also cause problems for feature-based structure-from-motion methods =-=[37, 43, 60, 61]-=-, due to the difficulty of locating and tracking a sufficient number of features throughout the sequence. While volume intersection [10, 21, 56] and other contour-based techniques [6, 8, 9, 41, 42, 62   </text>
<query_num> 6904 </query_num>
<text>   e in a 3D scene space and is therefore related to other scene-space stereo algorithms that have been recently proposed [27–34]. Of these, most closely related are recent mesh-based [27] and level-set =-=[35]-=- algorithms, as well as algorithms that sweep a plane or other manifold through a discretized scene space [28–30, 33]. While the algorithms in [27, 35] generate high-quality reconstructions and perfor   </text>
<query_num> 6905 </query_num>
<text>   ependence on viewpoint. Since our notion of photo-consistency implicitly ensures that all these 3D shape cues are taken into account in the recovery process, our approach is related to work on stereo =-=[1, 14, 20]-=-, shape-from-contour [8, 9, 21], as well as shape-from-shading [22–24]. These approaches rely on studying a single 3D shape cue under the assumptions that (1) other sources of variability can be safel   </text>
<query_num> 6906 </query_num>
<text>   g smoothness constraints or other geometric heuristics, there are many cases where it may be advantageous to impose apriori constraints, especially when the scene is known to have a certain structure =-=[53, 54]-=-. Least-commitment reconstruction suggests a new way of incorporating such constraints: rather than imposing them as early as possible in the reconstruction process, we can impose them after first rec   </text>
<query_num> 6907 </query_num>
<text>   ific examples include (1) using a mobile camera mounted with a light source to capture photographs of a scene whose reflectance can be expressed in closed form (e.g., using the Torrance-Sparrow model =-=[17, 47]-=-), and (2) using multiple cameras to capture photographs of an approximately Lambertian scene under arbitrary unknown illumination (Fig. 1). 8s(a) (b) Fig. 3: (a) Illustration of the Visibility Lemma.   </text>
<query_num> 6908 </query_num>
<text>   is reconstructing the shape of a complex 3D scene from multiple photographs. While current techniques work well under controlled conditions (e.g., small stereo baselines [1], active viewpoint control =-=[2]-=-, spatial and temporal smoothness [3–5], or scenes containing curved lines [6], planes [7], or texture-less surfaces [8–12]), very little is known about scene reconstruction under general conditions.   </text>
<query_num> 6909 </query_num>
<text>   nces, known vs. unknown camera motion), make the 1 Examples include the use of the small baseline assumption in stereo to simplify correspondence-finding and maximize joint visibility of scene points =-=[26]-=-, the availability of easily-detectable image contours in shape-from-contour reconstruction [9], and the assumption that all views are taken from the same viewpoint in photometric stereo [24]. 3sgeome   </text>
<query_num> 6910 </query_num>
<text>   o global reconstruction algorithms [12, 37] that recover 3D shape information from all photographs in a single step. This eliminates the need for complex partial reconstruction and merging operations =-=[38, 39]-=- in which partial 3D shape information is extracted from subsets of the photographs [32, 40–42], and where global consistency with the entire set of photographs is not guaranteed for the final shape.   </text>
<query_num> 6911 </query_num>
<text>   ods [37, 43, 60, 61], due to the difficulty of locating and tracking a sufficient number of features throughout the sequence. While volume intersection [10, 21, 56] and other contour-based techniques =-=[6, 8, 9, 41, 42, 62]-=- are often used successfully in similar experiments, they require the detection of silhouettes or occluding contours. For the gargoyle sequence, the background was unknown and heterogeneous, making th   </text>
<query_num> 6912 </query_num>
<text>   om cameras distributed throughout the inside and outside of the house. 4. Because no constraints on the camera viewpoints are imposed, our approach leads naturally to global reconstruction algorithms =-=[12, 37]-=- that recover 3D shape information from all photographs in a single step. This eliminates the need for complex partial reconstruction and merging operations [38, 39] in which partial 3D shape informat   </text>
<query_num> 6913 </query_num>
<text>   om cameras distributed throughout the inside and outside of the house. 4. Because no constraints on the camera viewpoints are imposed, our approach leads naturally to global reconstruction algorithms =-=[12, 37]-=- that recover 3D shape information from all photographs in a single step. This eliminates the need for complex partial reconstruction and merging operations [38, 39] in which partial 3D shape informat  changes in visibility, due to the complete rotation of the object in front of the camera. The low-texture and occlusion properties also cause problems for feature-based structure-from-motion methods =-=[37, 43, 60, 61]-=-, due to the difficulty of locating and tracking a sufficient number of features throughout the sequence. While volume intersection [10, 21, 56] and other contour-based techniques [6, 8, 9, 41, 42, 62   </text>
<query_num> 6914 </query_num>
<text>   owever, does operate in a 3D scene space and is therefore related to other scene-space stereo algorithms that have been recently proposed [27–34]. Of these, most closely related are recent mesh-based =-=[27]-=- and level-set [35] algorithms, as well as algorithms that sweep a plane or other manifold through a discretized scene space [28–30, 33]. While the algorithms in [27, 35] generate high-quality reconst   </text>
<query_num> 6915 </query_num>
<text>   they rely on the presence of specific image features such as edges and hence generate only sparse reconstructions [28], or they place strong constraints on the input viewpoints relative to the scene =-=[29, 30]-=-. Our implementation of the Space Carving Algorithm also uses plane sweeps, but unlike all previous methods the algorithm guarantees complete reconstructions in the general case. Our approach offers s his implementation can be seen as a generalization of silhouette-based techniques like volume intersection [21, 44, 56, 57] to the case of gray-scale and full-color images, and extends voxel coloring =-=[29]-=- and plenoptic decomposition [30] to the case of arbitrary 5sFig. 2: Viewing geometry. camera geometries. 3 Section 5 concludes with experimental results on real and synthetic images. 2 Picture Constr ant improvements in the state of the art. For instance, silhouette-based algorithms require identification of silhouettes, fail at surface concavities, and treat only the case of binary images. While =-=[29, 30]-=- also used a volumetric algorithm, their method worked only when the scene was outside the convex hull of the cameras. This restriction strongly limits the kinds of environments that can be reconstruc g contours. For the gargoyle sequence, the background was unknown and heterogeneous, making the contour detection problem extremely difficult. Note also that Seitz and Dyer’s voxel coloring technique =-=[29]-=- would not work for this sequence because of the camera configuration, i.e., the scene intersects the convex hull of the camera centers. The Space Carving algorithm succeeds for this sequence because  stic changes in visibility between interior and exterior cameras. 10 The voxel space was initialized to a 200 170 200 block, containing roughly 7 million voxels. The 10 For example, the algorithms in =-=[29, 30]-=- fail catastrophically for this scene because the unconstrained distribution of the input views 19sFig. 10: Cameras for the building scene. Cameras were placed in both the interior and exterior of a b   </text>
</top>
<top>
<paper_num> 70 </paper_num>
<paper_title>   Logic-Based Knowledge Representation.  </paper_title>
<paper_abstract>   . After a short analysis of the requirements that a knowledge  representation language must satisfy, we introduce Description Logics,  Modal Logics, and Nonmonotonic Logics as formalisms for representing  terminological knowledge, time-dependent or subjective knowledge, and  incomplete knowledge respectively. At the end of each section, we briefly  comment on the connection to Logic Programming.  1 Introduction  This section is concerned with the question under which conditions one may rightfully claim to have represented knowledge about an application domain, and not just stored data occurring in this domain.  1  In the early days of Artificial Intelligence and Knowledge Representation, there was a heated discussion on whether logic can at all be used as a formalism for Knowledge Representation (see e.g. [135, 91, 92]). One aspect of the requirements on knowledge representation formalisms that can be derived from the considerations in this section is very well satisfied by logical for...  </paper_abstract>
<query_num> 7001 </query_num>
<text>   . By employing appropriate translation techniques and resolution strategies, general purpose resolution provers can, however, be used to obtain decision procedures for subsumption in the language ALC =-=[65, 103, 164]-=-. Decidability of the inference problems for ALC can also be obtained as a consequence of the known decidability result for the two variable fragment of first-order predicate logic. The language L 2 c   </text>
<query_num> 7002 </query_num>
<text>   55, 56] and &amp;quot;Logic Programming and Nonmonotonic Reasoning&amp;quot; [125, 54]. Connection with Description Logics The integration of default rules into Description Logics was investigated in [153, 16, 17]. In =-=[60]-=-, an epistemic operator K is added to the DL ALC. This operator is similar to the modal operators employed in modal nonmonotonic logic, and it can, for example, be used to impose a &amp;quot;local&amp;quot; closed worl   </text>
<query_num> 7003 </query_num>
<text>   ABox individuals) did not focus on decidability issues. The extension of results from propositional modal and dynamic logic to logics allowing for number restrictions and individuals was addressed in =-=[77, 46, 79]-=-. Axiomatizations. If one wants to assign the modal operators with a specific meaning (like &amp;quot;knowledge of an intelligent agent&amp;quot; or &amp;quot;in the future&amp;quot;), then using the basic modal logic Kn is not sufficie   </text>
<query_num> 7004 </query_num>
<text>   ], Sb-one [106]), which have been used in different application domains such as natural language processing [154], configuration of technical systems [178, 42, 158, 133], software information systems =-=[50]-=-, optimizing queries to databases [41, 25, 24], or planning [107]. Syntax and semantics. Figure 2 introduces syntax and semantics of some of the concept constructors employed in systems or investigate   </text>
<query_num> 7005 </query_num>
<text>   and to various description languages extending ALC (see, e.g., [95, 94, 20, 21, 8] for languages with number restrictions; [6] for transitive closure of roles and [159, 99, 101] for transitive roles; =-=[12, 89, 87, 22]-=- for constructs that allow to refer to concrete domains such as numbers; and [10, 40, 8] for the treatment of general axioms of the form C : = D, where C; D may both be complex concept terms). Undecid   </text>
<query_num> 7006 </query_num>
<text>   and to various description languages extending ALC (see, e.g., [95, 94, 20, 21, 8] for languages with number restrictions; [6] for transitive closure of roles and [159, 99, 101] for transitive roles; =-=[12, 89, 87, 22]-=- for constructs that allow to refer to concrete domains such as numbers; and [10, 40, 8] for the treatment of general axioms of the form C : = D, where C; D may both be complex concept terms). Undecid the languages used in early DL systems were too expressive, which led to undecidability of the subsumption problem [165, 147]. More recent undecidability results for extensions of ALC can be found in =-=[13, 89, 20, 21, 87, 22]-=-. The first complexity results [115, 139] showed that, even for very small languages, there cannot exist subsumption algorithms that are both complete and polynomial. In the meantime, the worst-case c   </text>
<query_num> 7007 </query_num>
<text>   available (see [88] for detailed proofs of these and other complexity results). The properties of modal operators that model time-dependent knowledge have also been investigated in detail (see, e.g., =-=[63, 171]-=- for overview articles on this topic). Integration with Description Logics. In order to represent time-dependent and subjective knowledge in Description Logics, one can integrate modal operators into   </text>
<query_num> 7008 </query_num>
<text>   axioms (in ALC) even causes the subsumption problem to become ExpTime-complete [163]. It has also been shown that for certain languages the instance problem can be harder than the subsumption problem =-=[160, 61]-=-. Optimizations. Considering these complexity results, one may ask whether incomplete, but polynomial algorithms should be preferred over the complete ones, which are necessarily of high worst-case co   </text>
<query_num> 7009 </query_num>
<text>   both complete and polynomial. In the meantime, the worst-case complexity of the subsumption problem in a large class of DL languages, the so-called AL-family, has (almost completely) been determined =-=[59, 58, 57]-=-. With the exception of a few polynomially decidable languages, the complexity results range between NP or coNP and PSPACE. Whereas these results are given with respect to an empty TBox (i.e., they co   </text>
<query_num> 7010 </query_num>
<text>   constructors that cannot be expressed within L 2 . Number restrictions can, however, be expressed in C 2 , the extension of L 2 by counting quantifiers, which has recently been shown to be decidable =-=[86, 144]-=-. Another distinguishing features of the formulae obtained by translating ALCconcept terms into first-order predicate logic is that quantifiers are used only in a very restricted way: the quantified v   </text>
<query_num> 7011 </query_num>
<text>   ded to the instance problem [93, 15] and to various description languages extending ALC (see, e.g., [95, 94, 20, 21, 8] for languages with number restrictions; [6] for transitive closure of roles and =-=[159, 99, 101]-=- for transitive roles; [12, 89, 87, 22] for constructs that allow to refer to concrete domains such as numbers; and [10, 40, 8] for the treatment of general axioms of the form C : = D, where C; D may   </text>
<query_num> 7012 </query_num>
<text>   domains such as natural language processing [154], configuration of technical systems [178, 42, 158, 133], software information systems [50], optimizing queries to databases [41, 25, 24], or planning =-=[107]-=-. Syntax and semantics. Figure 2 introduces syntax and semantics of some of the concept constructors employed in systems or investigated in the literature. Most of the systems do not provide for all o   </text>
<query_num> 7013 </query_num>
<text>   e constructors of Figure 2. Since then, this approach for designing subsumption algorithms was extended to the instance problem [93, 15] and to various description languages extending ALC (see, e.g., =-=[95, 94, 20, 21, 8]-=- for languages with number restrictions; [6] for transitive closure of roles and [159, 99, 101] for transitive roles; [12, 89, 87, 22] for constructs that allow to refer to concrete domains such as nu   </text>
<query_num> 7014 </query_num>
<text>   e constructors of Figure 2. Since then, this approach for designing subsumption algorithms was extended to the instance problem [93, 15] and to various description languages extending ALC (see, e.g., =-=[95, 94, 20, 21, 8]-=- for languages with number restrictions; [6] for transitive closure of roles and [159, 99, 101] for transitive roles; [12, 89, 87, 22] for constructs that allow to refer to concrete domains such as nu the languages used in early DL systems were too expressive, which led to undecidability of the subsumption problem [165, 147]. More recent undecidability results for extensions of ALC can be found in =-=[13, 89, 20, 21, 87, 22]-=-. The first complexity results [115, 139] showed that, even for very small languages, there cannot exist subsumption algorithms that are both complete and polynomial. In the meantime, the worst-case c   </text>
<query_num> 7015 </query_num>
<text>   een DL and modal logics, such an integrations appeared to be rather simple. It has turned out, however, that it is more complex than expected, both from the semantic and the algorithmic point of view =-=[113, 19, 18, 162, 175]-=-. It should be noted that some of these combined languages [18, 175] are first-order modal logics rather than propositional modal logics. Connection with Logic Programming. We close this section by me   </text>
<query_num> 7016 </query_num>
<text>   ges with number restrictions; [6] for transitive closure of roles and [159, 99, 101] for transitive roles; [12, 89, 87, 22] for constructs that allow to refer to concrete domains such as numbers; and =-=[10, 40, 8]-=- for the treatment of general axioms of the form C : = D, where C; D may both be complex concept terms). Undecidability and complexity results. Other important research contributions for DL are concer   </text>
<query_num> 7017 </query_num>
<text>   he finite model property. For these languages, reasoning with respect to finite models (which is, for example, of interest for database applications) differs from reasoning with respect to all models =-=[43]-=-. Given the translation of DL into first-order predicate logic mentioned above, one might ask whether general first-order theorem provers can be employed for reasoning in DL. In general, this approach   </text>
<query_num> 7018 </query_num>
<text>   her hand, it could mean that any frog has at least the colour green, but may have other colours too (it might be 4 Although there is no overview article that covers all aspects of this research area, =-=[177, 15, 62]-=- can serve as a starting point. - colour 6 6 6 Z Z Z Z - Animal Frog Kermit IS-A IS-A IS-A IS-A colour Brown Green Treefrog Grassfrog Fig. 1. A semantic network. green with red stripes). A partial rec   </text>
<query_num> 7019 </query_num>
<text>   her hand, it could mean that any frog has at least the colour green, but may have other colours too (it might be 4 Although there is no overview article that covers all aspects of this research area, =-=[177, 15, 62]-=- can serve as a starting point. - colour 6 6 6 Z Z Z Z - Animal Frog Kermit IS-A IS-A IS-A IS-A colour Brown Green Treefrog Grassfrog Fig. 1. A semantic network. green with red stripes). A partial rec nd the restricted expressiveness of these logics allows to design terminating procedures, i.e., for many description languages one obtains decision procedures for the relevant inference problems (see =-=[15]-=- for an introductory exposition of tableau-based inference methods in DL). The first tableau-based subsumption algorithm was developed in [166] for the language ALC, which allows for the first five co   </text>
<query_num> 7020 </query_num>
<text>   in different application domains such as natural language processing [154], configuration of technical systems [178, 42, 158, 133], software information systems [50], optimizing queries to databases =-=[41, 25, 24]-=-, or planning [107]. Syntax and semantics. Figure 2 introduces syntax and semantics of some of the concept constructors employed in systems or investigated in the literature. Most of the systems do no   </text>
<query_num> 7021 </query_num>
<text>   ke KR languages.&amp;quot; use a Logic Programming approach for KR. Overviews on the topic of extended Logic Programs with declarative semantics and their application to representing knowledge can be found in =-=[53, 134, 23, 4]-=-. 2 Description Logics The attempt to provide for a structured representation of information was one of the main motivations for introducing early KR formalisms such as Semantic Networks and Frames. D   </text>
<query_num> 7022 </query_num>
<text>   loped both from the theoretical and the practical point of view. In particular, there is a great variety of successor systems (e.g., Back [143, 149, 96], Classic [29, 33], Crack [36], DLP [148], FaCT =-=[98]-=-, Flex [152], K-Rep [127, 128], Kris [14], Loom [122, 121], Sb-one [106]), which have been used in different application domains such as natural language processing [154], configuration of technical s  these complexity results, one may ask whether incomplete, but polynomial algorithms should be preferred over the complete ones, which are necessarily of high worst-case complexity. First experiences =-=[11, 36, 98]-=- with implemented systems using complete algorithms show, however, that on realistic KBs the run time is comparable to that of Classic and Loom (i.e., mature systems using incomplete algorithms). Thes s. Whereas [11] concentrated mostly on reducing the number of subsumption tests during classification, more recent work in this direction is concerned with optimizing the subsumption algorithm itself =-=[83, 97, 98, 82, 100, 104]-=-. Connections with other logical formalisms. Before we turn to the connection between DL and Logic Programming, we should like to mention several interesting connections between DL and more traditiona   </text>
<query_num> 7023 </query_num>
<text>   n into account. The integration of Description Logics with a rule calculus that is able to express Horn rules has been investigated in [90]. Other work in this direction can, for example, be found in =-=[117, 116]-=-. 3 Modal Logics This is an area of logics that has been investigated for quite a while, and for which a great variety of methods and results are available. In the following, we give a very short intr   </text>
<query_num> 7024 </query_num>
<text>   nevertheless has the finite model property, which implies that satisfiability of formulae in GF is decidable. More precisely, the satisfiability problem for GF is complete for double exponential time =-=[84]-=-. Decidability of GF can also be shown with the help of resolution methods [48]. Connection with Logic Programming Since Logic Programming languages are computationally complete and DL languages are u   </text>
<query_num> 7025 </query_num>
<text>   onal Logic Programming languages. To overcome this deficit, extensions of Logic Programming languages by disjunction and classical negation (in contrast to &amp;quot;negation as failure&amp;quot;) have been introduced =-=[76, 151, 119, 23, 35]-=-. However, these extensions treat only some aspects of these constructors: for example, the &amp;quot;classical negation&amp;quot; in these approaches only represents the aspect that a set and its complement are disjoi   </text>
<query_num> 7026 </query_num>
<text>   pretation) or by employing tree automata (see, e.g, [174]). It should be noted, however, that some of the very expressive DL languages considered in this context (e.g., the language CIQ introduced in =-=[79]-=-) no longer satisfy the finite model property. For these languages, reasoning with respect to finite models (which is, for example, of interest for database applications) differs from reasoning with r ABox individuals) did not focus on decidability issues. The extension of results from propositional modal and dynamic logic to logics allowing for number restrictions and individuals was addressed in =-=[77, 46, 79]-=-. Axiomatizations. If one wants to assign the modal operators with a specific meaning (like &amp;quot;knowledge of an intelligent agent&amp;quot; or &amp;quot;in the future&amp;quot;), then using the basic modal logic Kn is not sufficie   </text>
<query_num> 7027 </query_num>
<text>   propositional modal logics. Connection with Logic Programming. We close this section by mentioning some work that is concerned with the connection between modal logics and logic programming. Gelfond =-=[73]-=- extends Disjunctive Logic Programs to Epistemic Logic Programs by introducing modal operators K and M: for a literal L, the expression KL should be read as &amp;quot;L is known&amp;quot; and ML should be read as &amp;quot;L ma   </text>
<query_num> 7028 </query_num>
<text>   roles. This idea has been further developed both from the theoretical and the practical point of view. In particular, there is a great variety of successor systems (e.g., Back [143, 149, 96], Classic =-=[29, 33]-=-, Crack [36], DLP [148], FaCT [98], Flex [152], K-Rep [127, 128], Kris [14], Loom [122, 121], Sb-one [106]), which have been used in different application domains such as natural language processing [ n approach, the use of non-standard semantics has been proposed. In [146], a four-valued semantics characterizing the behaviour of a structural subsumption algorithm is introduced. The system Classic =-=[29, 33]-=- employs an &amp;quot;almost&amp;quot; complete structural subsumption algorithm [30]. Its only incompleteness stems from the treatment of individuals inside concept terms, which can, however, again be characterized wi   </text>
<query_num> 7029 </query_num>
<text>   s. Whereas [11] concentrated mostly on reducing the number of subsumption tests during classification, more recent work in this direction is concerned with optimizing the subsumption algorithm itself =-=[83, 97, 98, 82, 100, 104]-=-. Connections with other logical formalisms. Before we turn to the connection between DL and Logic Programming, we should like to mention several interesting connections between DL and more traditiona   </text>
<query_num> 7030 </query_num>
<text>   signing subsumption algorithms was extended to the instance problem [93, 15] and to various description languages extending ALC (see, e.g., [95, 94, 20, 21, 8] for languages with number restrictions; =-=[6]-=- for transitive closure of roles and [159, 99, 101] for transitive roles; [12, 89, 87, 22] for constructs that allow to refer to concrete domains such as numbers; and [10, 40, 8] for the treatment of   [161] was the first to observe that the language ALC is a syntactic variant of the propositional multi-modal logic Kn (see next section), and that the extension of ALC by transitive closure of roles =-=[6]-=- corresponds to propositional dynamic logic (PDL) [67, 145]. In particular, the algorithms used in modal logics for deciding satisfiability are very similar to the tableau-based algorithms newly devel   </text>
<query_num> 7031 </query_num>
<text>   stant symbols (but without function symbols) using only the variables x; y. Decidability of L 2 has been shown in [137]. More precisely, satisfiability of L 2 -formulae is a NEXPTIME-complete problem =-=[85]-=-. It is easy to see that, by appropriately re-using variable names, any concept term of the language ALC can be translated into an L 2 -formula with one free variable. A direct translation of the conc   </text>
<query_num> 7032 </query_num>
<text>   stead of polynomial) for certain languages. In the presence of cyclic TBox definitions, so-called terminological cycles, the subsumption problem becomes PSPACE-complete even for these small languages =-=[140, 5, 7, 112]-=-. The use of general inclusion axioms (in ALC) even causes the subsumption problem to become ExpTime-complete [163]. It has also been shown that for certain languages the instance problem can be harde   </text>
<query_num> 7033 </query_num>
<text>   stead of using tableau-based algorithms, decidability of certain propositional modal logics (and thus of the corresponding DL), can also be shown by establishing the finite model property (see, e.g., =-=[68]-=-, Section 1.14) of the logic (i.e., showing that a formula/concept is satisfiable iff it is satisfiable in a finite interpretation) or by employing tree automata (see, e.g, [174]). It should be noted, owing, we give a very short introduction, which emphasizes the connection between Description Logics and Modal Logics. For more detailed introductions and overviews of the area we refer the reader to =-=[44, 102, 88, 68]-=-. The propositional multi-modal logic Kn extends propositional logic by n pairs of unary operators, which are called box and diamond operators. The K stands for the basic modal logic on which most mod   </text>
<query_num> 7034 </query_num>
<text>   t provide for all of these constructors, and vice versa, they may use additional constructors not introduced here. An extensive list of (most of) the constructors considered until now can be found in =-=[9]-=-. The first column of the figure shows the (Lisp-like) concrete syntax that is used in most (and C1 : : : Cn ) C1 u : : : u Cn C I 1 &amp;quot; : : : &amp;quot; C I n (or C1 : : : Cn) C1 t : : : t Cn C I 1 [ : : : [ C   </text>
<query_num> 7035 </query_num>
<text>   te that these deficiencies can be overcome with the help of other logic-based formalisms, namely Description Logics, Modal Logics, and Nonmonotonic Logics. More recently, it has been argued (see e.g. =-=[110, 23]-=-) that Logic Programming can serve as a convenient and universal formalism for Knowledge Representation. However, as indicated by their name, Logic Programming languages are programming languages, and ke KR languages.&amp;quot; use a Logic Programming approach for KR. Overviews on the topic of extended Logic Programs with declarative semantics and their application to representing knowledge can be found in =-=[53, 134, 23, 4]-=-. 2 Description Logics The attempt to provide for a structured representation of information was one of the main motivations for introducing early KR formalisms such as Semantic Networks and Frames. D onal Logic Programming languages. To overcome this deficit, extensions of Logic Programming languages by disjunction and classical negation (in contrast to &amp;quot;negation as failure&amp;quot;) have been introduced =-=[76, 151, 119, 23, 35]-=-. However, these extensions treat only some aspects of these constructors: for example, the &amp;quot;classical negation&amp;quot; in these approaches only represents the aspect that a set and its complement are disjoi   </text>
<query_num> 7036 </query_num>
<text>   ther option) explicit (see below). Systems and applications. Description Logics are descended from so-called &amp;quot;structured inheritance networks&amp;quot; [31, 32], which were first realized in the system kl-one =-=[34]-=-. Their main idea is to start with atomic concepts (unary predicates) and roles (binary predicates), and use a (rather small) set of epistemologically adequate constructors to build complex concepts a   </text>
<query_num> 7037 </query_num>
<text>   these complexity results, one may ask whether incomplete, but polynomial algorithms should be preferred over the complete ones, which are necessarily of high worst-case complexity. First experiences =-=[11, 36, 98]-=- with implemented systems using complete algorithms show, however, that on realistic KBs the run time is comparable to that of Classic and Loom (i.e., mature systems using incomplete algorithms). Thes   </text>
<query_num> 7038 </query_num>
<text>   to the existing approaches and the problems treated by these approaches. Overviews of this research area can, for example, be found in [71, 38]. In addition, there are several monographs on the topic =-=[37, 120, 126, 3]-=-. An annotated collection of influential papers in the area can be found in [80]. Motivation. Knowledge representation languages based on classical logics (e.g., first-order predicate logic) are monot   </text>
<query_num> 7039 </query_num>
<text>   which none seems to be &amp;quot;the best&amp;quot; approach. A very positive development is the fact that recently several results clarifying the connection between different approaches have been obtained (see, e.g., =-=[105, 108, 109, 118, 111]-=-). In the following, we briefly introduce the four most important types of approaches. Consistency-based approaches , of which Reiter&amp;apos;s Default Logic [157] is a typical example, consider nonmonotonic   </text>
<query_num> 7040 </query_num>
<text>   which none seems to be &amp;quot;the best&amp;quot; approach. A very positive development is the fact that recently several results clarifying the connection between different approaches have been obtained (see, e.g., =-=[105, 108, 109, 118, 111]-=-). In the following, we briefly introduce the four most important types of approaches. Consistency-based approaches , of which Reiter&amp;apos;s Default Logic [157] is a typical example, consider nonmonotonic  e 4 gives several examples of such reasonable properties. It has also turned out that there is a close connection between preferential semantics and inferences relations satisfying certain properties =-=[123, 111]-=-. OE j OE Reflexivity If OE j / and OE equivalent to OE 0 then OE 0 j / Left equivalence If OE j / and / implies / 0 then OE j / 0 Right weakening If OEsOE 0 j / and OE j OE 0 then OE j / Cut If OE j   </text>
<query_num> 7041 </query_num>
<text>   y; z)sA(z)) does not contain x, this variable can be re-used: renaming the bound variable z into x yields the equivalent formula 8y:(R(x; y) ! (9x:(R(y; x) A(x)))), which uses only two variables (see =-=[28]-=- for details). This connection between ALC and L 2 shows that any extension of ALC by constructors that can be expressed with the help of only two variables yields a decidable DL. Number restrictions   </text>
</top>
<top>
<paper_num> 71 </paper_num>
<paper_title>   A Performance Optimization Framework for Compilation of Tensor Contraction Expressions into Parallel Programs.  </paper_title>
<paper_abstract>   This paper discusses a program synthesis system to facilitate the generation of high-performance parallel programs for a class of computations encountered in quantum chemistry and physics. These computations are expressible as a set of tensor contractions and arise in electronic structure modeling. An overview is provided of the synthesis system under development, that will take as input a high-level specification of the computation and generate high-performance parallel code for a number of target architectures. Several components of the synthesis system are described, focusing on compile-time optimization issues that they address.  </paper_abstract>
<query_num> 7101 </query_num>
<text>   a systematic search through the space of possible implementations. Other efforts in automatically generating efficient implementations of programs include FFTW [6], the telescoping languages project =-=[9]-=-, ATLAS [22] for deriving efficient implementation of BLAS routines, the PHIPAC [1] project, and the TUNE project [21]. In addition, motivated by the difficulty of detecting and optimizing matrix oper   </text>
<query_num> 7102 </query_num>
<text>   and optimizing matrix operations hidden in array subscript expressions within loop nests, several projects have worked on efficient code generation from high-level languages such as MATLAB and Maple =-=[5, 17, 18, 19]-=-. While our effort shares some common goals with several of the projects mentioned above, there are also significant differences. Some of the optimizations we consider, such as the algebraic optimizat   </text>
<query_num> 7103 </query_num>
<text>   he lowest recomputation cost. 6. Data locality optimization Once a solution is found that fits onto disk, we optimize the data locality to reduce memory and disk access times. We developed algorithms =-=[2, 3]-=- that, given a memory-reduced (fused) version of the code, find the appropriate blocking of the loops in order to maximize data reuse. These algorithms can be applied at different levels of the memory   </text>
<query_num> 7104 </query_num>
<text>   he lowest recomputation cost. 6. Data locality optimization Once a solution is found that fits onto disk, we optimize the data locality to reduce memory and disk access times. We developed algorithms =-=[2, 3]-=- that, given a memory-reduced (fused) version of the code, find the appropriate blocking of the loops in order to maximize data reuse. These algorithms can be applied at different levels of the memory memory and disk (disk access minimization) , or to minimize data transfer between main memory and the cache (cache optimization). In this section, we briefly describe the main points of our algorithm =-=[3]-=-, focusing mostly on the cache management problem. For the disk access minimization problem, the same approach is used, replacing the cache size by the physical memory size. We introduce a memory acce   </text>
<query_num> 7105 </query_num>
<text>   ic search through the space of possible implementations. Other efforts in automatically generating efficient implementations of programs include FFTW [6], the telescoping languages project [9], ATLAS =-=[22]-=- for deriving efficient implementation of BLAS routines, the PHIPAC [1] project, and the TUNE project [21]. In addition, motivated by the difficulty of detecting and optimizing matrix operations hidde   </text>
<query_num> 7106 </query_num>
<text>   ire different loops to be made the outermost. In prior work, we addressed the problem of finding the choice of fusions for a given operator tree that minimized the total space required for all arrays =-=[10, 11, 12]-=-. 3. An example One of the most computationally intensive components of many quantum chemistry packages is the CCSD(T) scheme. It is a coupled cluster approximation that includes single and double exc   </text>
<query_num> 7107 </query_num>
<text>   ject bears similarities to some projects in other domains, such as the SPIRAL project which is aimed at the design of a system to generate efficient libraries for digital signal processing algorithms =-=[20, 8, 23]-=-. SPIRAL generates efficient implementations of algorithms expressed in a domain-specific language called SPL by a systematic search through the space of possible implementations. Other efforts in aut   </text>
<query_num> 7108 </query_num>
<text>   nt implementations of programs include FFTW [6], the telescoping languages project [9], ATLAS [22] for deriving efficient implementation of BLAS routines, the PHIPAC [1] project, and the TUNE project =-=[21]-=-. In addition, motivated by the difficulty of detecting and optimizing matrix operations hidden in array subscript expressions within loop nests, several projects have worked on efficient code generat   </text>
<query_num> 7109 </query_num>
<text>   shown for the operation-reduced form above. We have shown that the problem of determining the operator tree with minimal operation count is NP-complete, and have developed a pruning search procedure =-=[13, 14]-=- that is very efficient in practice. For the above example, although the latter form is far more economical in terms of the number of operations, its implementation will require the use of temporary i mization and memory considerations. In the next sections, we provide some details about the optimizations implemented in some of these modules. For details of the operation minimization algorithm see =-=[13, 14]-=-. 5. Memory minimization and space-time trade-offs As discussed in Section 2, the operation minimization procedure results in the creation of intermediate temporary arrays. For typical computations in   </text>
<query_num> 7110 </query_num>
<text>   uirements still exceed the disk capacity, we need to recompute some (parts of) temporary arrays in order to further reduce the space requirements. The space-time trade-off algorithm we have developed =-=[4]-=- employs a combination of fusion and tiling to achieve a good balance between recomputation and memory usage. The first step of the space-time trade-off algorithm uses a dynamic programming approach s   </text>
</top>
<top>
<paper_num> 72 </paper_num>
<paper_title>   Characterization of the E-commerce Storage Subsystem Workload.  </paper_title>
<paper_abstract>   This paper characterizes the workload seen at the storage subsystem of an e-commerce system. Measurements are conducted on multi-tiered systems running three different benchmarks, i.e., TPC-W, TPC-C, and RUBiS. In this environment, TPC-W and RUBiS are used to represent webbased e-commerce applications (i.e., on-line shopping and auctioning). They generate mostly READ-dominated workloads. The TPC-C benchmark, although not directly an e-commerce benchmark, is used to represent an e-commerce system under heavy on-line transactions processing activity. Different from the TPC-W and RUBiS benchmarks, TPC-C generates WRITE-dominated workloads. For all three benchmarks, in addition to the system load, the workload mix causes the system resources such as memory to saturate, IO traffic to increase, and, consequently, overall system throughput to reduce. Generally, when a workload shifts from web-site browsing (i.e, reading) to transaction processing (i.e writing) the IO load reduces but the footprint of the IO working set increases, which slows down the IO subsystem. File system and device driver scheduling represent elements in the IO path that for a given set of system resources further improve user-level throughput. Their impact is visible for medium to high utilization and diminishes for light load or overload. 1  </paper_abstract>
<query_num> 7201 </query_num>
<text>   109/QEST.2008.44 297 Authorized licensed use limited to: Seagate Technology. Downloaded on November 24, 2008 at 11:32 from IEEE Xplore. Restrictions apply.either in the IO path or somewhere above it =-=[3, 6]-=-. Our measurements indicate that the memory size of the database server (relative to the database size) largely determines the load seen by the IO subsystems under ecommerce workloads. Nevertheless, t arks are proposed to model on-line e-commerce services such as RUBiS [1] that models an on-line auction server. Various workload characterization studies are based on measurements on these benchmarks =-=[5, 20, 8, 7, 3, 6]-=- with some further generalizing the behavior via analytic models [18]. Our work differs from previous ones, because our focus is on the detailed characterization of the block-level workload in the und chmark to capture occasional heavy load situations on e-commerce sites. Consequently, our work differs from previous work that focuses solely on the behavior of the TPC-C benchmark and OLTP workloads =-=[8, 7, 3, 6]-=-. 3 Measurement System 3.1 Benchmarks Evaluated In this paper, three benchmarks are evaluated; TPC-W and RUBiS capturing the average behavior and TPC-C capturing the occasional heavy-transaction behav   </text>
<query_num> 7202 </query_num>
<text>   2, 21]. Better understanding of the behavior and workload in e-commerce systems has resulted in new and more effective resource management policies that are tailored for such applications and systems =-=[4, 11, 12, 14]-=-. Because it is difficult to obtain data from real ecommerce sites, one can only resort to synthetic workload generators to study such systems, with the most prominent ones being the benchmarks of the   </text>
<query_num> 7203 </query_num>
<text>   an on-line auction server. Various workload characterization studies are based on measurements on these benchmarks [5, 20, 8, 7, 3, 6] with some further generalizing the behavior via analytic models =-=[18]-=-. Our work differs from previous ones, because our focus is on the detailed characterization of the block-level workload in the underlying system that supports e-commerce applications rather than gene   </text>
<query_num> 7204 </query_num>
<text>   arks are proposed to model on-line e-commerce services such as RUBiS [1] that models an on-line auction server. Various workload characterization studies are based on measurements on these benchmarks =-=[5, 20, 8, 7, 3, 6]-=- with some further generalizing the behavior via analytic models [18]. Our work differs from previous ones, because our focus is on the detailed characterization of the block-level workload in the und   </text>
<query_num> 7205 </query_num>
<text>   ase server appears as a single process thread. 4 Overall system Performance For our measurements the IO subsystem is configured with its default settings, i.e., the Ext3 file system, the Anticipatory =-=[9]-=- IO scheduler, a maximum queue length of 4 at the disk, and enabled WRITE-caching at the disk. Figure 1 plots the user-level throughput, measured in transactions per minute, as a function of the numbe   </text>
<query_num> 7206 </query_num>
<text>   support daily commercial and personal activities, the complexity of the underlying systems has increased and various studies have been conducted to understand the main characteristics of such systems =-=[2, 13, 22, 21]-=-. Better understanding of the behavior and workload in e-commerce systems has resulted in new and more effective resource management policies that are tailored for such applications and systems [4, 11   </text>
<query_num> 7207 </query_num>
<text>   support daily commercial and personal activities, the complexity of the underlying systems has increased and various studies have been conducted to understand the main characteristics of such systems =-=[2, 13, 22, 21]-=-. Better understanding of the behavior and workload in e-commerce systems has resulted in new and more effective resource management policies that are tailored for such applications and systems [4, 11  focus is on the detailed characterization of the block-level workload in the underlying system that supports e-commerce applications rather than general understanding of such systems workloads as in =-=[1, 5, 21]-=-. We stress that the databases supporting e-commerce systems differ from the productionlevel ones on the way that they are accessed from the users. As a result, the IO e-commerce workloads are READ do   </text>
<query_num> 7208 </query_num>
<text>   the normal operation of an e-commerce site with the TPC-W and RUBiS benchmarks and the occasional high activity periods with the TPC-C benchmark. As a result, our work is different from previous ones =-=[8, 7]-=- that evaluate TPC-C with the goal of understanding the behavior of production databases where the normal operation is characterized by heavy online-transaction processing. Although we conduct measure arks are proposed to model on-line e-commerce services such as RUBiS [1] that models an on-line auction server. Various workload characterization studies are based on measurements on these benchmarks =-=[5, 20, 8, 7, 3, 6]-=- with some further generalizing the behavior via analytic models [18]. Our work differs from previous ones, because our focus is on the detailed characterization of the block-level workload in the und chmark to capture occasional heavy load situations on e-commerce sites. Consequently, our work differs from previous work that focuses solely on the behavior of the TPC-C benchmark and OLTP workloads =-=[8, 7, 3, 6]-=-. 3 Measurement System 3.1 Benchmarks Evaluated In this paper, three benchmarks are evaluated; TPC-W and RUBiS capturing the average behavior and TPC-C capturing the occasional heavy-transaction behav   </text>
</top>
<top>
<paper_num> 73 </paper_num>
<paper_title>   Towards IP geolocation using delay and topology measurements.  </paper_title>
<paper_abstract>   We present Topology-based Geolocation (TBG), a novel approach to estimating the geographic location of arbitrary Internet hosts. We motivate our work by showing that 1) existing approaches, based on end-to-end delay measurements from a set of landmarks, fail to outperform much simpler techniques, and 2) the error of these approaches is strongly determined by the distance to the nearest landmark, even when triangulation is used to combine estimates from different landmarks. Our approach improves on these earlier techniques by leveraging network topology, along with measurements of network delay, to constrain host position. We convert topology and delay data into a set of constraints, then solve for router and host locations simultaneously. This approach improves the consistency of location estimates, reducing the error substantially for structured networks in our experiments on Abilene and Sprint. For networks with insufficient structural constraints, our techniques integrate external hints that are validated using measurements before being trusted. Together, these techniques lower the median estimation error for our university-based dataset to 67 km vs. 228 km for the best previous approach.  </paper_abstract>
<query_num> 7301 </query_num>
<text>   ct their actual location. Hostnames, especially for routers, may follow naming conventions that include geographic hints in the form of cities or airport codes [23], but many do not or are misleading =-=[29]-=-. 2.4 Sensor Network Localization Our topology-based techniques are inspired by similar techniques from the sensor network community, originally meant for positioning sensor nodes using observed radio The issue with using router DNS name information is that they could be misnamed, with the naming errors introduced by router reconfiguration, router repairs/upgrades, and reassignment of IP addresses =-=[29]-=-. Some names may be correct but misleading or ambiguous: does chrl indicate Charlotte, NC, or Charleston, SC? Any geolocation technique that uses location hints must validate the location derived from   </text>
<query_num> 7302 </query_num>
<text>   e are not the first to do so, we began our research by studying techniques proposed by others. These techniques are based on end-to-end delay measurements from a set of landmarks with known locations =-=[18, 10]-=-. As we experimented with variations and evaluated them using a dataset gathered on PlanetLab, we were surprised to discover that much simpler delay-based algorithms were able to deliver performance t  any measurements that originate at targets. 2.2 Internet Measurement Techniques Two published geolocation techniques fit our problem and approach, GeoPing [18] and Constraint-Based Geolocation (CBG) =-=[10]-=-. Both use delay measurements from landmarks to estimate the location of Internet hosts. We present them in some detail below, because they show how delays may relate to location in non-trivial ways,  region is bounded by a bold perimeter. Experiments have shown CBG to provide better geolocation estimates than GeoPing and DNS-based approaches (e.g. [28]) on both United States and European datasets =-=[10]-=-. 2.3 Other IP Geolocation Techniques There are several geolocation techniques that can be used with Internet hosts but which we do not consider viable solutions to our problem. The Global Positioning ve been previously evaluated using measurements with a heavy academic bias. GeoPing was evaluated with university-based landmarks and targets [18], and CBG was evaluated with NLANR and PlanetLab data =-=[10]-=-. We expect these algorithms to be equally applicable to our dataset, and indeed judge them to perform on it comparatively as well as in their published evaluations. To evaluate each technique, we com   </text>
<query_num> 7303 </query_num>
<text>   enge, GPS does not function well indoors or in urban canyons, limiting its ubiquity, and it would preclude many applications because it is client-driven. Other systems such as Place Lab [13], Cricket =-=[20]-=- and RADAR [1] locate mobile hosts using 802.11 and GSM beacons with known locations. These systems have the potential for accurate estimates but their coverage is limited by the propagation range of   </text>
<query_num> 7304 </query_num>
<text>   ere needed. To this end, we investigate algorithms that combine delay with topology to factor out the circuitousness of Internet paths. Inspired by algorithms used for localization in sensor networks =-=[7, 4]-=-, we convert Internet route measurements into a set of constraints on the unknown locations of targets and intermediate routers en route to it, and then simultaneously geolocate the target and all of  etwork Localization Our topology-based techniques are inspired by similar techniques from the sensor network community, originally meant for positioning sensor nodes using observed radio connectivity =-=[7, 4]-=-. Observations that a pair of nodes are within or not within radio range induce a set of constraints on the locations of the nodes. The problem is then to solve for locations of target nodes, given a   </text>
<query_num> 7305 </query_num>
<text>   es for the link latency after clustering as there are likely to be more observations of paths traversing each induced link. We identify IP aliases using two existing techniques, Mercator [9] and Ally =-=[24]-=-. The Mercator technique works as follows. UDP probes are sent to high-numbered ports on a set of network interfaces. Routers typically send back a port-unreachable ICMP message with the source addres cation Hints Certain routers and end-hosts have DNS names that could be parsed to provide hints regarding where they are located, typically airport codes or abbreviated city names as part of the name =-=[24, 28]-=-. For example, bos-edge-03.inet.qwest.net is likely in Boston. The issue with using router DNS name information is that they could be misnamed, with the naming errors introduced by router reconfigurat   </text>
<query_num> 7306 </query_num>
<text>   etter estimates for the link latency after clustering as there are likely to be more observations of paths traversing each induced link. We identify IP aliases using two existing techniques, Mercator =-=[9]-=- and Ally [24]. The Mercator technique works as follows. UDP probes are sent to high-numbered ports on a set of network interfaces. Routers typically send back a port-unreachable ICMP message with the   </text>
<query_num> 7307 </query_num>
<text>   n the locations of the nodes. The problem is then to solve for locations of target nodes, given a set of reference nodes with known locations. This problem can be formulated as a semidefinite program =-=[25, 7, 4]-=-, allowing the use of powerful solvers such as Sedumi [21]. Our problem setting, however, requires a richer set of constraints. Nearby nodes may not be connected, and backbone links may connect nodes  ,j∈C l subject to : Cd, Cl. |eij| The above problem is not a convex optimization problem, but we follow a formulation from the sensor network community to recast the problem as a semidefinite program =-=[25]-=-, thereby allowing us to use to fast solvers (such as SeDuMi [21]) to perform the optimization. One could also use a modified form of Vivaldi [6] to solve for the coordinates of the routers and the en   </text>
<query_num> 7308 </query_num>
<text>   nd were hampered by a number of significant errors! We subsequently verified all locations to weed out these errors with a combination of USGS data [11] and Google Maps [12]. dinates (such as Vivaldi =-=[6]-=-) and/or network location services (such as Meridian [3]) to first identify a small number of nearby landmarks and then issue probes only from the nearby landmarks to the target. 3.3 Speed of Internet unity to recast the problem as a semidefinite program [25], thereby allowing us to use to fast solvers (such as SeDuMi [21]) to perform the optimization. One could also use a modified form of Vivaldi =-=[6]-=- to solve for the coordinates of the routers and the end-hosts given the topology and delay constraints. We chose not to do so because our experiments with a straightforward implementation of Vivaldi   </text>
<query_num> 7309 </query_num>
<text>   nistrative location, which may not reflect their actual location. Hostnames, especially for routers, may follow naming conventions that include geographic hints in the form of cities or airport codes =-=[23]-=-, but many do not or are misleading [29]. 2.4 Sensor Network Localization Our topology-based techniques are inspired by similar techniques from the sensor network community, originally meant for posit   </text>
<query_num> 7310 </query_num>
<text>   not function well indoors or in urban canyons, limiting its ubiquity, and it would preclude many applications because it is client-driven. Other systems such as Place Lab [13], Cricket [20] and RADAR =-=[1]-=- locate mobile hosts using 802.11 and GSM beacons with known locations. These systems have the potential for accurate estimates but their coverage is limited by the propagation range of APs and cell t   </text>
<query_num> 7311 </query_num>
<text>   of APs and cell towers. Large-scale coverage requires wide-spread and dense deployment of nodes with 802.11 or GSM hardware, at known locations. Systems such as IP2GEO [26], GeoCluster [18], GeoTrack =-=[17]-=-, and Netgeo [27] require no special hardware. However, all these methods depend largely on manually maintained databases and are prone to incomplete coverage, outdated information, and faulty data en   </text>
<query_num> 7312 </query_num>
<text>   t for which we have ground truth information 2 and from which we can make measurements. There are well-known network diversity issues associated with the use of PlanetLab and educational institutions =-=[2]-=-. We expect experiments with this dataset to show delay-based methods in a good light, and hence be appropriate for understanding the extent of their accuracy in controlled settings. This is because o   </text>
</top>
<top>
<paper_num> 74 </paper_num>
<paper_title>   Evolutionary Multi-Objective Optimisation Of Neural Networks For Face Detection.  </paper_title>
<paper_abstract>   For face recognition from video streams speed and accuracy are vital aspects. The first decision whether a preprocessed image region represents a human face or not is often made by a feed-forward neural network (NN), e.g., in the Viisage-FaceFINDER � video surveillance system. We describe the optimization of such a NN by a hybrid algorithm combining evolutionary multi-objective optimization (EMO) and gradient-based learning. The evolved solutions perform considerably faster than an expert-designed architecture without loss of accuracy. We compare an EMO and a single objective approach, both with online search strategy adaptation. It turns out that EMO is preferable to the single objective approach in several respects.  </paper_abstract>
<query_num> 7401 </query_num>
<text>   Bj &amp;gt; 0 ∧ DBj,Ai = 0, 1 ≤ j ≤ T � | · 1 /T , (14) P Ai||B := | � (Ai, Bj) : DAi,Bj &amp;gt; 0 ∧ DBj,Ai &amp;gt; 0, 1 ≤ j ≤ T � | · 1 /T , (15) PBj⊲A := | � (Ai, Bj) : DAi,Bj = 0 ∧ DBj,Ai &amp;gt; 0, 1 ≤ i ≤ T � | · 1 /T , =-=(16)-=- P Bj||A := | � (Ai, Bj) : DAi,Bj &amp;gt; 0 ∧ DBj,Ai &amp;gt; 0, 1 ≤ i ≤ T � | · 1 /T , (17) that is, in (14) the average number of trials from algorithm (B) that perform worse than trial Ai, in (15) the average n   </text>
<query_num> 7402 </query_num>
<text>   Then all individuals a ∈ P (t) ∪ O (t) are sorted in ascending order according to the partial order ≥n defined by � ai ≥n aj ⇔ R (t) (ai) &amp;lt; R (t) � � (aj) ∨ R (t) (ai) = R (t) � (aj) ∧ C(ai) ≥ C(aj) =-=(4)-=- and the first |P| individuals form the new parent population P (t+1) . We refer to the described selection method as NSGA-II selection throughout this article. 2.2.5. Search strategy adaptation: Adju   </text>
<query_num> 7403 </query_num>
<text>   absolute deviation of the quantities ∆Ai,Bj := DAi,Bj − DBj,Ai = −∆Bj,Ai and DAi,Bj . Furthermore we calculate for 1 ≤ i, j ≤ T PAi⊲B := | � (Ai, Bj) : DAi,Bj &amp;gt; 0 ∧ DBj,Ai = 0, 1 ≤ j ≤ T � | · 1 /T , =-=(14)-=- P Ai||B := | � (Ai, Bj) : DAi,Bj &amp;gt; 0 ∧ DBj,Ai &amp;gt; 0, 1 ≤ j ≤ T � | · 1 /T , (15) PBj⊲A := | � (Ai, Bj) : DAi,Bj = 0 ∧ DBj,Ai &amp;gt; 0, 1 ≤ i ≤ T � | · 1 /T , (16) P Bj||A := | � (Ai, Bj) : DAi,Bj &amp;gt; 0 ∧ DBj,   </text>
<query_num> 7404 </query_num>
<text>   an adaptation cycle. The average performance achieved by the operator o over an adaptation cycle is measured by q (t,τ) o := τ−1 � i=0 � a∈O (t−i) o max (0, B (t) (a)) � τ−1 � � �O i=0 (t−i) o � � , =-=(5)-=- where B (t) (a) represents a quality measure proportional to some kind of fitness improvement. This is for the scalar value based selection scheme, case (A), B (t) (a) := Φ(a) − Φ(parent(a)) (6) and   </text>
<query_num> 7405 </query_num>
<text>   ferenceindicator DA,B := HA+B − HB. 29 It reflects the size of the objective space that is weakly dominated by the set A but not by B, see Fig. 3 (right). It holds (DA,B &amp;gt; 0 and DB,A = 0) ⇔ (A ⊲ B) . =-=(12)-=- The coverage difference DA,B also allows to draw conclusions of the form (DA,B = 0) ∧ (DB,A = 0) ⇔ (A = B), and (DA,B &amp;gt; 0) ∧ (DB,A &amp;gt; 0) ⇔ (A||B), where A||B denotes that A and B are incomparable. Fol   </text>
<query_num> 7406 </query_num>
<text>   t. This is for the scalar value based selection scheme, case (A), B (t) (a) := Φ(a) − Φ(parent(a)) (6) and for the vector-valued selection scheme, case (B), B (t) (a) := R (t) (parent(a)) − R (t) (a) =-=(7)-=- respectively, where parent(a) denotes the parent of an offspring a. The operator probabilities p (t+1) o are adjusted every τ generations according to equations ˜p (t+1) � ζ · q o := (t,τ) o /q (t,τ)   </text>
</top>
<top>
<paper_num> 75 </paper_num>
<paper_title>   Feature Subset Selection Using a Genetic Algorithm.  </paper_title>
<paper_abstract>   Many practical pattern classi cation applications require a careful selection of attributes or features (from amuch larger set) to represent the patterns to be classi ed. This feature subset selection problem is a multicriterion optimization problem. We propose a solution to this problem using a genetic algorithm. Our experiments demonstrate the feasibility of this approach for feature subset selection in the automated design of neural network pattern classi ers. 1  </paper_abstract>
<query_num> 7501 </query_num>
<text>   ew, provably convergent, and relatively e cient constructive learning algorithms for multi-category real as well as discrete valued pattern classi cation tasks have begun to appear in the literature [=-=Parekh et al., 1995; Parekh et al., 1996; Yang et al., 1997; Yang et al., 1996-=-]. Many of these algorithms have demonstrated very good performance in terms of reduced network size, learning time, and generalization in a   </text>
<query_num> 7502 </query_num>
<text>   nt, and relatively e cient constructive learning algorithms for multi-category real as well as discrete valued pattern classi cation tasks have begun to appear in the literature [=-=Parekh et al., 1995;Parekh et al., 1996; Yang et al., 1997; Yang et al., 1996-=-]. Many of these algorithms have demonstrated very good performance in terms of reduced network size, learning time, and generalization in a number of experiments   </text>
</top>
<top>
<paper_num> 76 </paper_num>
<paper_title>   Integrating active information delivery and reuse repository systems.  </paper_title>
<paper_abstract>   Although software reuse can improve both the quality and productivity of software development, it will not do so until software developers stop believing that it is not worth their effort to find a component matching their current problem. In addition, if the developers do not anticipate the existence of a given component, they will not even make an effort to find it in the first place.  Even the most sophisticated and powerful reuse repositories will not be effective if developers don&amp;apos;t anticipate a certain component exists, or don&amp;apos;t deem it worthwhile to seek for it. We argue that this crucial barrier to reuse is overcome by integrating active information delivery, which presents information without explicit queries from the user, and reuse repository systems. A prototype system, CodeBroker, illustrates this integration and raises several issues related to software reuse.  </paper_abstract>
<query_num> 7601 </query_num>
<text>   because the same representation is also used by developers to specify their reuse queries, the format of representation is limited by the developers’ willingness to formulate long and precise queries =-=[30]-=-. 2.1.4 Retrieval by Reformulation Retrieved components only match reuse queries. It is difficult for most reusers to create a well-defined query on their first attempt. Retrieval by reformulation is  ies are willing to pay the costs associated with setting up, evolving and sustaining reuse repositories. Traditionally, there are two distinctive roles: producers and consumers of reusable components =-=[30]-=-. Producers of components apply domain analysis and other techniques to identify and develop reusable components, which are then consumed (reused) by developers. This dichotomous perspective on reuse   </text>
<query_num> 7602 </query_num>
<text>   development method of using existing reusable software components to create new systems, has been shown through empirical studies to improve both the quality and productivity of software development =-=[1]-=-. Software reuse also increases the evolvability of software systems because complex systems evolve faster when they are built from stable subsystems [35]. However, before developers can take advantag   </text>
<query_num> 7603 </query_num>
<text>   ey will evolve over time, and this will increase the size of (L4 – L3). Many reports about reuse experience of industrial software companies illustrate this inhibiting factor of reuse. Devanbu et al. =-=[5]-=- report that because developers are unaware of reusable components, they repeatedly re-implement the same function—in one case, this occurred ten times. This kind of behavior is also observed as typic use multiple facets to represent a reusable component and a conceptual distance graph to reflect the semantic relationships among terms describing reusable components. Both CodeFinder [23] and LaSSIE =-=[5]-=- represent reusable components as frames. CodeFinder organizes those frames into an associated network and uses spreading activation to find reusable components. Frames in LaSSIE are organized into hi   </text>
<query_num> 7604 </query_num>
<text>   he introduction of a new word into the English language that increases our ability to reason and communicate, reusable components increase the ability of the developer to create and interpret designs =-=[26]-=-. However, developers must learn the syntax and the semantics of the new vocabulary to take advantage of reuse repositories and to be able to form and express their reuse intentions. Vocabulary learni   </text>
<query_num> 7605 </query_num>
<text>   ion schemas [6] use multiple facets to represent a reusable component and a conceptual distance graph to reflect the semantic relationships among terms describing reusable components. Both CodeFinder =-=[23]-=- and LaSSIE [5] represent reusable components as frames. CodeFinder organizes those frames into an associated network and uses spreading activation to find reusable components. Frames in LaSSIE are or   </text>
<query_num> 7606 </query_num>
<text>   ly challenge the assumed dichotomy of producers and consumers of reuse repository systems, and provide an alternative way to the production (setting up, evolving and sustaining) of reuse repositories =-=[11]-=-. One important factor in the success of OSS is to get users involved as co-developers. We must provide immediate benefits to the developers so that they will help build better repository systems. The   </text>
<query_num> 7607 </query_num>
<text>   m the editor, and can be reused without further modification. 4.2 Fetcher The retrieval mechanism used by Fetcher is the combination of Latent Semantic Analysis (LSA) [27] and Signature Matching (SM) =-=[39]-=-. LSA is used to compute the conceptual similarity between conceptual queries and functional descriptions of reusable components. SM is used to determine the constraint compatibility between constrain ieve reusable components on the basis of conceptual content. 4.2.2 Signature Matching Signature matching is the process of determining the compatibility of two components in terms of their signatures =-=[39]-=-. It is an indexing and retrieval mechanism based on type constraints. The basic form of a signature of a method is: Method:InTypeExp-&amp;gt;OutTypeExp where InTypeExp and OutTypeExp are type expressions re  difficult to construct such systems because they need to formalize expert knowledge that is difficult to elicit, and they are difficult to scale. Formal method-based approaches use either signatures =-=[39]-=-, or formal specifications [40] to represent components. Signaturebased approaches are easy to use, but suffer from the impreciseness of representation; formal specification approaches are time consum   </text>
<query_num> 7608 </query_num>
<text>   of reusable components, they repeatedly re-implement the same function—in one case, this occurred ten times. This kind of behavior is also observed as typical among the four companies investigated in =-=[9]-=-. From the experience of promoting reuse, Rosenbaum and DuCastel [34] conclude that making components known to developers is a key factor for successful reuse. Developers will reuse those components r s structural indentation, comments, and identifier names. Comments and identifier names are important beacons for the understanding of programs, because they reveal the important concepts of programs =-=[2, 9, 29]-=-. One important constraint of a program is its type compatibility, Developer Working Products Listener Analyzes Program Editor RCI-display Refine Conceptual Queries Constraint Queries Update Automatic   </text>
<query_num> 7609 </query_num>
<text>   s structural indentation, comments, and identifier names. Comments and identifier names are important beacons for the understanding of programs, because they reveal the important concepts of programs =-=[2, 9, 29]-=-. One important constraint of a program is its type compatibility, Developer Working Products Listener Analyzes Program Editor RCI-display Refine Conceptual Queries Constraint Queries Update Automatic evelopers choose components from different repositories by providing different views. It uses name matching and information retrieval-based similarity matching to identify similar reusable components =-=[29]-=-. 6. DISCUSSION Software development is difficult because of the irreducible essential difficulties brought by the complexity, conformity,schangeability and invisibility of software systems [3]. Softw   </text>
<query_num> 7610 </query_num>
<text>   stems because they need to formalize expert knowledge that is difficult to elicit, and they are difficult to scale. Formal method-based approaches use either signatures [39], or formal specifications =-=[40]-=- to represent components. Signaturebased approaches are easy to use, but suffer from the impreciseness of representation; formal specification approaches are time consuming and cognitively challenging   </text>
</top>
<top>
<paper_num> 77 </paper_num>
<paper_title>   Document-Word Co-regularization for Semi-supervised Sentiment Analysis.  </paper_title>
<paper_abstract>   The goal of sentiment prediction is to automatically identify whether a given piece of text expresses positive or negative opinion towards a topic of interest. One can pose sentiment prediction as a standard text categorization problem. However, gathering labeled data turns out to be a bottleneck in the process of building high quality text classifiers. Fortunately, background knowledge is often available in the form of prior information about the sentiment polarity of words in a lexicon. Moreover, in many applications abundant unlabeled data is also available. In this paper, we propose a novel semi-supervised sentiment prediction algorithm that utilizes lexical prior knowledge in conjunction with unlabeled examples. Our method is based on joint sentiment analysis of documents and words based on a bipartite graph representation of the data. We present an empirical study on a diverse collection of sentiment prediction problems which confirms that our semi-supervised lexical models significantly outperform purely supervised and competing semisupervised techniques.  </paper_abstract>
<query_num> 7701 </query_num>
<text>   a) unsupervised lexical classification (LEX) which gives a baseline, (b) Linear SVMs which are considered state-of-the-art for text classification, and (c) two implementations of the Transductive SVM =-=[14, 26]-=-, one based on label switching (TSVM) and another based on deterministic annealing (DA) [26]. We carefully tune the regularization parameter for linear SVMs (in the range γ = c l where c = {0.001,0.01 meant to represent the best possible results one can hope to obtain with a state-ofthe-art purely supervised learner. We report the best performance of TSVM and DA over the parameter settings used in =-=[26]-=-. Furthermore, TSVM and DA require an accurate estimate of positive class fraction. In practical semisupervised settings, a noisy estimate of this fraction is obtained from the labeled data. In our ex   </text>
<query_num> 7702 </query_num>
<text>   ce out-of-sample predictions through a linear model. Finally, our algorithm is also closely related to a large class of multi-view learning algorithms, in particular the co-regularization approach of =-=[27]-=-.5 Empirical Study 5.1 Data sets In order to test the generality of our approach we experimented on three qualitatively different domains — blogs discussing enterprise software, blogs about US Presid   </text>
<query_num> 7703 </query_num>
<text>   ication algorithms implement the classical cluster assumption [3] which states the following: if two documents are in the same cluster, they are likely to be of the same class. Low-density techniques =-=[14, 5, 4]-=- implement this assumption by attempting to find separators that do not cut thru unlabeled data clusters. Similarly, Graph-based techniques [15, 34, 1] use unlabeled examples to find classifiers that  he practical rank of Q) in practice. We call our approach Semi-supervised Lexical Regularized Least Squares (SS+LEX+RLS) classification. Advantages of the Proposed Algorithm: Unlike Transductive SVMs =-=[14, 4]-=- our algorithm is based on convex optimization and therefore does not suffer from local minima issues. Unlike, typical graph-based methods [15, 34, 1] which require an expensive construction of a near   </text>
<query_num> 7704 </query_num>
<text>   ication algorithms implement the classical cluster assumption [3] which states the following: if two documents are in the same cluster, they are likely to be of the same class. Low-density techniques =-=[14, 5, 4]-=- implement this assumption by attempting to find separators that do not cut thru unlabeled data clusters. Similarly, Graph-based techniques [15, 34, 1] use unlabeled examples to find classifiers that  he practical rank of Q) in practice. We call our approach Semi-supervised Lexical Regularized Least Squares (SS+LEX+RLS) classification. Advantages of the Proposed Algorithm: Unlike Transductive SVMs =-=[14, 4]-=- our algorithm is based on convex optimization and therefore does not suffer from local minima issues. Unlike, typical graph-based methods [15, 34, 1] which require an expensive construction of a near a) unsupervised lexical classification (LEX) which gives a baseline, (b) Linear SVMs which are considered state-of-the-art for text classification, and (c) two implementations of the Transductive SVM =-=[14, 26]-=-, one based on label switching (TSVM) and another based on deterministic annealing (DA) [26]. We carefully tune the regularization parameter for linear SVMs (in the range γ = c l where c = {0.001,0.01   </text>
<query_num> 7705 </query_num>
<text>   is area is the use and generation of dictionaries capturing the sentiment of words. These methods range from manual approaches of developing domain-dependent lexicons [7] to semi-automated approaches =-=[13, 35, 17]-=-, and even an almost fully automated approach [30]. As observed by Ng et al. [19], most semi-automated approaches yield unsatisfactory lexicons, with either high coverage and low precision or vice ver   </text>
<query_num> 7706 </query_num>
<text>   lean meaningful insights therein. One key component of this process is to be able to gauge the sentiment expressed in blogs around selected topics of interest. The emerging area of Sentiment Analysis =-=[31, 21, 13]-=- focuses on this task of automatically identifying whether a piece of text expresses a positive or negative opinion towards the subject matter. Detecting the sentiment expressed in documents turns out   </text>
<query_num> 7707 </query_num>
<text>   lean meaningful insights therein. One key component of this process is to be able to gauge the sentiment expressed in blogs around selected topics of interest. The emerging area of Sentiment Analysis =-=[31, 21, 13]-=- focuses on this task of automatically identifying whether a piece of text expresses a positive or negative opinion towards the subject matter. Detecting the sentiment expressed in documents turns out t classify the sentiment of texts based on lexicons defining the sentiment-polarity of words, and simple linguistic patterns. There have been some recent studies that take a machine learning approach =-=[21, 11]-=-, and build text classifiers trained on documents that have been human-labeled as positive or negative. The knowledgebased approaches tend to be non-adaptive, while the learning approaches do not effe mated approach [30]. As observed by Ng et al. [19], most semi-automated approaches yield unsatisfactory lexicons, with either high coverage and low precision or vice versa. More recently, Pang et al. =-=[21]-=- successfully applied a machine learning approach to classifying sentiment for movie reviews. They cast the problem as a text classification task, using a bag-of-words representation of each movie rev m “clinton” and 1000 containing “obama” in their URLs. Movie reviews: Apart from the blog data that we collected, we also used the publicly available data set of movie reviews provided by Pang et al. =-=[21]-=-. This data set consists of 1000 positive and 1000 negative reviews from the Internet Movie Database. Positive labels were assigned to reviews that had a rating above 3.5 stars and negative labels wer generally considered negative in other contexts. The down-weighting of positive lexicon terms, such as talent for MOVIES is also consistent with the “thwarted expectation” narratives that Pang et al. =-=[21]-=- observed in this data. In Tables 3 and 4, we look at words not in the lexicon that recieve the highest positive and negative sentiment polarity scores as given by the magnitude of fw i . It is clear   </text>
<query_num> 7708 </query_num>
<text>   methods range from manual approaches of developing domain-dependent lexicons [7] to semi-automated approaches [13, 35, 17], and even an almost fully automated approach [30]. As observed by Ng et al. =-=[19]-=-, most semi-automated approaches yield unsatisfactory lexicons, with either high coverage and low precision or vice versa. More recently, Pang et al. [21] successfully applied a machine learning appro   </text>
<query_num> 7709 </query_num>
<text>   nerated from a list of relevant features to label pseudo-examples. They modify the boosting objective function to fit the training data, and the prior model based on these pseudo-examples. Liu et al. =-=[18]-=- use background knowledge to generate labeled training examples from a large pool of unlabeled examples. In their work, they focus on the process of selecting features to be labeled by humans — using   </text>
<query_num> 7710 </query_num>
<text>   orating parts-of-speech and n-gram language models, do not improve over the simple unigram bag-of-words representation. In keeping with their findings, we also adopt a unigram text model. Pang et al. =-=[20]-=- extend their work, by first classifying sentences as subjective versus objective, and then classifying only the subjective sentences based on sentiment polarity. They demonstrate that by focusing onl   </text>
<query_num> 7711 </query_num>
<text>   proach to also include unlabeled data. 1 The bias weight is often excluded from the regularizer‖w‖ 2 , though including it brings about simplifications without any performance consequences. See, e.g.,=-=[16]-=- for a discussion.4.1 Incorporating Lexical Knowledge in Supervised Regularization Models It is well-known that RLS may be interpreted as maximum aposteriori (MAP) estimation under a Gaussian likelih   </text>
<query_num> 7712 </query_num>
<text>   ta dependent norm equals α 2 i r(λi) for an increasing function r(·). Different choices of r enforce different amounts of smoothness generating a whole family of graph-based smoothness operators. See =-=[28]-=- for typical choices. In particular, we use M = ˜ L p which corresponds to the choice r(λ) = λ p (note that the explicit eigendecomposition of ˜ L is not required) where p parameterizes the amount of   </text>
<query_num> 7713 </query_num>
<text>   ted to ours. However, theyreport that this method was unreliable; since, it occasionally produced the best, but usually produced the worst results compared to other approaches. Finally, Druck et al. =-=[10]-=- incorporate prior knowledge through labeled features, which are used to directly constrain the model’s predictions on unlabeled instances. Their Generalized Expectation criteria approach is applicabl   </text>
<query_num> 7714 </query_num>
<text>   the subject matter. Detecting the sentiment expressed in documents turns out be an extremely difficult task, and the performance of sentiment classifiers can vary a great deal depending on the domain =-=[30]-=-. As such, one of the grand challenges of sentiment analysis is to build a robust system that provides insights across a growing list of different products and topics of interest. Such a system needs  g the sentiment of words. These methods range from manual approaches of developing domain-dependent lexicons [7] to semi-automated approaches [13, 35, 17], and even an almost fully automated approach =-=[30]-=-. As observed by Ng et al. [19], most semi-automated approaches yield unsatisfactory lexicons, with either high coverage and low precision or vice versa. More recently, Pang et al. [21] successfully a   </text>
<query_num> 7715 </query_num>
<text>   tures to particular classes. Most of this work has focused on using such prior class-bias of features to generate labeled examples that are then used for standard supervised learning. Schapire et al. =-=[24]-=- propose one such framework for boosting logistic regression, that uses handcrafted rules generated from a list of relevant features to label pseudo-examples. They modify the boosting objective functi   </text>
<query_num> 7716 </query_num>
<text>   uster assumptions while incorporating semi-supervision along both dimensions. We begin by introducing a bipartite graph representation of the data, previously utilized in the context of co-clustering =-=[9]-=-. We then formulate joint sentiment classification of documents and words in terms of transductive prediction on this graph whose nodes are viewed to be partially labeled. However, since this approach hird property can be enforced also over unlabeled documents and unlabeled words. It turns out that the third property has close connections to the classical SVD applied to document-term matrices (see =-=[9]-=- for more details). These three properties can be enforced through the terms of the objective function in the following minimization problem, argmin f d ,f w 1 ld i=1 ld∑ V (f d i ,y d i ) + 1 +µ n∑ i  document-word associations in order to use lexical prior knowledge in the presence of unlabeled examples. Joint document-word analysis has previously been explored in the context of co-clustering in =-=[9]-=-. Our algorithm may be seen as providing two additional capabilities on top of the bipartite co-clustering approach: (a) the capibility to use semi-supervision for both document and words, and (b) the   </text>
<query_num> 7717 </query_num>
<text>   y to be of the same class. Low-density techniques [14, 5, 4] implement this assumption by attempting to find separators that do not cut thru unlabeled data clusters. Similarly, Graph-based techniques =-=[15, 34, 1]-=- use unlabeled examples to find classifiers that give smooth predictions on data clusters. In the presence of lexical knowledge we may further qualify the cluster assumption as follows: if two documen nsider the problem of completing the labeling of the rest of the vertices of the graph. Such prediction problems on graphs have been wellstudied in the graph-based semi-supervised learning literature =-=[15, 34, 1]-=-, but to the best of our knowledge they have never been applied to solve joint prediction problems on document and words. Our goal is to learn a real-valued sentiment-polarity score vector, f d , over s of the Proposed Algorithm: Unlike Transductive SVMs [14, 4] our algorithm is based on convex optimization and therefore does not suffer from local minima issues. Unlike, typical graph-based methods =-=[15, 34, 1]-=- which require an expensive construction of a nearest neighbor graph, our algorithm uses regularization operators defined on the biparite document-word graph. Thus, there is no expensive graph constru   </text>
</top>
<top>
<paper_num> 78 </paper_num>
<paper_title>   Towards an active network architecture.  </paper_title>
<paper_abstract>   Active networks allow users to inject customized programs into the nodes of the network. In this paper, we describe our vision of an active network architecture, outline our approach to its design, and survey the technologies that can be brought to bear on its implementation. In the course of this presentation we identify a number of research questions to be addressed and propose that the research community mount a joint effort to develop and deploy a wide area ActiveNet. 1. INTRODUCTION Traditional data networks passively transport bits from one end system to another. Ideally, the user data is transferred opaquely, i.e., the network is insensitive to the bits it carries and they are transferred between end systems without modification. The role of computation within such networks is extremely limited, e.g., header processing in packet-switched networks and/or signaling in connection-oriented networks. Active Networks break with tradition by allowing the network to perform customized c...  </paper_abstract>
<query_num> 7801 </query_num>
<text>   allow the compiler to generate code that is more efficient. In both cases, we would expect the directly executable binary code to out-perform an interpreted format. On-the-fly Compilation Recent work =-=[18]-=- has enabled &amp;quot;on-the-fly&amp;quot; compilation with a dialect of the C programming language. This allows source programs to be automatically tailored, or even wholly generated, at run-time. In conjunction with   </text>
<query_num> 7802 </query_num>
<text>   ance of TCP connections by retaining per-connection state information at wireless base stations. Application-specific services performed at gateways include file caching and the transcoding of images =-=[4]-=-. The InfoPad [5] takes the process even further, by instantiating user-specific &amp;quot;pad servers&amp;quot; supporting a range of applications, such as voice and hand-writing recognition, at intermediate nodes. Ne   </text>
<query_num> 7803 </query_num>
<text>   ardware activities, such as the development of switching technology that &amp;quot;caches&amp;quot; fast paths and is highly responsive to active capsules. What about the end-to-end argument? The &amp;quot;end-to-end argument&amp;quot; =-=-=-ot;end-to=end argument&amp;quot; [28] concerns -=-the design of intermediaries, such as networks, that provide services that cannot be made perfectly reliable. 1 Since users of these services must provide &amp;quot;end-to-end&amp;quot; mechanisms that cope w   </text>
<query_num> 7804 </query_num>
<text>   as those described in [NOW] might be used to &amp;quot;page&amp;quot; soft state to/from nearby nodes. accessing replicated storage services located elsewhere on the network. 1 We hope to leverage technologies such as =-=[20]-=- for this purpose. Logical Resources Although there are a relatively small number of physical resources, a node may support a large number of logical resources of many different types. This suggests t   </text>
<query_num> 7805 </query_num>
<text>   egic points that bridge networks with vastly different bandwidth and reliability characteristics, such as the junctions between wired and wireless networks. Application-neutral work on TCP &amp;quot;snooping&amp;quot; =-=[3]-=- improves the performance of TCP connections by retaining per-connection state information at wireless base stations. Application-specific services performed at gateways include file caching and the t   </text>
<query_num> 7806 </query_num>
<text>   nment. When a program is presented for execution, the run-time system verifies that the instruction sequence was generated by a trusted compiler and has not been modified. . The approach described in =-=[16, 17]-=- prescribes a set of rules that instruction sequences must adhere to, such as restrictions on how address arithmetic is performed. In conjunction with a modicum of run-time support and a collection of   </text>
<query_num> 7807 </query_num>
<text>   or complex, did not address the safety issues relevant to shared infrastructures. The advent of heterogeneous distributed systems and internetworking has accelerated the pace of research. The xkernel =-=[9]-=- supports the composition of protocol handlers by providing a regular architecture for stacking them and by automating the dispatch process. Other efforts [10-12] have focused on less friendly environ   </text>
<query_num> 7808 </query_num>
<text>   sly considered within the context of time-sharing systems [22, 23] but we are not aware of work that addresses delegation in as complex a system as the ActiveNet. Work on the cascading [24] and logic =-=[25]-=- of authentication, which has some of the delegation flavor we are looking for, may provide a starting point for further research. Ultimately, this may be one of the most important &amp;quot;open&amp;quot; questions wi   </text>
<query_num> 7809 </query_num>
<text>   to the firewall and inject the appropriate modules into it. Web Proxies Web proxies are an example of an application-specific service that is tailored to the caching of World Wide Web pages. Harvest =-=[1]-=- employs a hierarchical caching scheme that can reduce the latencies experienced by individual users and the aggregate bandwidth that is consumed. The cache nodes are presently located near the edges   </text>
</top>
<top>
<paper_num> 79 </paper_num>
<paper_title>   Managing and querying transaction-time databases under schema evolution.  </paper_title>
<paper_abstract>   The old problem of managing the history of database information is now made more urgent and complex by fast spreading web information systems, such as Wikipedia. Our PRIMA system addresses this difficult problem by introducing two key pieces of new technology. The first is a method for publishing the history of a relational database in XML, whereby the evolution of the schema and its underlying database are given a unified representation. This temporally grouped representation makes it easy to formulate sophisticated historical queries on any given schema version using standard XQuery. The second key piece of technology is that schema evolution is transparent to the user: she writes queries against the current schema while retrieving the data from one or more schema versions. The system then performs the labor-intensive and error-prone task of rewriting such queries into equivalent ones for the appropriate versions of the schema. This feature is particularly important for historical queries spanning over potentially hundreds of different schema versions and it is realized in PRIMA by (i) introducing Schema Modification Operators (SMOs) to represent the mappings between successive schema versions and (ii) an XML integrity constraint language (XIC) to efficiently rewrite the queries using the constraints established by the SMOs. The scalability of the approach has been tested against both synthetic data and real-world data from the Wikipedia DB schema evolution history. 1.  </paper_abstract>
<query_num> 7901 </query_num>
<text>   MV-documents use a temporally grouped data model [8] that naturally and concisely represents temporal data and also simplify the task of expressing powerful temporal queries using standard XQuery. In =-=[27]-=-, it also has been shown that advanced RDBMS technology can be 1 PRIMA stands for Panta Rhei Information Management &amp; ArchivalTable 1: Schema evolution in an employee database (V1 to V5) Schema Versi poral extensions to the relational data model, the temporal difficulties of the relational model and SQL dissolve when the database information is viewed and queried using XML and its query languages =-=[21, 27]-=-. In particular, in [27] it was shown that the history of a relational database can be published in XML, and viewed under a temporally grouped representation whereby complex historical queries can be  These temporal queries and views can be managed in any DBMS that provides native support for XML and XQuery, or more efficiently managed in a Relational DBMS that supports the SQL/XML standards as in =-=[27]-=-. We describe this XML-based temporal model, called V-document, and show how we extend it to model and query temporal databases with evolving schemas. 2.2.1 V-Document In Figure 2, an example V-docume these in turn contain those of the column-name elements under them. XQuery can be used as an effective temporal query language, over this representation, without requiring extensions to the standards =-=[27]-=-. This is due to (i) the expressive power of XQuery which is Turing-complete [14], and (ii) the fact that by supporting a temporally grouped representation XML reduces the need for coalescing [8]. We  the situation dramatically by building on two recent advances on database technology, whereby: • XML and XQuery are i) well-supported in DBMS and ii) very conducive to temporal information management =-=[13, 18, 21, 27]-=-. • Powerful mapping techniques are now emerging [12, 26, 29, 30, 6] which have made it possible to map a query expressed against one schema into equivalent ones expressed on other schemas. By exploit   </text>
<query_num> 7902 </query_num>
<text>   Rondo, model management is often feasible by using a small body of high-level code; however, the problem of evolving schema versions is not discussed in [17]. Lastly, we mention Panta Rhei framework =-=[11]-=- that seeks to provide an integrated support for schema evolution. In this framework, PRISM [10] supports the DBAs in the schemaevolution process by managing and preserving user-provided SMOs to auto   </text>
<query_num> 7903 </query_num>
<text>   an efficient manner, by means of query reformulation. Taxonomy-based Query Reformulation. Query reformulation, recently, has been studied in the context of snapshot data with a single source version =-=[12, 29]-=-. We, however, address a different and more challenging problem of rewriting temporal queries into equivalent ones that will be evaluated against one or more source versions of evolv2 This feature is  ics We have database history under multiple schema versions where we let the user query the database history as if it were under a single schema version that she queried, or the target version. As in =-=[29]-=-, we define the semantics of query answering as following: given a target version Vk, we migrate each database Di under Vi (i ≤ k) to Vk according to the schema mappings and obtain Di→k valid on Vk. W database technology, whereby: • XML and XQuery are i) well-supported in DBMS and ii) very conducive to temporal information management [13, 18, 21, 27]. • Powerful mapping techniques are now emerging =-=[12, 26, 29, 30, 6]-=- which have made it possible to map a query expressed against one schema into equivalent ones expressed on other schemas. By exploiting these recent advances of database technology PRIMA aims at combi   </text>
<query_num> 7904 </query_num>
<text>   an efficient manner, by means of query reformulation. Taxonomy-based Query Reformulation. Query reformulation, recently, has been studied in the context of snapshot data with a single source version =-=[12, 29]-=-. We, however, address a different and more challenging problem of rewriting temporal queries into equivalent ones that will be evaluated against one or more source versions of evolv2 This feature is  n translate these SMOs into XICs which we use for query reformulation XML Integrity Constraint Within our transactiontime database, which is modeled in XML, we employ XML Integrity Constraints (XICs) =-=[12]-=- as our mapping language. XIC is a language for expressing intra- or inter-schema integrity constraints in XML, using a relational representation. It is similar to first-order logic, except that it us  with which an input query is expanded to another equivalent query under the given set of rules. For chase execution, rather than reinventing the wheel, we adopt an efficient chase engine called MARS =-=[12]-=-. With chase at the core, we reformulate the input query Q in three stages: chase preparation, chase, and output XQuery reconstruction. Chase Preparation We first translate SMOs into XICs as discussed de MARS with the target version where query is posed and the source versions to rewrite the query. Target version comes from the 7 Note that XQuery is composed of navigation part and tagging template =-=[12]-=-user, while the source version is provided by MinSourceFind algorithm. Chase With the input of XBind query, XIC, and a source version to rewrite to, we perform a chase, to rewrite an input query on t database technology, whereby: • XML and XQuery are i) well-supported in DBMS and ii) very conducive to temporal information management [13, 18, 21, 27]. • Powerful mapping techniques are now emerging =-=[12, 26, 29, 30, 6]-=- which have made it possible to map a query expressed against one schema into equivalent ones expressed on other schemas. By exploiting these recent advances of database technology PRIMA aims at combi   </text>
<query_num> 7905 </query_num>
<text>   database technology, whereby: • XML and XQuery are i) well-supported in DBMS and ii) very conducive to temporal information management [13, 18, 21, 27]. • Powerful mapping techniques are now emerging =-=[12, 26, 29, 30, 6]-=- which have made it possible to map a query expressed against one schema into equivalent ones expressed on other schemas. By exploiting these recent advances of database technology PRIMA aims at combi   </text>
<query_num> 7906 </query_num>
<text>   different and more challenging problem of rewriting temporal queries into equivalent ones that will be evaluated against one or more source versions of evolv2 This feature is called schema versioning =-=[22]-=-. ing data. To address this problem, we propose a novel solution capable of finding correct and efficient rewritten queries based on a query taxonomy: we characterize an input temporal query, based on  letting users specify the schema version for query writing using SQL extension, and iii) supporting the query by migrating data to the queried version [16, 4, 7, 25]. Comprehensive survey appears in =-=[22, 20]-=-. These are valid solutions, as long as the data that need to be migrated remain relatively small in size. We provide a more practical solution for the problem by means of query reformulation, instead for transaction-time databases had long been viewed as a solution to the schema-evolution problem that, although ideal in theory, in practice could not be effectively realized for real-life databases =-=[22]-=-. Our PRIMA prototype is changing the situation dramatically by building on two recent advances on database technology, whereby: • XML and XQuery are i) well-supported in DBMS and ii) very conducive t   </text>
<query_num> 7907 </query_num>
<text>   e problem of evolving schema versions is not discussed in [17]. Lastly, we mention Panta Rhei framework [11] that seeks to provide an integrated support for schema evolution. In this framework, PRISM =-=[10]-=- supports the DBAs in the schemaevolution process by managing and preserving user-provided SMOs to automate the tasks of data migration, legacy application query rewriting, and schema history recordi   </text>
<query_num> 7908 </query_num>
<text>   ing databases, and they support versioning for their content (text and multimedia). In addition to content evolution [3], these systems experience intense evolution of database schema: as reported in =-=[9]-=-, Wikipedia has experienced more than 170 schema changes in its 4.5 years of lifetime. Thus, schema evolution, which represents a serious problem for traditional information systems [24, 15], is even  tself needs to be efficient also. This turns out to be another serious challenge, as the real-world data, such as Wikipedia schema evolution, insist that we need to handle hundreds of schema versions =-=[9]-=-. In order to perform query rewriting between two schema versions that are hundreds of versions away, within a practical time bound, we propose two optimization techniques in Section 5. Experimental S ersion is introduced, as shown in Table 1. It has been shown that complex schema evolution scenario in Wikipedia schema’s 171 versions, can be completely and naturally described using these operators =-=[9]-=-. As shown in the following sections, SMOs are also supportive of efficient storage of historical data under schema evolution (Section 2.2.3) and efficient query reformulation (Section 5). 2.2 XML-Bas w.r.t. real-life schema evolution and measure the rewriting time of our system against a long schema evolution history. The MediaWiki schema evolution has been expressed in terms of SMOs, as shown in =-=[9]-=-. For queries, we derived real-world query workload to the Wikipedia site from the Wikipedia on-line profiler 15 and use the 20 most common queries in the Wikipedia. Since they are in SQL, we translat   </text>
<query_num> 7909 </query_num>
<text>   instead of data migration, into the source versions that are minimized by careful analysis of input queries. More recently, schema evolution has also been studied in the framework of model management =-=[5, 17]-=-, where model refers to metadata of various types, including relational and XML schemas, SQL view definitions, and mediator specifications. In [17], Melnik et al. present a prototype system called Ron   </text>
<query_num> 7910 </query_num>
<text>   instead of data migration, into the source versions that are minimized by careful analysis of input queries. More recently, schema evolution has also been studied in the framework of model management =-=[5, 17]-=-, where model refers to metadata of various types, including relational and XML schemas, SQL view definitions, and mediator specifications. In [17], Melnik et al. present a prototype system called Ron model management. The authors show that, using Rondo, model management is often feasible by using a small body of high-level code; however, the problem of evolving schema versions is not discussed in =-=[17]-=-. Lastly, we mention Panta Rhei framework [11] that seeks to provide an integrated support for schema evolution. In this framework, PRISM [10] supports the DBAs in the schemaevolution process by mana   </text>
<query_num> 7911 </query_num>
<text>   the situation dramatically by building on two recent advances on database technology, whereby: • XML and XQuery are i) well-supported in DBMS and ii) very conducive to temporal information management =-=[13, 18, 21, 27]-=-. • Powerful mapping techniques are now emerging [12, 26, 29, 30, 6] which have made it possible to map a query expressed against one schema into equivalent ones expressed on other schemas. By exploit   </text>
<query_num> 7912 </query_num>
<text>   used as an effective temporal query language, over this representation, without requiring extensions to the standards [27]. This is due to (i) the expressive power of XQuery which is Turing-complete =-=[14]-=-, and (ii) the fact that by supporting a temporally grouped representation XML reduces the need for coalescing [8]. We next present examples of temporal queries on our employee database. For these que   </text>
<query_num> 7913 </query_num>
<text>   ving schema versions by means of timestamps, ii) letting users specify the schema version for query writing using SQL extension, and iii) supporting the query by migrating data to the queried version =-=[16, 4, 7, 25]-=-. Comprehensive survey appears in [22, 20]. These are valid solutions, as long as the data that need to be migrated remain relatively small in size. We provide a more practical solution for the proble   </text>
</top>
<top>
<paper_num> 80 </paper_num>
<paper_title>   Secure verification of location claims.  </paper_title>
<paper_abstract>   With the growing prevalence of sensor and wireless networks comes a new demand for location-based access control mechanisms. We introduce the concept of secure location verification, and we show how it can be used for location-based access control. Then, we present the Echo protocol, a simple method for secure location verification. The Echo protocol is extremely lightweight: it does not require time synchronization, cryptography, or highly accurate clocks. Hence, we believe that it is well suited for use in small, cheap, mobile devices.  </paper_abstract>
<query_num> 8001 </query_num>
<text>   [4] as a defense against man-in-the-middle attacks on cryptographic identification schemes. Hu, et al., proposed using temporal packet leashes for wireless networks to defend against similar attacks =-=[12]-=-. However, a major limitation of these schemes is that both the prover and verifier send RF signals, requiring the access to a much more accurate timing system at the verifier as well as tight real-ti   </text>
<query_num> 8002 </query_num>
<text>   ant hardware. Location-limited channels provide a communication mechanism that is restricted to a short range and provides both endpoints a mechanism to guarantee the authenticity of each participant =-=[16]-=-. Balfanz, et al., have proposed using location-limited channels for locationbased access control [3], and many others have also proposed use of limited-range radio broadcasts as a way to verify proxi   </text>
<query_num> 8003 </query_num>
<text>   ease uncertainty by about 1/3 of a meter. If the prover and verifier are 50m apart, the protocol runtime is about 150ms; clock skew is unlikely to be more than a few microseconds during this interval =-=[8]-=-, so the uncertainty added would be on the order of millimeters, which is acceptable for our application domain. 5 Related Work A number of authors have proposed using time-offlight measurements and t   </text>
<query_num> 8004 </query_num>
<text>   fanz, et al., have proposed using location-limited channels for locationbased access control [3], and many others have also proposed use of limited-range radio broadcasts as a way to verify proximity =-=[13, 6, 5]-=-. However, there are no strong security guarantees that the communication range will always be limited as desired: an adversary with more powerful equipment may be able to participate in the protocols algorithm. Thus, insecure localization protocols should be seen as complementary to our work on secure location verification. Many authors have commented on the value of location-based access control =-=[7, 5, 3, 13, 6]-=-. 6 Future Work One area for future work is performing more precise region verification using intersection. The idea is that if verifier v1 can verify ROA(v1, #) and v2 can verify ROA(v2, #), then tog   </text>
<query_num> 8005 </query_num>
<text>   more powerful equipment may be able to participate in the protocols even if they are substantially further away than non-malicious parties. Finally, there are many techniques to help localize devices =-=[2, 14, 15, 10, 18, 1]-=-, GPS being one of the most widely deployed. However, none of those works addressed security, and in fact, GPS signals can be spoofed [17, 3.2.2]. Nonetheless, we have noted that combining a localizat   </text>
<query_num> 8006 </query_num>
<text>   onsequence of considering this domain is that many techniques, such as public-key cryptography, are infeasible. The Berkeley Mica sensor nodes, for example, have 4MHz 8-bit processors with 4KB of RAM =-=[11]-=-. What we need, then, is a lightweight way to perform location verification given many sensor-class nodes. The principal trying to prove its location need not be a sensor-class node, though we do not   </text>
<query_num> 8007 </query_num>
<text>   t range and provides both endpoints a mechanism to guarantee the authenticity of each participant [16]. Balfanz, et al., have proposed using location-limited channels for locationbased access control =-=[3]-=-, and many others have also proposed use of limited-range radio broadcasts as a way to verify proximity [13, 6, 5]. However, there are no strong security guarantees that the communication range will a algorithm. Thus, insecure localization protocols should be seen as complementary to our work on secure location verification. Many authors have commented on the value of location-based access control =-=[7, 5, 3, 13, 6]-=-. 6 Future Work One area for future work is performing more precise region verification using intersection. The idea is that if verifier v1 can verify ROA(v1, #) and v2 can verify ROA(v2, #), then tog   </text>
</top>
<top>
<paper_num> 81 </paper_num>
<paper_title>   A Framework for Multi-Agent Belief Revision.  </paper_title>
<paper_abstract>   c ○ Copyright by Wei Liu (M.Eng), 2002 I hereby certify that the work embodied in this thesis is that the result of original research and has not been submitted for a higher degree to any other University or Institution. ii  </paper_abstract>
<query_num> 8101 </query_num>
<text>   96 that their work[68] is more amenable to solve the heterogeneity problem than Van der Meyden’s by stating that agent knowledge could fall into private and shared domains. Recently, other researchers=-=[78]-=-[102], following Fagin et al.’s approach, defined various concepts such as team knowledge and shared knowledge. Dragoni et al. also distinguishes global knowledge from local knowledge[20]. Actually, d   </text>
<query_num> 8102 </query_num>
<text>   Nwana and Ndumn[89], they do not in themselves offer solutions to multi-agent systems for the following reasons: • Distributed computing modules are designed to be passive/cooperative and homogeneous=-=[25]-=-. 70s• The communications between distributed computing modules are usually lowlevel while multi-agent systems require high-level messages. • Multi-agent system applications require a cooperative-know ef (c) for his further decision making. Assume agent x himself uses the above mechanism for γ to defraud others, then he is likely to make the assumption that y is using similar mechanisms for himself=-=[25]-=-. A reasonable guess can be achieved if we assume, degree of deception ≡ 1 - degree of sincerity (i.e. γ ≡ 1 − s) Substituting γ with 1−s in the above definition sentence for γ, which is the subjectiv   </text>
<query_num> 8103 </query_num>
<text>   and filter the often large retrieved repository for relevant information. The next generation ontologies, will facilitate the revolution of the world wide web into Berners-Lee’s vision of semantic web=-=[7]-=- 1 , which makes a huge amount of information available in machine-readable format. This will not only free human-beings from the information overflow but also enable the automatic online B2B trading.  agents interact with an open system sharing intrinsically the same ontology is therefore not realistic. This is also one of the main driving forces for current research activities in the Semantic Web=-=[7]-=- and service description languages[84]. An ontology based communication envisaged by FIPA is illustrated in Figure 6.1[29]. In a more general and ambitious setting, FIPA’s ontology service specificati   </text>
<query_num> 8104 </query_num>
<text>   credible and not honest) • G4: Nonsense Agents (agents who are neither truthful nor credible) 7.3 Trustworthiness and Degree of Beliefs In the fields of data fusion[3], weighted knowledge base merging=-=[76]-=-, multiple source conflict resolution [19] and multi-agent belief revision[68], the value of trustworthiness is always assumed to be given. In this section, based on the observation in Section 7.2 abo   </text>
<query_num> 8105 </query_num>
<text>   er(s). Therefore, we can treat the layer’s functionality as sub-capability of the higher layer’s capability. This provide a modularised design that meets the requirements of agent-oriented programming=-=[43, 91]-=-. It also facilitates the possibility of plug-in compatible capabilities in order to incorporate specific revision techniques. In JADE, each functionality/service provided by an agent should be implem   </text>
<query_num> 8106 </query_num>
<text>   is understood to be the property that agents can achieve designed and delegated goals without human 5sintervention[113]. A further interpretation is that agents may be developed by different vendors =-=[27]-=-[92] under different standards, using different platforms, technologies and toolkits, bearing different assumptions. A good example that illustrates the diversity of agent technologies, Agenticities 3   </text>
<query_num> 8107 </query_num>
<text>   known, in this or similar topics. 2. The agent is known, but not in this or similar topics. 3. The agent is known and trusted in this or similar topics. Inspired by Jonker and Trenur’s trust dynamics=-=[64]-=-, the above three situations can be handled respectively as shown below: 116s1. In the first case, without previous trust influencing experiences but with a given competency or sincerity threshold, cδ rity) using a competency (sincerity) evaluation or update function. Such functions can be defined to suit various applications by following the properties and constraints proposed by Jonker and Trenur=-=[64]-=-. Although such functions are applicable for both competency and sincerity evaluation, one may find that an agent’s competency value is unlikely to change as dramatically as the sincerity value does.   </text>
<query_num> 8108 </query_num>
<text>   le to Web applications which are highly heterogenous and autonomous in the sense that no single system has global control or holds all the global data. The first generation of web agents were shopbots=-=[67]-=- that compare prices (e.g. Bargain Finder from Anderson Consulting[71]) and provide recommendations etc. It is also predicated that with the help of research in agent research communities, in three or   </text>
<query_num> 8109 </query_num>
<text>   nition of independence - namely, that A and B are independent if P rA ∩ B = P r(A) × P r(B) - is a consequence of the intuitive notion, but not considered as sufficient or complete definition over it.=-=[53]-=- 57sIf there is no such pair B1, B2 that meets the combination condition, then m1 ⊕m2 is left undefined, and the corresponding belief functions Bel1 and Bel2 are said to be not combinable. Using the r new information as a generalized probability (c.f. an evidence). Halpern proposed another version of conditional belief function. For those who are interested, please refer to Theorem 3.4 in his paper=-=[53]-=-. 4.4 Possibility Theory Possibility theory[23] is a new form of information theory which is related to, but independent of, both fuzzy sets and probability theory. Technically, a possibility distribu oint but are also singletons, 64sa belief function turns out to be a special type of probability function, in which every element in the sample space is measurable. Therefore, we have, Theorem 4.5.1 (=-=[53]-=-). A belief function is a discrete probability function if only if its focal elements are disjoint singletons. 4.6 Possibility as Special Belief Functions We can show via the following proof 6 that th   </text>
<query_num> 8110 </query_num>
<text>   process is carried out in a multi-agent environment, where the new information may come from multiple sources and may be in conflict. Belief revision in this sense is called MSBR by Dragoni et al.[18]=-=[22]-=-[19]. Cantwell[11] tries to resolve conflicting information by ordering the information sources on the basis of their trustworthiness. This could be viewed as a rational way of generating the new info   </text>
<query_num> 8111 </query_num>
<text>   roperties (i.e. attributes), as compared to KIF and CycL, whose central primitives are predicates. On the other hand, recently, there is much interest in ontology languages based on Description Logics=-=[9]-=-, such as CLASSIC and DAML. A distinguishing feature of Description Logics is that classes (usually called concepts) can be defined intentionally using descriptions. The descriptions are used to speci   </text>
<query_num> 8112 </query_num>
<text>   s in agent knowledge bases. In the early 1990’s, Fagin et al.[31] semantically defined mutual belief and common knowledge. Van der Meyden extended this modal logic approach into MABR in 1994. Malheiro=-=[81]-=- in the same year defined private and shared belief to model the belief revision process in a distributed truth maintenance system. Similarly, Kfir-dahav and 86sTennenholtz claimed in 1996 that their   </text>
<query_num> 8113 </query_num>
<text>   s of belief revision. Hence, inconsistency across the knowledge bases of the society is permitted. This is called inconsistency principle. This is different from the 141sLiberal Belief Revision Policy=-=[20]-=- where the final distributed belief revision goal is to achieve global consistency. Observation: If sentence p is inconsistently believed in the agent society, then neither p ∈ Kc nor ¬p ∈ Kc This obs   </text>
<query_num> 8114 </query_num>
<text>   t or representation of some part 94sof a conceptualisation[48] is usually called an ontology. In other words, ontologies are explicit specifications of the terms in the domain and relations among them=-=[46]-=-. In addition to conceptualisation, there are two other important features of an explicit ontology: • Vocabulary: this involves assigning symbols or terms to refer to those objects, attributes and rel re not panacea, they can be both constructive and destructive. Information distortion could occur during the generalisation process of ontology design. Many criteria exist to prescribe a good ontology=-=[46]-=- and more are still under investigation. Essentially, the key is to thoroughly understand the problem domain. To design an ontological knowledge structure and model trust, both philosophical and psych   </text>
<query_num> 8115 </query_num>
<text>   technologies gain maturity, it becomes widely accepted that agent-oriented approaches represent the mainstream software engineering, particularly for tackling issues presented in large complex systems=-=[63]-=-. Agent-oriented software engineering has gained tremendous attention in recent times. Its practical and theoretical elements are proving to be helpful in understanding and modelling complex systems,  liefs. 3. ...” — Michael Wooldridge (1999) in [113] 5.1 Brief Introduction to Multi-Agent Systems Agent technologies and multi-agent systems have been identified as the mainstream software engineering=-=[63]-=- for large distributed domains, such as ecommerce/ebusiness, distributed sensor networks, air traffic control[77], resource allocation for Internet access[80] etc. The major focus of a multi-agent sys   </text>
</top>
<top>
<paper_num> 82 </paper_num>
<paper_title>   Capturing Anatomical Shape Variability Using B-Spline Registration.  </paper_title>
<paper_abstract>   Abstract. Registration based on B-spline transformations has attracted much attention in medical image processing recently. Non-rigid registration provides the basis for many important techniques, such as statistical shape modeling. Validating the results, however, remains difficult- especially in intersubject registration. This work explores the ability of Bspline registration methods to capture intersubject shape deformations. We study the effect of different established and new shape representations, similarity measures and optimization strategies on the matching quality. To this end we conduct experiments on synthetic shapes representing deformations which typically may arise in intersubject registration, as well as on real patient data of the liver and pelvic bone. The experiments clearly reveal the influence of each component on the registration performance. The results may serve as a guideline for assessing intensity based registration. 1  </paper_abstract>
<query_num> 8201 </query_num>
<text>   averaged intensity image obtained by their registration. Frangi et al. [8] consider landmark correspondence. Instead of surfaces, lower dimensional structures such as landmarks [13] or feature curves =-=[14]-=- are in use. Unfortunately, for many organs like the liver such descriptions are difficult to derive due to a lack of characteristic shape features.sIncorporation of geometric features into ICP can be   </text>
<query_num> 8202 </query_num>
<text>   e. Unfortunately, for many organs like the liver such descriptions are difficult to derive due to a lack of characteristic shape features.sIncorporation of geometric features into ICP can be found in =-=[15, 16]-=-. Wang et al. [17] base their semi-automatic matching on curvature classifiers. A fundamentally different approach to non-rigid matching is based on mappings of two-dimensional manifolds [18–20], as o   </text>
<query_num> 8203 </query_num>
<text>   investigation. Previous Work. Fleute et al. [6] first used intersubject non-rigid registration for building a statistical shape model of the knee. They employed the algorithm by Szeliski and Lavallée =-=[1]-=- using asymmetric surface distance as similarity measure. Frangi et al. applied Rückert’s registration [2] based on label fields for CT bone, MRI brain data [7], and cardiac images [8]. They compare t d to irregular B-spline deformations. Therefore, in some applications we use a regularization term R in the cost function, which models the bending energy of a thin metal plate (biharmonic model, see =-=[1, 2]-=-). Boundary Constraint (Landmarks). We encountered situations in intersubject registration where all of the above similarity measures with or without regularization fail to achieve a reasonable regist   </text>
<query_num> 8204 </query_num>
<text>   ion Taffine as well as a B-spline deformation TB−spline. The latter is defined on a 3D discrete uniform control point grid (CPG) with cubic B-spline interpolation between adjacent control points, see =-=[2, 4]-=- for details. The B-spline deformation model appears suitable for intersubject registration, because it provides smooth deformations when a physical model is not known. The optimal transformation T fo   </text>
<query_num> 8205 </query_num>
<text>   many organs like the liver such descriptions are difficult to derive due to a lack of characteristic shape features.sIncorporation of geometric features into ICP can be found in [15, 16]. Wang et al. =-=[17]-=- base their semi-automatic matching on curvature classifiers. A fundamentally different approach to non-rigid matching is based on mappings of two-dimensional manifolds [18–20], as opposed to volumetr   </text>
<query_num> 8206 </query_num>
<text>   method to the work of Brett et al. [9], who use a symmetric variant of the rigid iterative closest-points algorithm (ICP) [10] for brain data. Non-rigid extensions to ICP have been reported recently =-=[11, 12]-=-. Rohlfing et al. [3] employed the algorithm by Rückert for construction of an anatomical atlas of the honey bee. The capability of the deformation model has not been analyzed thoroughly up to now. It   </text>
<query_num> 8207 </query_num>
<text>   milarity measure. Frangi et al. applied Rückert’s registration [2] based on label fields for CT bone, MRI brain data [7], and cardiac images [8]. They compare their method to the work of Brett et al. =-=[9]-=-, who use a symmetric variant of the rigid iterative closest-points algorithm (ICP) [10] for brain data. Non-rigid extensions to ICP have been reported recently [11, 12]. Rohlfing et al. [3] employed   </text>
<query_num> 8208 </query_num>
<text>   pness, entropy) of the averaged intensity image obtained by their registration. Frangi et al. [8] consider landmark correspondence. Instead of surfaces, lower dimensional structures such as landmarks =-=[13]-=- or feature curves [14] are in use. Unfortunately, for many organs like the liver such descriptions are difficult to derive due to a lack of characteristic shape features.sIncorporation of geometric f   </text>
</top>
<top>
<paper_num> 83 </paper_num>
<paper_title>   Storing Multidimensional XML Documents in Relational Databases.  </paper_title>
<paper_abstract>   Abstract. The problem of storing and querying XML data using relational databases has been considered a lot and many techniques have been developed. MXML is an extension of XML suitable for representing data that assume different facets, having different value and structure under different contexts, which are determined by assigning values to a number of dimensions. In this paper, we explore techniques for storing MXML documents in relational databases, based on techniques previously proposed for conventional XML documents. Essential characteristics of the proposed techniques are the capabilities a) to reconstruct the original MXML document from its relational representation and b) to express MXML context-aware queries in SQL. 1  </paper_abstract>
<query_num> 8301 </query_num>
<text>   XML data in relational databases Many researchers have investigated how an RDBMS can be used to store and query XML data. Work has also been directed towards the storage of temporal extensions of XML =-=[15, 1, 2]-=-. The techniques proposed for XML storage can be divided in two categories, depending on the presence or absence of a schema: 1. Schema-Based XML Storage techniques: the objective here is to find a re   </text>
<query_num> 8302 </query_num>
<text>   ce or absence of a schema: 1. Schema-Based XML Storage techniques: the objective here is to find a relational schema for storing an XML document, guided by the structure of a schema for that document =-=[9, 13, 5, 14, 10, 3, 11]-=-. 2. Schema-Oblivious XML Storage techniques: the objective is to find a relational schema for storing XML documents independent of the presence or absence of a schema [13, 5, 14, 16, 10, 6, 4]. The a   </text>
<query_num> 8303 </query_num>
<text>   cument [9, 13, 5, 14, 10, 3, 11]. 2. Schema-Oblivious XML Storage techniques: the objective is to find a relational schema for storing XML documents independent of the presence or absence of a schema =-=[13, 5, 14, 16, 10, 6, 4]-=-. The approaches that we propose in this paper do not take schema information into account, and therefore belong to the Schema-Oblivious category. 3 Properties of MXML documents 3.1 A graphical model   </text>
<query_num> 8304 </query_num>
<text>   cument from its relational representation and b) to express MXML context-aware queries in SQL. 1 Introduction The problem of storing XML data in relational databases has been intensively investigated =-=[4, 10, 11, 13]-=- during the past 10 years. The objective is to use an RDBMS in order to store and query XML data. First, a relational schema is chosen for storing the XML data, and then XML queries, produced by appli ce or absence of a schema: 1. Schema-Based XML Storage techniques: the objective here is to find a relational schema for storing an XML document, guided by the structure of a schema for that document =-=[9, 13, 5, 14, 10, 3, 11]-=-. 2. Schema-Oblivious XML Storage techniques: the objective is to find a relational schema for storing XML documents independent of the presence or absence of a schema [13, 5, 14, 16, 10, 6, 4]. The a   </text>
<query_num> 8305 </query_num>
<text>   cument from its relational representation and b) to express MXML context-aware queries in SQL. 1 Introduction The problem of storing XML data in relational databases has been intensively investigated =-=[4, 10, 11, 13]-=- during the past 10 years. The objective is to use an RDBMS in order to store and query XML data. First, a relational schema is chosen for storing the XML data, and then XML queries, produced by appli cument [9, 13, 5, 14, 10, 3, 11]. 2. Schema-Oblivious XML Storage techniques: the objective is to find a relational schema for storing XML documents independent of the presence or absence of a schema =-=[13, 5, 14, 16, 10, 6, 4]-=-. The approaches that we propose in this paper do not take schema information into account, and therefore belong to the Schema-Oblivious category. 3 Properties of MXML documents 3.1 A graphical model   </text>
<query_num> 8306 </query_num>
<text>   imensional XML In MXML, data assume different facets, having different value or structure, under different contexts according to a number of dimensions which may be applied to elements and attributes =-=[7, 8]-=-. The notion of “world” is fundamental in MXML. A world represents an environment under which data obtain a meaning. A world is determined by assigning to every dimension a single value, taken from th   </text>
<query_num> 8307 </query_num>
<text>   ly, the inherited context ic(q) of a node q is defined as ic(q) = ic(p) ∩ c ec(q), where ic(p) is the inherited context of its parent node p. ∩ c is an operator called context intersection defined in =-=[12]-=- which combines two context specifiers and computes a new context specifier which represents the intersection of the worlds specified by the original context specifiers. The evaluation of the inherite  a leaf node then icc(n) = ic(n); otherwise icc(n) = icc(n1) ∪ c icc(n2) ∪ c ... ∪ c icc(nk), where n1, . . . , nk are the child element nodes of n. ∪ c is an operator called context union defined in =-=[12]-=- which combines two context specifiers and computes a new one which represents the union of the worlds specified by the original context specifiers. The inherited context coverage gives the true conte   </text>
</top>
<top>
<paper_num> 84 </paper_num>
<paper_title>   Large Scale Distributed Deep Networks.  </paper_title>
<paper_abstract>   Recent work in unsupervised feature learning and deep learning has shown that being able to train large models can dramatically improve performance. In this paper, we consider the problem of training a deep network with billions of parameters using tens of thousands of CPU cores. We have developed a software framework called DistBelief that can utilize computing clusters with thousands of machines to train large models. Within this framework, we have developed two algorithms for large-scale distributed training: (i) Downpour SGD, an asynchronous stochastic gradient descent procedure supporting a large number of model replicas, and (ii) Sandblaster, a framework that supports a variety of distributed batch optimization procedures, including a distributed implementation of L-BFGS. Downpour SGD and Sandblaster L-BFGS both increase the scale and speed of deep network training. We have successfully used our system to train a deep network 30x larger than previously reported in the literature, and achieves state-of-the-art performance on ImageNet, a visual object recognition task with 16 million images and 21k categories. We show that these same techniques dramatically accelerate the training of a more modestly- sized deep network for a commercial speech recognition service. Although we focus on and report performance of these methods as applied to training large neural networks, the underlying algorithms are applicable to any gradient-based machine learning algorithm. 1  </paper_abstract>
<query_num> 8401 </query_num>
<text>   cademic machine learning data sets have grown at an unprecedented pace. In response, a great many authors have explored scaling up machine learning algorithms through parallelization and distribution =-=[11, 12, 13, 14, 15, 16, 17]-=-. Much of this research has focused on linear, convex models, where distributed gradient computation is the natural first step. Within this area, some groups have relaxed synchronization requirements,   </text>
<query_num> 8402 </query_num>
<text>   coordinates of the gradient vector are non-zero for any given training example) have explored lock-less asynchronous stochastic gradient descent on shared-memory architectures (i.e. single machines) =-=[5, 18]-=-. We are interested in an approach that captures the best of both worlds, allowing the use of a cluster of machines asynchronously computing gradients, but without requiring that the problem be either   </text>
<query_num> 8403 </query_num>
<text>   e combined with clever distributed optimization techniques that leverage data parallelism. We considered a number of existing large-scale computational tools for application to our problem, MapReduce =-=[23]-=- and GraphLab [24] being notable examples. We concluded that MapReduce, designed for parallel data processing, was ill-suited for the iterative computations inherent in deep network training; whereas  r schedules multiple copies of the outstanding portions and uses the result from whichever model replica finishes first. This scheme is similar to the use of “backup tasks” in the MapReduce framework =-=[23]-=-. Prefetching of data, along with supporting data affinity by assigning sequential 5portions of data to the same worker makes data access a non-issue. In contrast with Downpour SGD, which requires re   </text>
<query_num> 8404 </query_num>
<text>   ever distributed optimization techniques that leverage data parallelism. We considered a number of existing large-scale computational tools for application to our problem, MapReduce [23] and GraphLab =-=[24]-=- being notable examples. We concluded that MapReduce, designed for parallel data processing, was ill-suited for the iterative computations inherent in deep network training; whereas GraphLab, designed   </text>
<query_num> 8405 </query_num>
<text>   ion accuracy [3, 4, 7]. These results have led to a surge of interest in scaling up the training and inference algorithms used for these models [8] and in improving applicable optimization procedures =-=[7, 9]-=-. The use of GPUs [1, 2, 3, 8] is a significant advance in recent years that makes the training of modestly sized deep networks practical. A known limitation of the GPU approach is that the training s   </text>
<query_num> 8406 </query_num>
<text>   learning and unsupervised feature learning have shown great promise in many practical applications. State-of-the-art performance has been reported in several domains, ranging from speech recognition =-=[1, 2]-=-, visual object recognition [3, 4], to text processing [5, 6]. It has also been observed that increasing the scale of deep learning, with respect to the number of training examples, the number of mode These results have led to a surge of interest in scaling up the training and inference algorithms used for these models [8] and in improving applicable optimization procedures [7, 9]. The use of GPUs =-=[1, 2, 3, 8]-=- is a significant advance in recent years that makes the training of modestly sized deep networks practical. A known limitation of the GPU approach is that the training speed-up is small when the mode  more parameters benefit more from the use of additional machines than do models with fewer parameters. 625 Accuracy on Training Set 25 Accuracy on Test Set Average Frame Accuracy (%) 20 15 10 5 SGD =-=[1]-=- DownpourSGD [20] DownpourSGD [200] w/Adagrad Sandblaster L−BFGS [2000] 0 0 20 40 60 80 100 120 Time (hours) Average Frame Accuracy (%) 20 15 10 SGD [1] GPU [1] 5 DownpourSGD [20] DownpourSGD [20] w/A   </text>
<query_num> 8407 </query_num>
<text>   learning have shown great promise in many practical applications. State-of-the-art performance has been reported in several domains, ranging from speech recognition [1, 2], visual object recognition =-=[3, 4]-=-, to text processing [5, 6]. It has also been observed that increasing the scale of deep learning, with respect to the number of training examples, the number of model parameters, or both, can drastic   </text>
<query_num> 8408 </query_num>
<text>   nts varying the number of identically connected nodes from 8 to 36. The output layer consisted of 21 thousand one-vs-all logistic classifier nodes, one for each of the ImageNet object categories. See =-=[29]-=- for similar deep network configurations and training procedures. Model parallelism benchmarks: To explore the scaling behavior of DistBelief model parallelism (Section 3), we measured the mean time t xploring the capabilities of very large neural networks, we used Downpour SGD to train the 1.7 billion parameter image model described above on the ImageNet object classification task. As detailed in =-=[29]-=-, this network achieved a cross-validated classification accuracy of over 15%, a relative improvement over 60% from the best performance we are aware of on the 21k category ImageNet classification tas   </text>
<query_num> 8409 </query_num>
<text>   observed that increasing the scale of deep learning, with respect to the number of training examples, the number of model parameters, or both, can drastically improve ultimate classification accuracy =-=[3, 4, 7]-=-. These results have led to a surge of interest in scaling up the training and inference algorithms used for these models [8] and in improving applicable optimization procedures [7, 9]. The use of GPU liminated stability concerns in training deep networks using Downpour SGD (see results in Section 5). 4.2 Sandblaster L-BFGS Batch methods have been shown to work well in training small deep networks =-=[7]-=-. To apply these methods to large models and large datasets, we introduce the Sandblaster batch optimization framework and discuss an implementation of L-BFGS using this framework. A key idea in Sandb   </text>
<query_num> 8410 </query_num>
<text>   promise in many practical applications. State-of-the-art performance has been reported in several domains, ranging from speech recognition [1, 2], visual object recognition [3, 4], to text processing =-=[5, 6]-=-. It has also been observed that increasing the scale of deep learning, with respect to the number of training examples, the number of model parameters, or both, can drastically improve ultimate class   </text>
<query_num> 8411 </query_num>
<text>   promise in many practical applications. State-of-the-art performance has been reported in several domains, ranging from speech recognition [1, 2], visual object recognition [3, 4], to text processing =-=[5, 6]-=-. It has also been observed that increasing the scale of deep learning, with respect to the number of training examples, the number of model parameters, or both, can drastically improve ultimate class  coordinates of the gradient vector are non-zero for any given training example) have explored lock-less asynchronous stochastic gradient descent on shared-memory architectures (i.e. single machines) =-=[5, 18]-=-. We are interested in an approach that captures the best of both worlds, allowing the use of a cluster of machines asynchronously computing gradients, but without requiring that the problem be either  form of the model. In special cases where one layer dominates computation, some authors have considered distributing computation in that one layer and replicating computation in the remaining layers =-=[5]-=-. But in the general case where many layers of the model are computationally intensive, full model parallelism in a spirit similar to [22] is required. To be successful, however, we believe that model   </text>
<query_num> 8412 </query_num>
<text>   results about large-scale nonconvex optimization. Firstly, asynchronous SGD, rarely applied to nonconvex problems, works very well for training deep networks, particularly when combined with Adagrad =-=[10]-=- adaptive learning rates. Secondly, we show that given sufficient resources, L-BFGS is competitive with or faster than many variants of SGD. With regard to specific applications in deep learning, we r s, but in practice we found relaxing consistency requirements to be remarkably effective. One technique that we have found to greatly increase the robustness of Downpour SGD is the use of the Adagrad =-=[10]-=- adaptive learning rate procedure. Rather than using a single fixed learning rate on the parameter sever (η in Figure 2), Adagrad uses a separate adaptive learning rate for each parameter. Let ηi,K be   </text>
<query_num> 8413 </query_num>
<text>   ures. For visual object recognition we trained a larger neural network with locally-connected receptive fields on the ImageNet data set of 16 million images, each of which we scaled to 100x100 pixels =-=[28]-=-. The network had three stages, each composed of filtering, pooling and local contrast normalization, where each node in the filtering layer was connected to a 10x10 patch in the layer below. Our infr   </text>
</top>
<top>
<paper_num> 85 </paper_num>
<paper_title>   Revisiting the visit: : understanding how technology can shape the museum visit.  </paper_title>
<paper_abstract>   This paper reports findings from a study of how a guidebook was used by pairs of visitors touring a historic house. We describe how the guidebook was incorporated into their visit in four ways: shared listening, independent use, following one another, and checking in on each other. We discuss how individual and groupware features were adopted in support of different visiting experiences, and illustrate how that adoption was influenced by social relationships, the nature of the current visit, and any museum visiting strategies that the couples had. Finally, we describe how the guidebook facilitated awareness between couples, and how awareness of non-guidebook users (strangers) influenced use.  </paper_abstract>
<query_num> 8501 </query_num>
<text>   While previous studies have often focused on the awareness colleagues need in order to collaborate, some studies have observed that people are aware of others around them. For example, Isaacs et al. =-=[19]-=- report that office colleagues used IM when they wanted to talk without disturbing others. Grinter and Eldridge [13] find that teenagers sometimes use SMS to avoid disturbing other family members. Sim   </text>
<query_num> 8502 </query_num>
<text>   ation. Two couples described longer checking in activities. These resemble reports of monitoring activities, such as those described by Heath and Luff in their study of London Underground controllers =-=[16]-=-. One couple, P3 described checking in on his companion because he did not want to lose the awareness of where she was while he went and did something else. In this case, one member of P3 wanted to ta   </text>
<query_num> 8503 </query_num>
<text>   ects have audio descriptions associated with them; when a tap fails to hit a target, the guidebook briefly highlights all of the available targets (Figure 1, lower left), a technique we call tap tips =-=[5]-=-. Eavesdropping Sotto Voce supports synchronized sharing of descriptive audio content between pairs of visitors. The audio content is presented through headsets to reduce the impact on others in the e   </text>
<query_num> 8504 </query_num>
<text>   f Eleanor Roosevelt’s home using reel-to-reel tape players [3]. They continue to be the subject of research and development, with recent work typically focused on the use of portable computers (e.g., =-=[1, 9]-=-). In spite of this long history, electronic guidebooks still raise many design challenges. One design challenge for audio-based guidebooks is the physical delivery of content. Use of open speakers is s such, a number of research projects have been carried out. Many of these have focused on exploiting sensing and location awareness technologies, to explore ubiquitous computing concerns (see, e.g., =-=[1, 9]-=-). Sotto Voce differs from these, and other commercial efforts, in its focus on supporting interaction among visitors. THE SYSTEM: SOTTO VOCE The design of Sotto Voce was primarily motivated by the co   </text>
<query_num> 8505 </query_num>
<text>   f Eleanor Roosevelt’s home using reel-to-reel tape players [3]. They continue to be the subject of research and development, with recent work typically focused on the use of portable computers (e.g., =-=[1, 9]-=-). In spite of this long history, electronic guidebooks still raise many design challenges. One design challenge for audio-based guidebooks is the physical delivery of content. Use of open speakers is s such, a number of research projects have been carried out. Many of these have focused on exploiting sensing and location awareness technologies, to explore ubiquitous computing concerns (see, e.g., =-=[1, 9]-=-). Sotto Voce differs from these, and other commercial efforts, in its focus on supporting interaction among visitors. THE SYSTEM: SOTTO VOCE The design of Sotto Voce was primarily motivated by the co nd to find us when they had completed the three rooms contained in the guidebook (in the sense that they returned to their initial activity, this study is similar to that conducted by Cheverst et al. =-=[9]-=-). During their exploration of two of the three rooms, we used video cameras to record their interactions with each other, the guidebook, and (on some occasions) with other visitors and pairs of guide   </text>
<query_num> 8506 </query_num>
<text>   g us to make the technology as intuitive as possible to adopt. Additional issues arise when the system is compared to other collaborative technologies for public settings. For example, Benford et al. =-=[7]-=- have designed a number of technologies to support Inhabited TV, a blending of broadcast television with Collaborative Virtual Environments (CVEs). In one of their experiments (“The Mirror”), the tele   </text>
<query_num> 8507 </query_num>
<text>   ile remaining relatively co-located. Sotto Voce is also a mobile technology. Mobility and the collaborative use of mobile technologies have also been a focus of recent CSCW interest (see for example, =-=[6, 13, 17, 21, 25]-=-). However, Sotto Voce differs from these in two ways. First, Sotto Voce is limited in use to a particular setting, the museum. Findings presented show that the setting shaped how the technology was a   </text>
<query_num> 8508 </query_num>
<text>   ile remaining relatively co-located. Sotto Voce is also a mobile technology. Mobility and the collaborative use of mobile technologies have also been a focus of recent CSCW interest (see for example, =-=[6, 13, 17, 21, 25]-=-). However, Sotto Voce differs from these in two ways. First, Sotto Voce is limited in use to a particular setting, the museum. Findings presented show that the setting shaped how the technology was a range of the audience but not necessarily engaged in talk with them” ([12], p107). Strategies for technology adoption based on awareness of potential strangers have also been reported by Palen et al. =-=[21]-=-. Specifically, their study finds that people who had just purchased a mobile phone were sensitive to using it in public settings because they did not want to disturb strangers. However, the authors r   </text>
<query_num> 8509 </query_num>
<text>   ile remaining relatively co-located. Sotto Voce is also a mobile technology. Mobility and the collaborative use of mobile technologies have also been a focus of recent CSCW interest (see for example, =-=[6, 13, 17, 21, 25]-=-). However, Sotto Voce differs from these in two ways. First, Sotto Voce is limited in use to a particular setting, the museum. Findings presented show that the setting shaped how the technology was a ve observed that people are aware of others around them. For example, Isaacs et al. [19] report that office colleagues used IM when they wanted to talk without disturbing others. Grinter and Eldridge =-=[13]-=- find that teenagers sometimes use SMS to avoid disturbing other family members. Similar to these findings, some of our study participants reported not wanting to disturb others in the museum, and tha   </text>
<query_num> 8510 </query_num>
<text>   mechanism that enabled visitors wearing headphones to be able to share audio content in a similar (but not identical) way. We evaluated this mechanism, which we call eavesdropping, in a second study =-=[4]-=-. This led to the third study — the subject of this paper. In the first two studies we invited people to try the technology in a historic home setting during a closed day when they were alone in the h seum interactions were considerably enriched by additional resources such as the visitors’ control over their content playback and the objects themselves – including those without descriptive content =-=[4]-=-.) One other feature of Inhabited TV has a salience for Sotto Voce. In their discussion of challenges for Inhabited TV, they observe that in their first experiment, a public poetry performance, the po s. In the rest of this section, we describe our system prototype, focusing on the features that are relevant for this paper. The system and its design rationale are presented in more detail elsewhere =-=[4]-=-. Guidebook User Interface Sotto Voce is implemented on a handheld computer with a color touchscreen display – the current prototype uses the Compaq iPAQ™ 3650. Each visitor obtains information about  stening can facilitate conversations. The guidebook enables shared listeners to engage in mutual conversation by giving them activity cues – information about what and when to talk to their companion =-=[4, 26]-=-. This study reinforces these previous findings and extends our understanding of visitors’ social motivations to engage insshared listening activity. For example, P13 described shared listening as “th   </text>
<query_num> 8511 </query_num>
<text>   project has been through three design-evaluation iterations. The first evaluation established that visitors, without prompting, integrated audio clips played through open air into their interactions =-=[26]-=-. Based on these findings, we designed a mechanism that enabled visitors wearing headphones to be able to share audio content in a similar (but not identical) way. We evaluated this mechanism, which w vations of museum visitors, and our experience with a previous electronic guidebook, we knew that many visitors wish to be able to listen to descriptive information together and share their reactions =-=[26]-=-. However, at the same time, visitors generally want control over their experience. Consequently, the design strikes a balance between these (as well as other) concerns. In the rest of this section, w ree rooms have audio descriptions. The length and structure of the audio descriptions are specifically designed to provide frequent and natural opportunities for visitors to take conversational turns =-=[26]-=-. Visitors wear modified telephone headsets with a single, over-ear earphone (Figure 1, right). We chose this configuration to maximize ease of use, extended-wear comfort, and the ability to converse  stening can facilitate conversations. The guidebook enables shared listeners to engage in mutual conversation by giving them activity cues – information about what and when to talk to their companion =-=[4, 26]-=-. This study reinforces these previous findings and extends our understanding of visitors’ social motivations to engage insshared listening activity. For example, P13 described shared listening as “th ch supports an earlier finding that whether or not visitors were engaged in shared listening activity, their talk was organized around their listening activity with the guidebook’s audio descriptions =-=[26]-=-. In fact, some visitors expressed a feeling of becoming absorbed into the listening activity. As P19 described “I wasn’t asking her too much, I was trying to push my buttons as fast as possible ... I   </text>
</top>
<top>
<paper_num> 86 </paper_num>
<paper_title>   Why Real-World Multimedia Assets Fail to Enter the Semantic Web.  </paper_title>
<paper_abstract>   Making multimedia assets on the one hand first-class objects on the Semantic Web, while keeping them on the other hand conforming to existing multimedia standards is a non-trivial task. Most proprietary media asset formats are binary, optimized for streaming or storage. However, the semantics carried by the media assets are not accessible directly. In addition, multimedia description standards lack the expressiveness to gain a semantic understanding of the media assets. There exists an array of requirements regarding media assets and the Semantic Web, already. Based on a critical review of these requirements we investigate how ontology languages fit into the picture. We finally analyse the usefulness of formal accounts to describe spatio-temporal aspects of multimedia assets in a practical context.  </paper_abstract>
<query_num> 8601 </query_num>
<text>   OWL and WSML can be perceived comparable. 5.3.5 Trust and IPR In an interdependent, interconnected environment as the Semantic Web, two important aspects immediately arise: data provenance and trust =-=[4]-=-. Requirements regarding trust issues gathered from [30, 16] contain costs and benefits w.r.t. implementation, technology-driven vs. social networking, etc. Both WSML and OWL do not have explicit prov   </text>
<query_num> 8602 </query_num>
<text>   adata features of multimedia standards not enough? 2. REQUIREMENTS FOR THE DESCRIPTION OF MULTIMEDIA ASSETS Requirements for multimedia content descriptions have been researched in a number of papers =-=[15, 39, 29, 5]-=- before and investigations of the combination of multimedia descriptions with features of the Semantic Web are yet numerous [22, 3, 36, 37, 2]. In the following, we give a summarisation of the propose   </text>
<query_num> 8603 </query_num>
<text>   and IPR In an interdependent, interconnected environment as the Semantic Web, two important aspects immediately arise: data provenance and trust [4]. Requirements regarding trust issues gathered from =-=[30, 16]-=- contain costs and benefits w.r.t. implementation, technology-driven vs. social networking, etc. Both WSML and OWL do not have explicit provisions for handling trust and IPR issues, respectively. Howe   </text>
<query_num> 8604 </query_num>
<text>   forgotten when summarizing the requirements for a metadata schema. For the metadata creator it should be clear beforehand for what purpose the metadata will be used and what benefits he gains from it =-=[28]-=-, ie., using this part of the metadata scheme enhances retrieval, raises social attention or helps you protect your assets. This in turn also applies to the consumer of the metadata, functional descri   </text>
<query_num> 8605 </query_num>
<text>   les. • Libraries &amp; Applications. When developing applications, the availability of APIs is a core requirement. In special for Semantic Web applications, interface and mapping issues are of importance =-=[17]-=-. • Deployment Multimedia containers as HTML, SMIL, etc. require the metadata either being referenced from within the media assets, or being embedded into it. As the data model needs to be RDF—in cont   </text>
<query_num> 8606 </query_num>
<text>   s for modeling and description of ontologies but also functional (service) descriptions, i.e. the description of a service capability by means of precondition, assumptions, postconditions and effects =-=[24]-=-. OWL does not have support for such kind of descriptions. 5.3.7 Engineering Support Tool Support for WSML and especially OWL is constantly growing. However, the amount of tools available for OWL [41]   </text>
<query_num> 8607 </query_num>
<text>   ser-defined datatypes and restrictions involving datatype predicates, and a weak form of meta-modelling known as punning. The usage of rules in combination with DL has been investigated for some time =-=[12, 19]-=-—in the Semantic Web stack, it is expected that a rule language will complement the ontology layer. 11 http://www.w3.org/TR/rdfprimer/ 12 http://owl1_1.cs.manchester.ac.uk/owl_ specification.html5.1.   </text>
<query_num> 8608 </query_num>
<text>   tack of Semantic Web languages and technologies provided by the W3C is well suited to the formal, semantic descriptions of the terms in a multimedia document’s annotation. But, as also pointed out in =-=[35]-=-, the Semantic Web based languages lack the structural advantages of the XML-based approaches. Additionally, there is a huge amount of work already done on multimedia document annotation within the fr   </text>
<query_num> 8609 </query_num>
<text>   tent descriptions have been researched in a number of papers [15, 39, 29, 5] before and investigations of the combination of multimedia descriptions with features of the Semantic Web are yet numerous =-=[22, 3, 36, 37, 2]-=-. In the following, we give a summarisation of the proposed requirements and add two additional ones (Authoring &amp; Consumption and Performance &amp; Scalability). Representational Issues. A basic prerequis   </text>
</top>
<top>
<paper_num> 87 </paper_num>
<paper_title>   Detecting application-level failures in component-based Internet services.  </paper_title>
<paper_abstract>   Pinpoint is an application-generic framework for using statistical learning techniques to detect and localize likely application-level failures in component-based Internet services. Assuming that most of the system is working most of the time, Pinpoint looks for anomalies in low-level behaviors that are likely to reflect high-level application faults, and correlates these anomalies to their potential causes within the system. In our experiments, Pinpoint correctly detected and localized over 70-88% of the faults, depending on the type of fault, we injected into our testbed system, as compared to the 50-70% detected by current techniques. By demonstrating the applicability of statistical learning and providing an application-generic platform on which additional machine learning techniques can be applied to the problem of fast failure detection, we hope to hasten the adoption of statistical approaches to dependability for complex software systems.  </paper_abstract>
<query_num> 8701 </query_num>
<text>   (1) arises from the service being written using one of several standard component frameworks, such as .NET or J2EE, and from the threetier structure (Web servers, application logic, persistent store) =-=[25]-=- of many such services. (2) arises from the combination of using a component-based framework and HTTP’s requestreply nature. (3) arises because of the combination of large numbers of (presumably indep   </text>
<query_num> 8702 </query_num>
<text>   30] as a base, and can inject several different simple bugs, summarized in Table 3. While these are all minor bugs, evidence suggests that no bug is so trivial that it does not occur in real software =-=[22]-=-. For our experiments, we first generate an exhaustive list of the spots where a bug can be injected within a component, after eliminating locations in unused code. Then, we iterate over these “bug sp   </text>
<query_num> 8703 </query_num>
<text>   a tool for detecting “bad” behaviors in systems where many assumedgood behaviors can be observed, including intrusion detection [18, 36], Windows Registry debugging [37], finding bugs in system code =-=[19-=-], and detecting possible violation of runtime program invariants regarding variable assignment [21] or or assertions [26]. Although Ward et al. previously proposed anomaly detection as a way to ident   </text>
<query_num> 8704 </query_num>
<text>   ading faults. Pinpoint does not provide a detailed root-cause diagnosis, but rather localization of the fault within the system. Combined with a simple generic recovery mechanism such as microreboots =-=[8]-=-, simply knowing the location of a fault is often sufficient for fast recovery. Section 2 describes Pinpoint’s approach to detecting and localizing anomalies, and the two types of low-level behaviors   </text>
<query_num> 8705 </query_num>
<text>   alysis and component-interaction analysis, and evaluates when and how these fault detection algorithms work. Request paths have also been used for black-box monitoring for performance troubleshooting =-=[1]-=- and performance prediction and debugging [4]. In contrast to those efforts, we focus on the initial detection of failures and also focus on failures in application functionality rather than performan enview) and Tivoli (www.tivoli.com) typically rely on either expert systems with human-generated rules or on the use of dependency models to assist in fault localization [39, 17, 20]. Aguilera et al. =-=[1]-=- have used packetsniffing and statistical analysis to derive the call-graphs between black-boxes. In contrast to the direct tracing done by Pinpoint, this only produces a view of the majority behavior   </text>
<query_num> 8706 </query_num>
<text>   component behaviors, we localize these anomalies to their potential causes by looking for features of the paths and components that are highly correlated with the anomalies. We use the ID3 algorithm =-=[32]-=- for learning a decision tree, a data structure that represents a discrete-valued function, where each branch of the tree is a test of some attribute of the input, and where the leaves of the tree hol   </text>
<query_num> 8707 </query_num>
<text>   d simply omit it. If the function should have returned a value, we return 0 or a null value. Source code bug injection: Even simple programming bugs remain uncaught and cause problems in real software=-=[35, 33]-=-, so introducing them can be a useful method of simulating faults due to software bugs [15]. We do not inject low-level hardware or OS faults, such as CPU register bit-flips, memory corruptions and IO   </text>
<query_num> 8708 </query_num>
<text>   d.edu 1 has a serious effect on the overall reliability of Internet services: a study of three sites found that earlier detection might have mitigated or avoided 65% of reported user-visible failures =-=[31]-=-. We present Pinpoint, an application-generic framework for monitoring component-based Internet services, and detecting and localizing application-level failures without requiring a priori knowledge a t up-to-date as the application changes, otherwise the monitor will both miss real failures and cause false-alarms. For these reasons, neither the sites that we have spoken with, nor those studied in =-=[31]-=- make extensive use of these monitors. Thirdly, user-activity monitors watch simple statistics about the gross behavior of users and compare them to historical trends. Such a monitor might track the s vel behavior but do not cause obvious failures at lower-levels of the system stack. We surveyed studies of faults in deployed systems and the faults injected by other researchers in their experiments =-=[31, 23, 16]-=-. Java exceptions: Because Java coerces many different kinds of failures, from I/O errors to programmer errors, to manifest as exceptions, injecting exceptions is an appropriate method of testing an a   </text>
<query_num> 8709 </query_num>
<text>   er Pinpoint would be more likely to detect faults introduced by minor bugs in source code, we wrote a source-code bug injector for the Java language. We use the Polyglot extensible compiler framework =-=[30]-=- as a base, and can inject several different simple bugs, summarized in Table 3. While these are all minor bugs, evidence suggests that no bug is so trivial that it does not occur in real software [22   </text>
<query_num> 8710 </query_num>
<text>   g these failures is a significant problem: one large site estimates that about 93% of the time they spend recovering from application-level failures is spent detecting (75%) and diagnosing them (18%) =-=[12, 2]-=-. Other sites agreed that brown-outs can sometimes take days to detect, though they are usually repaired quickly once found. This situation Emre Kıcıman and Armando Fox {emrek, fox}@cs.stanford.edu 1  luster, we evaluate how well Pinpoint discovers failures and characterize Pinpoint’s strengths and weaknesses. We introduced path-analysis for localizing failures in Internet services in [14], and in =-=[13, 12]-=- broadened our pathanalysis techniques to show how this visibility into a system can aid fault management, fault impact analysis, and evolution management. The primary differentiator between Pinpoint  id HTML. cessful). In contrast Pinpoint assumes no prior knowledge of faults, and starts with fault detection. While we briefly discussed fault detection using path-shape analysis in Section 4.1.2 of =-=[12]-=-, this paper provides the details of the behavior modeling and anomaly detection algorithms for both pathshape analysis and component-interaction analysis, and evaluates when and how these fault detec ection. 3.1 PCFGs Model Path Shapes The shape of a path is the ordered set of logical software components (as opposed to instances of components on specific machines) used to service a client request =-=[12]-=-. We represent the shape of a path in a call-tree-like structure, except that each node in the tree is a component rather than a call site (i.e., calls that do not cross component boundaries are hidde  injected and secondary faults in realistic but small-scale test applications. The value of path-based analysis for failure management has been demonstrated in one production Internet service already =-=[12]-=-, and we are currently working to deploy with a large Internet service to apply Pinpoint monitoring to their systems. In addition, versions of Pinpoint have been integrated as fault monitors for two s   </text>
<query_num> 8711 </query_num>
<text>   identification of components. Anomaly detection has gained currency as a tool for detecting “bad” behaviors in systems where many assumedgood behaviors can be observed, including intrusion detection =-=[18, 36-=-], Windows Registry debugging [37], finding bugs in system code [19], and detecting possible violation of runtime program invariants regarding variable assignment [21] or or assertions [26]. Although   </text>
<query_num> 8712 </query_num>
<text>   lack-boxes. In contrast to the direct tracing done by Pinpoint, this only produces a view of the majority behavior of the system, and thus hides any anomalies existant within the system. Brown et al. =-=[6]-=- have also used dynamic observation to automatically build such dependency models. This approach can produce a rank-ordered list of potential causes, but they are intrusive and require a human to firs   </text>
<query_num> 8713 </query_num>
<text>   ncluding intrusion detection [18, 36], Windows Registry debugging [37], finding bugs in system code [19], and detecting possible violation of runtime program invariants regarding variable assignment [=-=21]-=- or or assertions [26]. Although Ward et al. previously proposed anomaly detection as a way to identify possible failures for Internet sites [38], they start with a statistical model based on 24 hours   </text>
<query_num> 8714 </query_num>
<text>   nged. 6 Discussion 6.1 The Base-Rate Fallacy The base-rate fallacy declares that, when looking for rare events, any non-zero false positive rate will overwhelm a de11 tector with even perfect recall. =-=[3]-=- argues that this makes most existing intrusion detection systems unusable. In the context of our broader project, Recovery Oriented Computing, we argue that false positives are only a problem when de   </text>
<query_num> 8715 </query_num>
<text>   nging. Sensor networks are a useful subclass of peer-to-peer systems whose applications are often data-analysis-centered by nature and in which dataaggregation machinery is already being put in place =-=[28]-=-, making sensor nets a potential appealing target as well. 9 Conclusions Pinpoint’s key insight is that aggregating low-level behavior over a large collection of requests, using it to establish a base   </text>
<query_num> 8716 </query_num>
<text>   our testbed cluster, we evaluate how well Pinpoint discovers failures and characterize Pinpoint’s strengths and weaknesses. We introduced path-analysis for localizing failures in Internet services in =-=[14]-=-, and in [13, 12] broadened our pathanalysis techniques to show how this visibility into a system can aid fault management, fault impact analysis, and evolution management. The primary differentiator   whether any feature-weighting additions to SVMs might improve the SVM analysis of path-shapes. As an alternative to decision-trees, we have used data clustering to correlate components with a failure=-=[14]-=-. While this worked reasonably well for single-component faults, we believe the decision-tree model can more naturally expresses localizations in the presence of multiple independent failures and late   </text>
<query_num> 8717 </query_num>
<text>   r Petstore 1.3, but its increased functionality makes it a useful test application. RUBiS is an auction website, developed at Rice University for experimenting with different design patterns for J2EE =-=[10]-=-. RUBiS contains over 500 Java files and over 25k lines of code. More relevant for our purposes, RUBiS has 6 EJBs and several servlets. While RUBiS comes with a custom load generator, we built our own   </text>
<query_num> 8718 </query_num>
<text>   time program invariants regarding variable assignment [21] or or assertions [26]. Although Ward et al. previously proposed anomaly detection as a way to identify possible failures for Internet sites [=-=38]-=-, they start with a statistical model based on 24 hours of observing the system, whereas Pinpoint builds and adjusts its model dynamically. 8 Future Directions Pinpoint detects injected and secondary   </text>
<query_num> 8719 </query_num>
<text>   urce code bug injection: Even simple programming bugs remain uncaught and cause problems in real software[35, 33], so introducing them can be a useful method of simulating faults due to software bugs =-=[15]-=-. We do not inject low-level hardware or OS faults, such as CPU register bit-flips, memory corruptions and IO errors because, empirically, these faults do not manifest as applicationlevel failures tha   </text>
<query_num> 8720 </query_num>
<text>   y detection has gained currency as a tool for detecting “bad” behaviors in systems where many assumedgood behaviors can be observed, including intrusion detection [18, 36], Windows Registry debugging =-=[37-=-], finding bugs in system code [19], and detecting possible violation of runtime program invariants regarding variable assignment [21] or or assertions [26]. Although Ward et al. previously proposed a   </text>
</top>
<top>
<paper_num> 88 </paper_num>
<paper_title>   Symbiosis in the Intranet: How Document Retrieval Benefits from Database Information.  </paper_title>
<paper_abstract>   The enterprise information space is split in two hemispheres. Documents contain unstructured or semistructured information; structured information is stored in databases. As regards the content, both kinds of information are complementary parts. However, enterprise information systems usually focus on one part, only. Our approach improves document retrieval in the intranet by exploiting the enterprise’s databases. In particular, we exploit database information to describe the context of documents and exploit this context to enhance common full text search. In this paper, we show how to model and compute document context and present results on runtime performance. 1  </paper_abstract>
<query_num> 8801 </query_num>
<text>   A somewhat similar approach enriches the result of SQL queries with documents retrieved by text search [18]. Recently, there has been considerable research in the area of XML retrieval systems, e.g. =-=[2, 8, 11, 20]-=-. None of them considers information from relational database systems. In general, they aim at exploiting both, the content and the logical structure of XML documents. In particular, they focus on the   </text>
<query_num> 8802 </query_num>
<text>   es on database exploration. The typical scenario comprises a user who is running keyword queries against a database with unknown schema. Goldman et al. propose the exploration of a Lorel/OEM database =-=[9]-=-. There, the user needs to specify two sets of objects: find and near objects. The system retrieves objects from the find-set according to their distance from objects in the near-set. In the relationa pe of this paper. Rocha [17] proposed to model all attributes of a tuple as one single graph node. However, we favor the smaller node granularity, as it has been proposed, e.g., in the OEM data model =-=[9]-=- for two reasons: First, it permits to assign different semantic distances (see below) to attributes. And secondly, a graph that contains atomic nodes only is more flexible regarding the integration o   </text>
<query_num> 8803 </query_num>
<text>   om the find-set according to their distance from objects in the near-set. In the relational database area, related approaches for keyword search support the user to find relationships in the database =-=[1,12,13]-=-. For a given set of search keywords, the systems return a set of joined tupels that contain the keyword, each. The BANKS system [3] is closely related to our approach since it represents a relational   </text>
<query_num> 8804 </query_num>
<text>   order of increasing weights. The restriction stops the algorithm when path weights exceed the cRange limit, i.e., when the context is identified. We additionally considered the Hidden Path algorithm =-=[14]-=- that does not benefit from the first optimization but computes the context of all nodes. It works as |V | individual Dijkstra algorithms in parallel, i.e., one for each graph node. Its running time b   </text>
<query_num> 8805 </query_num>
<text>   rtant consequences on our system architecture. E.g., the system can retrieve documents with and without context at the same time. There are ideas to integrate search engines with relational databases =-=[7,10]-=-. In contrast to our approach, they consider the results of the text search engine as FS a (virtual) table in the database system. The user interacts with the system by means of SQL queries. I.e., whe   </text>
<query_num> 8806 </query_num>
<text>   search support the user to find relationships in the database [1,12,13]. For a given set of search keywords, the systems return a set of joined tupels that contain the keyword, each. The BANKS system =-=[3]-=- is closely related to our approach since it represents a relational database as a graph where tuples and foreign key relationships are nodes and edges, respectively. For a given keyword query, the sy   </text>
</top>
<top>
<paper_num> 89 </paper_num>
<paper_title>   Query Processing and Optimization on the Web.  </paper_title>
<paper_abstract>   The advent of the Internet and the Web and their subsequent ubiquity have brought forth opportunities to connect information sources across all types of boundaries (local, regional, organizational, etc.). Examples of such information sources include databases, XML documents, and other unstructured sources. Uniformly querying those information sources has been extensively investigated. A major challenge relates to query optimization. Indeed, querying multiple information sources scattered on the Web raises several barriers for achieving efficiency. This is due to the characteristics of Web information sources that include volatility, heterogeneity, and autonomy. Those characteristics impede a straightforward application of classical query optimization techniques. They add new dimensions to the optimization problem such as the choice of objective function, selection of relevant information sources, limited query capabilities, and unpredictable events. In this paper, we survey the current research on fundamental problems to efficiently process queries over Web data integration systems. We also outline a classification for optimization techniques and a framework for evaluating them.  </paper_abstract>
<query_num> 8901 </query_num>
<text>   . Global schema integration was one of the first attempts at data sharing across HDDBS. It is based on the complete integration of multiple databases to provide a single view (global schema) [52]. In =-=[8]-=-, an exhaustive survey on schema integration is provided along with a comparison of several methodologies. In federated databases [29], a certain amount of autonomy for individual database systems is   </text>
<query_num> 8902 </query_num>
<text>   Most of the systems and techniques in this survey fall into that category. There are also other approaches based on the use of agents [41], ontology [2, 37, 43], and information retrieval techniques =-=[5, 33, 38]-=-. This section gives some details about the mediator approach. Then, it highlights major research issues related to query optimization for data integration on the Web. 3.1. Mediator based approaches A   </text>
<query_num> 8903 </query_num>
<text>   an is built based on a cost model using the previously gathered information. In the last step, the query execution plan is distributed to relevant cycle providers and executed using an iterator model =-=[26]-=-. Queries for the lookup service are extracted by the parser based on themes and attributes specified in the different clauses of a query. The lookup service uses the generated queries to locate relev form two actions: statistics gathering for the optimizer, and event handler invocation in case a significant event occurs. The operator tree execution follows the top-down iterator model described in =-=[26]-=-. For the second level of adaptiveness, two operators are used: dynamic collectors and the double pipelined hash join operator. The collector operator dynamically chooses relevant sources when a union   </text>
<query_num> 8904 </query_num>
<text>   can be directly executed by the Garlic execution engine. 4.1.3. Ariadne. Ariadne [3] is an integration system that uses the local-as-view mediation approach. The LOOM knowledge representation system =-=[36]-=- is used to construct the domain model that represents an integrated view of the sources. User queries are formulated against that model. Query processing has two phases: a preprocessing phase and a q   </text>
<query_num> 8905 </query_num>
<text>   cile the semantic differences between sources. The idea is to use axioms that describe how attributes and classes can be obtained by combining information from a set of sources. 4.1.4. Hermes. Hermes =-=[1]-=- considers the construction of mediators based on two major distinct tasks: domain integration—physical linking of the information sources, and thes198 OUZZANI AND BOUGUETTAYA semantic integration—coh   </text>
<query_num> 8906 </query_num>
<text>   cture and query processor. View Expander Logical Plan Matcher Sequencer Optimizer Physical Plan Execution Engine Sources Queries View Definitions Source Descriptions 4.4.1. Tsimmis. Tsimmis prototype =-=[24]-=- integrates heterogeneous information sources using the global-as-view approach for mediation (figure 4). The proposed mediation is supported through a number of concepts. An object model called OEM (   </text>
<query_num> 8907 </query_num>
<text>   dded to the ordering if its input requirements are satisfied with respect to the list of available parameters and source requirements, and it was not already considered. 4.4.3. Infomaster. Infomaster =-=[21]-=- provides integrated access to distributed heterogeneous information sources, including databases and semistructured data (figure 6). Infomaster uses the knowledge interchange format (KIF) as an inter   </text>
<query_num> 8908 </query_num>
<text>   describing data semantics and specifying metadata. Techniques for automatic data and metadata extraction and classification (ontologies, for example) are crucial for building tomorrow’s Semantic Web =-=[11]-=-. Query languages and query processing and optimization techniques need to be extended to exploit semantic information. Users also need adaptive systems to help them explore the Web and discover inter still relevant in some specific domains, providing query facilities over Web services covers a broader scope. The Web is geared towards an important paradigm shift through the concept of Semantic Web =-=[11]-=-. This basically means a transition from documents for human consumption to documents with machine-processable semantics. Allowing machine-processable content would be mainly based via another new par   </text>
<query_num> 8909 </query_num>
<text>   disjuncts is discarded. (5) Grouping: The conjuncts within each conjunction are grouped so that the atoms in each group share a common provider. 4.4.4. Annotated query plans. The approach proposed in =-=[22]-=- extends the System-R optimizer to deal with sources having limited query capabilities. It uses annotated query plans in the search space. Annotations describe the input that must be given to each sub   </text>
<query_num> 8910 </query_num>
<text>   e-optimize the query, or use specific operators that deal more flexibly with unpredictable events (e.g., data delivery rates). An interesting characterization of adaptive query processors is given in =-=[30]-=-. It states that a query processing is adaptive if: • it receives information from its environment, • it uses that information to determine its behavior, and • this process iterates over time, generat   </text>
<query_num> 8911 </query_num>
<text>   econd phase, the algorithm finds the best sub-plans for each cluster and then combines them to obtain the best feasible plan for the query. 4.4.6. Capability sensitive plans for selection queries. In =-=[23]-=-, two algorithms for generating plans for selection queries over Web data sources are presented. Sources exhibit restrictions on the size and structure of condition expressions along with limitations   </text>
<query_num> 8912 </query_num>
<text>   ent providers. Authorization information is recorded in a compatibility matrix that will annotate the query execution plan. The optimizer enumerates alternative query execution plans using a System-R =-=[49]-=- like dynamic algorithm. Costs of plans are estimated using information from the lookup service. If an information is missing it is set to a default value. The optimal plan is then executed by distrib   </text>
<query_num> 8913 </query_num>
<text>   es a generic cost model with specific cost information exported by wrappers. The generic cost model uses cost formulas established by the calibrating approach developed in the IRO-DB federated system =-=[25]-=-. The data source interface is specified using a subset of CORBA IDL extended with a cardinality section for data source statistics and a cost formula section for specific formulas. The wrapper writer   </text>
<query_num> 8914 </query_num>
<text>   hat cannot fulfill its QoS constraints as soon as possible. 4.2.2. HiQIQ. HiQIQ (High Quality Information Querying) uses information quality criteria to support query optimization in mediator systems =-=[40]-=-. The focus is on incorporating information quality into query planning. This could be useful in environments such as biological information systems where users are more sensitive to quality criteria   </text>
<query_num> 8915 </query_num>
<text>   he first phase prunes low-quality sources. The second phase finds all plans, i.e., combinations of QCAs, that produce semantically correct answers. The query planning strategy of Information Manifold =-=[34]-=- (see Section 4.4.2) is used. This phase does not use any quality related information. Finally, plans obtained from the previous phase are qualitatively ranked. This phase starts by determining IQ sco  Database Answers Object-oriented database Finally, the optimizer chooses the most efficient feasible plan using a cost based optimization algorithm. 4.4.2. Information manifold. Information Manifold =-=[34]-=- provides a uniform interface to structured information sources (figure 5). It proposes a mechanism to describe declaratively the content and query capabilities of information sources. The system uses   </text>
<query_num> 8916 </query_num>
<text>   imization information from autonomous information sources. All approaches either assume the availability of such statistics, or that they can be estimated or provided by wrappers. 4.1.1. Disco. Disco =-=[53]-=- is a mediator system based on the global-as-view approach (figure 2). The mediator generates multiple access plans involving local operations at the information sources and global operations at the m   </text>
<query_num> 8917 </query_num>
<text>   implemented in the context of a shared-nothing parallel query processing framework called River [6]. 4.3.2. Tukwila. Tukwila is another system addressing adaptiveness in data integration environment =-=[32]-=-. Adaptiveness is introduced at two levels: (1) between the optimizer and the execution engine, and (2) within the execution engine. In the first level, adaptiveness is deployed by annotating initial   </text>
<query_num> 8918 </query_num>
<text>   l plan. The algorithm applies an exhaustive enumeration of all possible placements of selections in the obtained plans. This process leads to the optimal plan. 4.4.5. Source sequencing algorithms. In =-=[55]-=-, a simple greedy strategy and a partitioning scheme are proposed to deal with limited capabilities of sources. Both algorithms use a cost model based on the number of source queries in the execution   </text>
<query_num> 8919 </query_num>
<text>   ources and data. The argument is that such quality parameters are more or as much important, in the context of the Web, than classical parameters like response time. 4.2.1. ObjectGlobe. ObjectGlobe’s =-=[17]-=- data integration is centered around three types of suppliers: data suppliers, function providers for query processing operators, and cycle providers for operators execution. The execution of a query   </text>
<query_num> 8920 </query_num>
<text>   participation in the system. It translates between the source’s local language, model, and concepts and those at the mediator level. To resolve a query, a mediator typically performs three main tasks =-=[27]-=-: ��� Database selection. Locate and select the databases that are relevant to the query. • Query translation. Decompose the query into sub-queries with respect to the previously selected databases. Eac   </text>
<query_num> 8921 </query_num>
<text>   ption and production rate, join implementation, and initial delays of input relations. Eddies have been implemented in the context of a shared-nothing parallel query processing framework called River =-=[6]-=-. 4.3.2. Tukwila. Tukwila is another system addressing adaptiveness in data integration environment [32]. Adaptiveness is introduced at two levels: (1) between the optimizer and the execution engine,   </text>
<query_num> 8922 </query_num>
<text>   r initially introduced in a seminal paper [54] by Gio Wiederhold. Most of the systems and techniques in this survey fall into that category. There are also other approaches based on the use of agents =-=[41]-=-, ontology [2, 37, 43], and information retrieval techniques [5, 33, 38]. This section gives some details about the mediator approach. Then, it highlights major research issues related to query optimi   </text>
<query_num> 8923 </query_num>
<text>   roduced in a seminal paper [54] by Gio Wiederhold. Most of the systems and techniques in this survey fall into that category. There are also other approaches based on the use of agents [41], ontology =-=[2, 37, 43]-=-, and information retrieval techniques [5, 33, 38]. This section gives some details about the mediator approach. Then, it highlights major research issues related to query optimization for data integr   </text>
<query_num> 8924 </query_num>
<text>   selects a winning plan, that plan is translated into an executable form. GarlicPOPs are translated into operators that can be directly executed by the Garlic execution engine. 4.1.3. Ariadne. Ariadne =-=[3]-=- is an integration system that uses the local-as-view mediation approach. The LOOM knowledge representation system [36] is used to construct the domain model that represents an integrated view of the   </text>
<query_num> 8925 </query_num>
<text>   ta integration systems. Wealso present a classification of the different presented techniques and a comprehensive framework to evaluate them. Most of the described systems adopt the mediator approach =-=[54]-=- for data integration. Those are systems that match information requests from consumers, individual users or applications, to information providers. This survey classifies the different systems accord  optimization techniques. Different approaches have been used for Web-based data integration. The most widely used approach is based on the concept of mediator initially introduced in a seminal paper =-=[54]-=- by Gio Wiederhold. Most of the systems and techniques in this survey fall into that category. There are also other approaches based on the use of agents [41], ontology [2, 37, 43], and information re   </text>
<query_num> 8926 </query_num>
<text>   tables used in the plan, output columns, and estimated cost. Garlic extends the traditional cost-based optimization approach by involving wrappers as important components in the optimization process =-=[46]-=-. Wrappers cooperate in the estimation of the total cost of a given query plan. The proposed framework aims to provide the necessary means to extend the traditional cost-based optimization to a hetero   </text>
<query_num> 8927 </query_num>
<text>   te unpredictable and bursty data-flows through computing resources. The query processor continuously reorders applications of pipelined operators in a query plan at run-time on a tuple-by-tuple basis =-=[7]-=-. Is uses the concept of eddy, defined as a n-ary tuple router interposed between n data sources and a set of query processing operators. An eddy encapsulates the ordering of operators by dynamically   </text>
<query_num> 8928 </query_num>
<text>   the leaves and then from the leaves to the root. In the first traversal, cost formulas are associated with nodes. In the second traversal, the cost of each operator is computed. 4.1.2. Garlic. Garlic =-=[28, 47]-=- provides an integrated view over heterogeneous information sources using the global-as-view approach (figure 3). Query processing is based on dynamically determining the middleware and wrapper roles   </text>
<query_num> 8929 </query_num>
<text>   to data managed by multiple databases. A mechanism to achieve that goal is through an integrated schema that involves all or parts of the component schemas. The taxonomy presented by Sheth and Larson =-=[51]-=- classifies the existing solutions in three categories: global schema integration, federated databases, and multidatabase language approach. These categories are presented according to how tightly int   </text>
</top>
<top>
<paper_num> 90 </paper_num>
<paper_title>   Run-Time Support for Distributed Sharing in Typed Languages.  </paper_title>
<paper_abstract>   This paper addresses run-time support for sharing objects in a typed language between the different computers within a cluster. Typing must be strong enough that it is possible to determine unambiguously whether a memory location contains an object reference or not. Many modern languages fall under this category, including Java and Modula-3. Direct access through a reference to object data is supported, unlike Java/RMI or Orca [2], where remote object access is restricted to method invocation. Furthermore, in languages with suitable multithreading support, such as Java, distributed execution is transparent: no new API is introduced for distributed sharing. This transparency distinguishes this work from many earlier distributed object sharing systems [2, 6, 13, 11].  </paper_abstract>
<query_num> 9001 </query_num>
<text>   (e.g., [1, 12]). Conventional VM-based DSM systems have only achieved good performance on relatively coarse-grained applications, because of their reliance on VM pages. Although relaxed memory models =-=[8]-=- and multiple-writer protocols [5] reduce the impact of the large page size, fine-grained sharing and false sharing remain problematic [1]. Fine-grained DSM systems have been built using code instrume ts, and therefore their elements can be written concurrently. Of course, for correctness, the different processes must write to disjoint elements in the arrays. The object space is release consistent =-=[8]-=-. In essence, under release consistency, the propagation of updates from one processor to another may be delayed until the processors synchronize. Parallel programs that are properly synchronized (i.e   </text>
<query_num> 9002 </query_num>
<text>   d garbage collector on both TreadMarks and DOSA that is representative of the state-of-the-art, and evaluated our distributed garbage collectors using a modified version of the OO7 database benchmark =-=[4]-=-. Our evaluation shows that DOSA considerably outperforms TreadMarks for garbage-collected appli7 cations. Again, due to the page limit, the detailed result of this performance evaluation has been omi   </text>
<query_num> 9003 </query_num>
<text>   er, maintained in terms of objects rather than pages. In other words, consistency messages specify object identifiers instead of page numbers. For individual objects, a single writer protocol is used =-=[10]-=-. For arrays of objects, whether of a scalar type or a reference type, a multiple writer protocol is used [1]. This permits the use of a single OID for the entire array, while still allowing concurren   </text>
<query_num> 9004 </query_num>
<text>   gure 7 shows various statistics from the execution of these applications on 32 processors for both problem sizes. 8 Related Work Two other systems have used VM mechanisms for finegrain DSM: Millipede =-=[9]-=- and the Region Trapping Library [3]. The fundamental difference between DOSA and these systems is that DOSA takes advantage of a typed language to distinguish a pointer from data at run-time and thes   </text>
<query_num> 9005 </query_num>
<text>   h pointers from data at run-time enables more efficient fine-grained sharing than is possible with conventional distributed shared memory (DSM) implementations that do not use type information (e.g., =-=[1, 12]-=-). Conventional VM-based DSM systems have only achieved good performance on relatively coarse-grained applications, because of their reliance on VM pages. Although relaxed memory models [8] and multip   </text>
<query_num> 9006 </query_num>
<text>   h pointers from data at run-time enables more efficient fine-grained sharing than is possible with conventional distributed shared memory (DSM) implementations that do not use type information (e.g., =-=[1, 12]-=-). Conventional VM-based DSM systems have only achieved good performance on relatively coarse-grained applications, because of their reliance on VM pages. Although relaxed memory models [8] and multip pared its performance to that of TreadMarks, a state-of-the-art pageDepartment of Computer Science Rice University Houston, Texas 77005 ychu, weimin, alc, dwallach, willy¡ @cs.rice.edu 1 based system =-=[1]-=-. We have derived our implementation from the TreadMarks code base, thereby avoiding performance differences due to irrelevant code differences. Our performance evaluation substantiates the following  e cache locality or during garbage collection, without affecting the other processors. 3.5 Consistency Protocol DOSA, like TreadMarks, uses a lazy invalidate protocol to implement release consistency =-=[1]-=-. Consistency is, however, maintained in terms of objects rather than pages. In other words, consistency messages specify object identifiers instead of page numbers. For individual objects, a single w   </text>
<query_num> 9007 </query_num>
<text>   ontains an object reference or not. Many modern languages fall under this category, including Java and Modula-3. Direct access through a reference to object data is supported, unlike Java/RMI or Orca =-=[2]-=-, where remote object access is restricted to method invocation. Furthermore, in languages with suitable multithreading support, such as Java, distributed execution is transparent: no new API is intro ions. This decision has the unfortunate side effect of forcing modifications made in the read-write region to be copied to the read region, every time protection changes from read-write to read. Orca =-=[2]-=-, Jade [11], COOL [6], and SAM [13] are parallel or distributed object-oriented langages. All of these systems differ from ours in that they present a new language or API to the programmer to express   </text>
<query_num> 9008 </query_num>
<text>   ry, and runs FreeBSD 2.2.6. We demonstrate the performance improvements of DOSA over TreadMarks for fine-grained applications, by using Barnes-Hut and Water-Spatial, both from the SPLASH-2 benchmarks =-=[15]-=-. SOR and Water-Nsquared from the SPLASH benchmarks [14] demonstrate only minimal performance loss for coarse-grained applications. For each of these applications, Table 1 lists each of the problem si   </text>
<query_num> 9009 </query_num>
<text>   se sharing remain problematic [1]. Fine-grained DSM systems have been built using code instrumentation, but they have been limited by the cost of instrumentation and lack of communication aggregation =-=[7]-=-. The system presented here, DOSA, uses the ability to distinguish pointers from data at run-time to achieve efficient fine-grained sharing using VM support and without using instrumentation. It does   DOSA aims to provide transparent object sharing for existing typed languages, such as Java. Furthermore, none of Orca, Jade, COOL, or SAM use VM-based mechanisms for object sharing. Dwarkadas et al. =-=[7]-=- compared Cashmere, a coarsegrained system, somewhat like TreadMarks, and Shasta, an instrumentation-based system, running on an identical platform – a cluster of four 4-way AlphaServers connected by   </text>
<query_num> 9010 </query_num>
<text>   t, such as Java, distributed execution is transparent: no new API is introduced for distributed sharing. This transparency distinguishes this work from many earlier distributed object sharing systems =-=[2, 6, 13, 11]-=-. The key insight in this paper is that the ability to distinguish pointers from data at run-time enables more efficient fine-grained sharing than is possible with conventional distributed shared memo  decision has the unfortunate side effect of forcing modifications made in the read-write region to be copied to the read region, every time protection changes from read-write to read. Orca [2], Jade =-=[11]-=-, COOL [6], and SAM [13] are parallel or distributed object-oriented langages. All of these systems differ from ours in that they present a new language or API to the programmer to express distributed   </text>
<query_num> 9011 </query_num>
<text>   t, such as Java, distributed execution is transparent: no new API is introduced for distributed sharing. This transparency distinguishes this work from many earlier distributed object sharing systems =-=[2, 6, 13, 11]-=-. The key insight in this paper is that the ability to distinguish pointers from data at run-time enables more efficient fine-grained sharing than is possible with conventional distributed shared memo unate side effect of forcing modifications made in the read-write region to be copied to the read region, every time protection changes from read-write to read. Orca [2], Jade [11], COOL [6], and SAM =-=[13]-=- are parallel or distributed object-oriented langages. All of these systems differ from ours in that they present a new language or API to the programmer to express distributed sharing, while DOSA doe   </text>
<query_num> 9012 </query_num>
<text>   the execution of these applications on 32 processors for both problem sizes. 8 Related Work Two other systems have used VM mechanisms for finegrain DSM: Millipede [9] and the Region Trapping Library =-=[3]-=-. The fundamental difference between DOSA and these systems is that DOSA takes advantage of a typed language to distinguish a pointer from data at run-time and these other systems do not. This allows   </text>
</top>
<top>
<paper_num> 91 </paper_num>
<paper_title>   Implementing Minimized Multivariate PKC on Low-Resource Embedded Systems.  </paper_title>
<paper_abstract>   Abstract. Multivariate (or MQ) public-key cryptosystems (PKC) are alternatives to traditional PKCs based on large algebraic structures (e.g., RSA and ECC); they usually execute much faster than traditional PKCs on the same hardware. However, one major challenge in implementing multivariates in embedded systems is that the key size can be prohibitively large for applications with stringent resource constraints such as low-cost smart cards, sensor networks (e.g., Berkeley motes), and radio-frequency identification (RFID). In this paper, we investigate strategies for shortening the key of a multivariate PKC. We apply these strategies to the Tame Transformation Signatures (TTS) as an example and quantify the improvement in key size and running speed, both theoretically and via implementation. We also investigate ways to save die space and energy consumption in hardware, reporting on our ASIC implementation of TTS on a TSMC 0.25µm process. Even without any key shortening, the current consumption of TTS is only 21 µA for computing a signature, using 22,000 gate equivalents and 16,000 100-kHz cycles (160 ms). With circulant-matrix key shortening, the numbers go down to 17,000 gates and 4,400 cycles (44 ms). We therefore conclude: besides representing a future-proofing investment against the emerging quantum computers, multivariates can be immediately useful in niches.  </paper_abstract>
<query_num> 9101 </query_num>
<text>   eting at RFID applications. We conclude with the discussion in Sec. 6. Some of the diversed results referenced are summarized in the appendix; the rest needs to be looked up from the original sources =-=[8, 10, 11, 13, 27, 37, 41, 43, 44]-=-. 2 Tame Transformation Signatures Before we move on, we need to emphasize that there is no reductionist proof of security today; the security of MQ-schemes is so far thrust-and-parry. Since this is m here are four basic ways [41] one can set up the trapdoor of a multivariate scheme for φ2 to be inverted: Imai-Matsumoto’s C ∗ [31] (and derivatives SFLASH [39] and PMI+ [10]), Hidden Field Equations =-=[37]-=-, “Unbalanced Oil-and-Vinegar” [27], and “Stepwise Triangular System” [43]. Basic trapdoors must be modified for security — see [41] for a summary of modifications. All have O(n 3 )-time (and space) p   </text>
<query_num> 9102 </query_num>
<text>   eting at RFID applications. We conclude with the discussion in Sec. 6. Some of the diversed results referenced are summarized in the appendix; the rest needs to be looked up from the original sources =-=[8, 10, 11, 13, 27, 37, 41, 43, 44]-=-. 2 Tame Transformation Signatures Before we move on, we need to emphasize that there is no reductionist proof of security today; the security of MQ-schemes is so far thrust-and-parry. Since this is m ks [43], including linear algebra attacks mentioned above, algebraic attacks based on Faugère’sImplementing Minimized Multivariate PKC on Low-Resource Embedded Systems 77 F5 and XL of Courtois et al =-=[8, 13]-=-, improved search methods [4], and some methods tailored to specific schemes [19, 20, 36]. We repeated experiments checking that no vulnerabilities had been reintroduced in our short-key versions. 2.3   </text>
<query_num> 9103 </query_num>
<text>   h is expected to be in the future releases of the device. Even with such an impediment, the results are quite promising compared with previous results obtained using more traditional PKCs such as ECC =-=[30]-=-. 3 From Previous Attempts at Key Shortening to Scheduling Many attempts to use subfields for smaller keys in multivariates have been made. The original version of SFLASH used a subfield but was broke   </text>
<query_num> 9104 </query_num>
<text>   luding the nomenclature and the state of the art. 1.1 Advantages That Accrue to Multivariates When quantum computers (QCs) with thousands of qubits arrive, RSA and discrete-log schemes will be broken =-=[40]-=-. MQ- and lattice-type schemes stand with halved logcomplexity ([22], e.g., 2200 → 2100 ). We also have quantum key exchange but not quantum signatures. So MQ-schemes seem like future-proofing insuran   </text>
<query_num> 9105 </query_num>
<text>   rating system [23]. TinyOS is a modular event-driven operating system designed specifically for sensor network applications. It has a component-based programming model, supported by the nesC language =-=[18]-=-. Applications are written as modules that are composed together through a set of predefined interfaces, which describe the events to be handled and the implemented commands. The composition of module   </text>
<query_num> 9106 </query_num>
<text>   sary to compute φ −1 2 . The security of a multivariate PKC thus depends on the infeasibility of decomposing maps and that of solving a large system of polynomial equations. The reader is referred to =-=[41]-=- for a comprehensive reference for multivariate PKCs, including the nomenclature and the state of the art. 1.1 Advantages That Accrue to Multivariates When quantum computers (QCs) with thousands of qu eting at RFID applications. We conclude with the discussion in Sec. 6. Some of the diversed results referenced are summarized in the appendix; the rest needs to be looked up from the original sources =-=[8, 10, 11, 13, 27, 37, 41, 43, 44]-=-. 2 Tame Transformation Signatures Before we move on, we need to emphasize that there is no reductionist proof of security today; the security of MQ-schemes is so far thrust-and-parry. Since this is m erience and history. In tweaking them appropriately, one can expect to inherit many of the good properties and reuse code, which introduces fewer opportunities for mistakes. There are four basic ways =-=[41]-=- one can set up the trapdoor of a multivariate scheme for φ2 to be inverted: Imai-Matsumoto’s C ∗ [31] (and derivatives SFLASH [39] and PMI+ [10]), Hidden Field Equations [37], “Unbalanced Oil-and-Vin  history, we examine two such earlier attempts in Sec. 4.1. Here we must mention the two recent other papers dealing with equivalency and normal forms of multivariates, [25] and [42]. As mentioned by =-=[41]-=-, sparse variants like TTS are not affected much by such equivalency concerns. 4.1 Earlier Attempts Two ideas heretofore seen in print tried replacing multiplication by a non-singular matrix by a sequ   </text>
<query_num> 9107 </query_num>
<text>   t Accrue to Multivariates When quantum computers (QCs) with thousands of qubits arrive, RSA and discrete-log schemes will be broken [40]. MQ- and lattice-type schemes stand with halved logcomplexity (=-=[22]-=-, e.g., 2200 → 2100 ). We also have quantum key exchange but not quantum signatures. So MQ-schemes seem like future-proofing insurance. Another reason is that the private map of RSA is intrinsically s   </text>
<query_num> 9108 </query_num>
<text>   ties and reuse code, which introduces fewer opportunities for mistakes. There are four basic ways [41] one can set up the trapdoor of a multivariate scheme for φ2 to be inverted: Imai-Matsumoto’s C ∗ =-=[31]-=- (and derivatives SFLASH [39] and PMI+ [10]), Hidden Field Equations [37], “Unbalanced Oil-and-Vinegar” [27], and “Stepwise Triangular System” [43]. Basic trapdoors must be modified for security — see   </text>
<query_num> 9109 </query_num>
<text>   use a stream cipher like RC4 (with a variable key size), or a fast linear feed-back pseudo random number generator (PRNG) such as TT800 (a precursor to the mersenne twister, up to 100-byte keys, cf. =-=[32]-=-). It seems natural to treat each routine as a PRNG or stream cipher. Two observations: – We store sub-keys (random numbers) in a buffer and take inputs as needed. – If we wish to output the public ke   </text>
</top>
<top>
<paper_num> 92 </paper_num>
<paper_title>   Probabilistic logic learning.  </paper_title>
<paper_abstract>   The past few years have witnessed an significant interest in probabilistic logic learning, i.e. in research lying at the intersection of probabilistic reasoning, logical representations, and machine learning. A rich variety of di#erent formalisms and learning techniques have been developed. This paper provides an introductory survey and overview of the stateof -the-art in probabilistic logic learning through the identification of a number of important probabilistic, logical and learning concepts.  </paper_abstract>
<query_num> 9201 </query_num>
<text>   - Early work on incorporating probabilities into logic programs was done by Clark and McCabe [9] and by Shapiro [97]. Subsequently, researchers such as Nilsson [75], Halpern [43], Ng and Subrahmanian =-=[71]-=-, and Poole [82] have studied probabilistic logics from a knowledge representational perspective. The aim of this research is more a probabilistic characterization of logic than suitable representatio   </text>
<query_num> 9202 </query_num>
<text>   Section 4.1 — the set of random variables defined by a Bayesian logic program corresponds to a Herbrand model. This is akin to the learning from interpretations setting in inductive logic programming =-=[18]-=-. The requirement is enforced when learning the structure of Bayesian logic programs by starting from an initial Bayesian logic programs that satisfies this requirement (such a hypothesis can be compu   </text>
<query_num> 9203 </query_num>
<text>   al representations or even logic programs. This however comes at a computational cost. Therefore, some recent approaches such as relational Markov models (RMMs) [2], hidden tree Markov models (HTMMs) =-=[28; 21]-=-, and logical hidden Markov models (LOHMMs) [55] have tried to upgrade some well understood probabilistic representations, such as (Hidden) Markov Models [84]. These approaches can also be viewed as d n Markov models (instead of the simpler Markov models presented above). Kersting et al. [55] present a bioinformatics application on predicting the fold class of proteins. • Hidden tree Markov models =-=[21]-=- and other hidden Markov models over data structures [28] focus on modelling probability distributions defined over spaces of trees and graphs. They do not use logical concepts such as predicates. Ins oretical concepts such as deterministic transduction to generically define the dependency structure among random variables. Diligenti et al. present an application on document image classification in =-=[21]-=-. Further possible choices for these approaches exist. One of these is concerned with the level at which to specify the domain probabilities. One can specify them locally, i.e., for each ”predicate” i   </text>
<query_num> 9204 </query_num>
<text>   al representations or even logic programs. This however comes at a computational cost. Therefore, some recent approaches such as relational Markov models (RMMs) [2], hidden tree Markov models (HTMMs) =-=[28; 21]-=-, and logical hidden Markov models (LOHMMs) [55] have tried to upgrade some well understood probabilistic representations, such as (Hidden) Markov Models [84]. These approaches can also be viewed as d sented above). Kersting et al. [55] present a bioinformatics application on predicting the fold class of proteins. • Hidden tree Markov models [21] and other hidden Markov models over data structures =-=[28]-=- focus on modelling probability distributions defined over spaces of trees and graphs. They do not use logical concepts such as predicates. Instead, they adapt automatatheoretical concepts such as det   </text>
<query_num> 9205 </query_num>
<text>   ation within databases [31]. 4.2 Probabilistic Proofs To define probabilities on proofs, there are at least two different options. They correspond to the approaches taken in stochastic logic programs =-=[24; 66; 12; 13; 15]-=- and PRISMs respectively [88; 90; 50; 89]. Let us start by introducing stochastic logic programs (SLPs). Stochastic logic programs are a direct upgrade of stochastic context free grammars. This is rea 4 · (0.2 · 0.5 · 0.2 · 0.2)] = 0.624 Thus, the normalized probability of :− mc(1m, r) is (2 · 0.004)/0.624 = 0.012820513. More details on inference (including approximative inference) can be found i=-=n [13; 15]-=-. Stochastic logic programs have been used to specify declarative priors for probabilistic models [3]. Whereas stochastic logic programs attach probability labels to all clauses, PRISMs and the earlie   </text>
<query_num> 9206 </query_num>
<text>   ation within databases [31]. 4.2 Probabilistic Proofs To define probabilities on proofs, there are at least two different options. They correspond to the approaches taken in stochastic logic programs =-=[24; 66; 12; 13; 15]-=- and PRISMs respectively [88; 90; 50; 89]. Let us start by introducing stochastic logic programs (SLPs). Stochastic logic programs are a direct upgrade of stochastic context free grammars. This is rea mc/2 is the sum of probabilities of all refutations of the most gen4 More general definitions of stochastic logic programs where e.g. the probability values do not sum to 1.0 are discussed by Cussens =-=[12; 14]-=-. Because learning of stochastic logic programs is considered only for ‘normalized’ programs we restrict our attention to them. SIGKDD Explorations. Volume 2, Issue 2 - page 8seral goal over mc(X, Y),   </text>
<query_num> 9207 </query_num>
<text>   c Proofs To define probabilities on proofs, there are at least two different options. They correspond to the approaches taken in stochastic logic programs [24; 66; 12; 13; 15] and PRISMs respectively =-=[88; 90; 50; 89]-=-. Let us start by introducing stochastic logic programs (SLPs). Stochastic logic programs are a direct upgrade of stochastic context free grammars. This is realized by associating to each clause a pro   </text>
<query_num> 9208 </query_num>
<text>   c Proofs To define probabilities on proofs, there are at least two different options. They correspond to the approaches taken in stochastic logic programs [24; 66; 12; 13; 15] and PRISMs respectively =-=[88; 90; 50; 89]-=-. Let us start by introducing stochastic logic programs (SLPs). Stochastic logic programs are a direct upgrade of stochastic context free grammars. This is realized by associating to each clause a pro , the improved estimates for each clause are obtained by dividing the clause’s expected counts by the sum of the expected counts of clauses for the same predicate. The EM algorithm for PRISM programs =-=[50; 91; 49; 92]-=- is similar in spirit. The main differences are that (1) no probability values are associated to (intensional) clauses, (2) failed derivations have a probability of zero, and (3) the disjoint represen   </text>
<query_num> 9209 </query_num>
<text>   c reasoning mechanisms with machine learning and data mining principles. In the past few years, this question has received a lot of attention and various different approaches have been developed, cf. =-=[82; 41; 88; 66; 73; 45; 60; 61; 57; 81; 52; 77; 2; 55; 87]-=-. This paper provides an introductory survey and overview of those developments that lie at the intersection of logical (or relational) representations, probabilistic reasoning and learning, cf. Figur   </text>
<query_num> 9210 </query_num>
<text>   c reasoning mechanisms with machine learning and data mining principles. In the past few years, this question has received a lot of attention and various different approaches have been developed, cf. =-=[82; 41; 88; 66; 73; 45; 60; 61; 57; 81; 52; 77; 2; 55; 87]-=-. This paper provides an introductory survey and overview of those developments that lie at the intersection of logical (or relational) representations, probabilistic reasoning and learning, cf. Figur  proofs. 4.1 Probabilistic Logical Models The first class of representations extends Bayesian networks with abilities to define probabilities on first order logical or relational interpretations, cf. =-=[82; 41; 73; 45; 57; 81; 31; 52]-=-. Important predecessors of these logical Bayesian networks discussed below, include the work by Breese, Charniak, Bacchus, Goldman, Poole and Wellmann [40; 6; 4; 7; 39]. These predecessors were large hat can be used to answer the query. Many of these approaches can represent Bayesian networks. However, they are not direct upgrades of Bayesian networks in the same way that the work by e.g. Haddawy =-=[41]-=- is. This may also explain why – to the authors’ knowledge – Bayesian network learning techniques have not been applied to these representations. A central contribution in combining clausal logic with n, i.e., they may have more than two possible values, or even be continuous. There are two ways to realize this. First, one could simply allow the domains of the random variables to be arbitrary, cf. =-=[41; 52]-=-. One side-effect is that boolean or logical random variables get mixed up with SIGKDD Explorations. Volume 2, Issue 2 - page 6sother ones. This may sometimes cause some confusion. Second, one could s al frameworks that have been developed around these principles. As mentioned before, the first approach to directly upgrade Bayesian networks to relational representations, seems due to Peter Haddawy =-=[41]-=-. In his initial approach, the random variables correspond to ground atoms. Furthermore, the need for combining rules was alleviated by requiring that for each random variable, there was at most one i   </text>
<query_num> 9211 </query_num>
<text>   c reasoning mechanisms with machine learning and data mining principles. In the past few years, this question has received a lot of attention and various different approaches have been developed, cf. =-=[82; 41; 88; 66; 73; 45; 60; 61; 57; 81; 52; 77; 2; 55; 87]-=-. This paper provides an introductory survey and overview of those developments that lie at the intersection of logical (or relational) representations, probabilistic reasoning and learning, cf. Figur  proofs. 4.1 Probabilistic Logical Models The first class of representations extends Bayesian networks with abilities to define probabilities on first order logical or relational interpretations, cf. =-=[82; 41; 73; 45; 57; 81; 31; 52]-=-. Important predecessors of these logical Bayesian networks discussed below, include the work by Breese, Charniak, Bacchus, Goldman, Poole and Wellmann [40; 6; 4; 7; 39]. These predecessors were large stantiations of q(X, Y) for a single X. Various solutions which might even be combined have been developed for dealing with this situation. First, one could introduce so-called combination rules, cf. =-=[73; 45; 52]-=-. Combination rules are functions that would – in the above case – take the corresponding conditional probability distributions P(p | q, r) and P(p | s, t) as inputs and would produce the desired P(p   </text>
<query_num> 9212 </query_num>
<text>   c reasoning mechanisms with machine learning and data mining principles. In the past few years, this question has received a lot of attention and various different approaches have been developed, cf. =-=[82; 41; 88; 66; 73; 45; 60; 61; 57; 81; 52; 77; 2; 55; 87]-=-. This paper provides an introductory survey and overview of those developments that lie at the intersection of logical (or relational) representations, probabilistic reasoning and learning, cf. Figur ic models with very expressive relational representations or even logic programs. This however comes at a computational cost. Therefore, some recent approaches such as relational Markov models (RMMs) =-=[2]-=-, hidden tree Markov models (HTMMs) [28; 21], and logical hidden Markov models (LOHMMs) [55] have tried to upgrade some well understood probabilistic representations, such as (Hidden) Markov Models [8 r the underlying representations can almost directly be applied. Let us illustrate this idea on Markov models (where we follow the ideas underlying RMMs and LOHMMs using an example by Anderson et al. =-=[2]-=-). A Markov model is essentially a finite state automaton with probabilities associated to each of the edges instead of symbols from an alphabet. Alternatively, it can be regarded as stochastic regula dro). This illustrates the key ideas underlying these representations: proof steps correspond to time steps. Nevertheless, various differences among these approaches exist: • Relational Markov Models =-=[2]-=- do not allow for variables and unification. Instead they employ a taxonomy to define the domains and probability estimation trees [83] to specify the domain distributions. Anderson et al. [83] presen   </text>
<query_num> 9213 </query_num>
<text>   c reasoning mechanisms with machine learning and data mining principles. In the past few years, this question has received a lot of attention and various different approaches have been developed, cf. =-=[82; 41; 88; 66; 73; 45; 60; 61; 57; 81; 52; 77; 2; 55; 87]-=-. This paper provides an introductory survey and overview of those developments that lie at the intersection of logical (or relational) representations, probabilistic reasoning and learning, cf. Figur incorporating probabilities into logic programs was done by Clark and McCabe [9] and by Shapiro [97]. Subsequently, researchers such as Nilsson [75], Halpern [43], Ng and Subrahmanian [71], and Poole =-=[82]-=- have studied probabilistic logics from a knowledge representational perspective. The aim of this research is more a probabilistic characterization of logic than suitable representations for learning.  proofs. 4.1 Probabilistic Logical Models The first class of representations extends Bayesian networks with abilities to define probabilities on first order logical or relational interpretations, cf. =-=[82; 41; 73; 45; 57; 81; 31; 52]-=-. Important predecessors of these logical Bayesian networks discussed below, include the work by Breese, Charniak, Bacchus, Goldman, Poole and Wellmann [40; 6; 4; 7; 39]. These predecessors were large  been used to specify declarative priors for probabilistic models [3]. Whereas stochastic logic programs attach probability labels to all clauses, PRISMs and the earlier representation by David Poole =-=[82]-=- attach probability labels to facts. Indeed, let a logic program consist of the set of facts E (this could be considered the extensional part) and the set of proper clauses I (the intension). The fram turn explains the strong link with 0.7 department 0.1 0.3 0.3 course 0.2 0.8 0.3 0.1 0.1 lecturer Figure 11: The graph structure of a Markov model for web navigation. abductive inference methods, cf. =-=[82]-=-. Finally, let us remark that we discussed simplified versions of Poole’s and Sato’s frameworks. Originally, they have been introduced to deal with situations where some of the facts are mutually excl   </text>
<query_num> 9214 </query_num>
<text>   c reasoning mechanisms with machine learning and data mining principles. In the past few years, this question has received a lot of attention and various different approaches have been developed, cf. =-=[82; 41; 88; 66; 73; 45; 60; 61; 57; 81; 52; 77; 2; 55; 87]-=-. This paper provides an introductory survey and overview of those developments that lie at the intersection of logical (or relational) representations, probabilistic reasoning and learning, cf. Figur ion rule. Finally, there is the quite popular formalism of probabilistic relational models 3 (PRMs) [57; 81; 31; 33]. This formalism 3 Several other extensions of Bayesian network have been developed =-=[60; 58; 61; 80; 81]-=- by Pfeffer et al. before introducing probabilistic relational models. These extensions have been termed frame-based and object-oriented, in which the relational and logical issues are less direct. Th   </text>
<query_num> 9215 </query_num>
<text>   c reasoning mechanisms with machine learning and data mining principles. In the past few years, this question has received a lot of attention and various different approaches have been developed, cf. =-=[82; 41; 88; 66; 73; 45; 60; 61; 57; 81; 52; 77; 2; 55; 87]-=-. This paper provides an introductory survey and overview of those developments that lie at the intersection of logical (or relational) representations, probabilistic reasoning and learning, cf. Figur lls :− alarm. marycalls :− alarm. Figure 2: The alarm program. the relational model is not expressive enough to model the logical component of many of the probabilistic logical representations (e.=-=g., [88; 66; 90; 52]-=-). Nevertheless, we have attempted to be self-contained and to reduce the required logical machinery and background as much as possible. The paper is organized as follows: Sections 2 and 3 introduce t  not Turingequivalent, i.e. they cannot be used as a programming language. Thus it is useful to lift the underlying grammar representation to that of logic programs, a Turing equivalent language, cf. =-=[66]-=-. SIGKDD Explorations. Volume 2, Issue 2 - page 5s4. FIRST ORDER PROBABILISTIC LOGICS By now, everything is in place to introduce the key representational frameworks that combine probabilistic reasoni ation within databases [31]. 4.2 Probabilistic Proofs To define probabilities on proofs, there are at least two different options. They correspond to the approaches taken in stochastic logic programs =-=[24; 66; 12; 13; 15]-=- and PRISMs respectively [88; 90; 50; 89]. Let us start by introducing stochastic logic programs (SLPs). Stochastic logic programs are a direct upgrade of stochastic context free grammars. This is rea   </text>
<query_num> 9216 </query_num>
<text>   c reasoning mechanisms with machine learning and data mining principles. In the past few years, this question has received a lot of attention and various different approaches have been developed, cf. =-=[82; 41; 88; 66; 73; 45; 60; 61; 57; 81; 52; 77; 2; 55; 87]-=-. This paper provides an introductory survey and overview of those developments that lie at the intersection of logical (or relational) representations, probabilistic reasoning and learning, cf. Figur lls :− alarm. marycalls :− alarm. Figure 2: The alarm program. the relational model is not expressive enough to model the logical component of many of the probabilistic logical representations (e.=-=g., [88; 66; 90; 52]-=-). Nevertheless, we have attempted to be self-contained and to reduce the required logical machinery and background as much as possible. The paper is organized as follows: Sections 2 and 3 introduce t c Proofs To define probabilities on proofs, there are at least two different options. They correspond to the approaches taken in stochastic logic programs [24; 66; 12; 13; 15] and PRISMs respectively =-=[88; 90; 50; 89]-=-. Let us start by introducing stochastic logic programs (SLPs). Stochastic logic programs are a direct upgrade of stochastic context free grammars. This is realized by associating to each clause a pro  distribution at the level of interpretations and so the formalisms by Sato and Poole, can also be interpreted from a model-theoretic perspective. Sato has called this a distributional semantics, cf. =-=[88]-=-, stressing the declarative character of PRISMs. However, the key difference with the Bayesian network approaches presented above is that the former approaches rely more on logical inference methods,   </text>
<query_num> 9217 </query_num>
<text>   d Bayesian network. BLPs deal with continuous random variables and employ a simpler form of combination rule. Finally, there is the quite popular formalism of probabilistic relational models 3 (PRMs) =-=[57; 81; 31; 33]-=-. This formalism 3 Several other extensions of Bayesian network have been developed [60; 58; 61; 80; 81] by Pfeffer et al. before introducing probabilistic relational models. These extensions have bee entity corresponding to the Mother. Furthermore, in the basic PRM setting the relations among entities (such as Mother or Father in the genetic example) are assumed to be deterministic and given, cf. =-=[33]-=-. This implies that it would be impossible to specify that the probability is 0.70 that Jef is the father of Mary. Nevertheless, some partial solutions for dealing with probabilistic relations have be  the key ideas and will not address any advanced issues such as efficiency concerns. 5.2.1 Model Theoretic Parameter estimations for probabilistic-logic programs [59], probabilistic relational models =-=[30; 31; 33]-=-, and Bayesian logic programs [51; 54] all follow the same principle: The given data and the current model induce a Bayesian network explaining each data cases. Then, the parameters of the induced Bay erent aggregate functions and where there is at most one clause defining a predicate (such as p/1). Structure learning with probabilistic relational models now occurs by refining this type of clauses =-=[30; 36; 34; 33; 31]-=- 11 . Aggregated literals such as τ1(q(X, Y)) can be added to or deleted from clauses. At this point, the graphical notation for probabilistic relational models (illustrated in Figure 10) is convenien   </text>
<query_num> 9218 </query_num>
<text>   eresting work lying on boundaries such as Flach and Lachiche’s work on Naïve Bayes for structured terms [27; 62] and Craven and Slattery’s work on combining Naïve Bayes with a relational rule learner =-=[11]-=-. Giving an overview of the pairwise intersections (let alone the underlying domains) would lead us too far (as it would require a book instead of a paper) and is therefore beyond the scope of this pa   </text>
<query_num> 9219 </query_num>
<text>   ferences among these approaches exist: • Relational Markov Models [2] do not allow for variables and unification. Instead they employ a taxonomy to define the domains and probability estimation trees =-=[83]-=- to specify the domain distributions. Anderson et al. [83] present impressive experimental results on adaptive web navigation. • Logical Hidden Markov Models [55] do allow for variables and unificatio nly select the most informative transitions — are needed. For RMMs, this is realized by 1) employing one abstract transition for each predicate pair, and 2) by assigning a probability estimation tree =-=[83]-=-, a kind of decision tree, to each such predicate. The probability estimation tree will assign to any given concrete state in the body part of the transition a probability distribution over the possib   </text>
<query_num> 9220 </query_num>
<text>   have been successfully applied to various problems such as relational clustering [98], hypertext classification [37], selectivity estimation within databases [38], and modelling identity uncertainty =-=[76]-=-, and led to impressive results within computational biology [95; 93; 94]. Recently, Getoor et al. [31; 35; 32] have introduced stochastic relational models (SRMs). At the syntactic level SRMs are alm   </text>
<query_num> 9221 </query_num>
<text>   in a natural way. This is advantageous because it reduces the size of the conditional probability distributions in a natural way, but – on the other hand – it may also result in a loss of information =-=[46]-=-. Above we introduced the principles underlying logical extensions of Bayesian networks. Let us now briefly survey some of the key representational frameworks that have been developed around these pri   </text>
<query_num> 9222 </query_num>
<text>   ion rule. Finally, there is the quite popular formalism of probabilistic relational models 3 (PRMs) [57; 81; 31; 33]. This formalism 3 Several other extensions of Bayesian network have been developed =-=[60; 58; 61; 80; 81]-=- by Pfeffer et al. before introducing probabilistic relational models. These extensions have been termed frame-based and object-oriented, in which the relational and logical issues are less direct. Th   </text>
<query_num> 9223 </query_num>
<text>   ional clustering [98], hypertext classification [37], selectivity estimation within databases [38], and modelling identity uncertainty [76], and led to impressive results within computational biology =-=[95; 93; 94]-=-. Recently, Getoor et al. [31; 35; 32] have introduced stochastic relational models (SRMs). At the syntactic level SRMs are almost identical to PRMs but SRMs possess a different semantics, which allow   </text>
<query_num> 9224 </query_num>
<text>   m probabilistic in our context refers to the use of probabilistic representations and reasoning mechanisms grounded in probability theory, such as Bayesian networks [78; 10; 47], hidden Markov models =-=[85; 84; 22]-=- and stochastic grammars [22; 64]. Such representations have been successfully used across a wide range of applications and have resulted in a number of robust models for reasoning about uncertainty.  r, and revolutionary novel products in computer vision, speech recognition, medical diagnostics, troubleshooting systems, etc. Overviews of and introductions to probabilistic learning can be found in =-=[20; 84; 44; 8; 48; 22; 64]-=-. - Inductive Logic Programming and multi-relational data mining studied logic learning, i.e. learning and data mining within first order logical or relational representations. Inductive Logic Program c grammars, most notably of (hidden) Markov models [84], which are essentially stochastic regular grammars. These are e.g. used to determine how likely it is that a given protein belongs to some fold =-=[22]-=-. From a computational and logical perspective, the key limitation of stochastic context free grammars concerns their expressive power. Indeed, context free grammars (and therefore stochastic context   </text>
<query_num> 9225 </query_num>
<text>   ntersection of probabilistic logic learning, cf. Figure 1. Consequently, we will not discuss interesting work lying on boundaries such as Flach and Lachiche’s work on Naïve Bayes for structured terms =-=[27; 62]-=- and Craven and Slattery’s work on combining Naïve Bayes with a relational rule learner [11]. Giving an overview of the pairwise intersections (let alone the underlying domains) would lead us too far   </text>
<query_num> 9226 </query_num>
<text>   pretations, cf. [82; 41; 73; 45; 57; 81; 31; 52]. Important predecessors of these logical Bayesian networks discussed below, include the work by Breese, Charniak, Bacchus, Goldman, Poole and Wellmann =-=[40; 6; 4; 7; 39]-=-. These predecessors were largely based on the idea of knowledge based model construction (KBMC). In knowledge based model construction, a knowledge base (sometimes in the form of an annotated logic p   </text>
<query_num> 9227 </query_num>
<text>   r, and revolutionary novel products in computer vision, speech recognition, medical diagnostics, troubleshooting systems, etc. Overviews of and introductions to probabilistic learning can be found in =-=[20; 84; 44; 8; 48; 22; 64]-=-. - Inductive Logic Programming and multi-relational data mining studied logic learning, i.e. learning and data mining within first order logical or relational representations. Inductive Logic Program   </text>
<query_num> 9228 </query_num>
<text>   r, and revolutionary novel products in computer vision, speech recognition, medical diagnostics, troubleshooting systems, etc. Overviews of and introductions to probabilistic learning can be found in =-=[20; 84; 44; 8; 48; 22; 64]-=-. - Inductive Logic Programming and multi-relational data mining studied logic learning, i.e. learning and data mining within first order logical or relational representations. Inductive Logic Program ining each data cases. Then, the parameters of the induced Bayesian network are estimated using standard Bayesian network parameter estimation methods, such as those discussed in Heckerman’s tutorial =-=[44]-=-. For instance the examples in Section 5.1.1 together with the genetic program would induce the Bayesian network shown in Figure 9. All nodes of the Bayesian network, which do not occur in the evidenc  applied. In one step, atoms are either added to or deleted from clauses. Furthermore, there is exactly one clause for each of proposition. Special care is taken that the resulting network is acyclic =-=[44]-=-. The state-of-the-art 10 At this point, note that we employ a simplified and idealized refinement operator. Some essential features, such as substitutions are not taken into account for reasons of si   </text>
<query_num> 9229 </query_num>
<text>   racterization of logic than suitable representations for learning. Finally, in the past few years there have been some important attempts that address the full problem of probabilistic logic learning =-=[59; 30; 67; 31; 14; 53; 51; 92; 68]-=-. The goal of this paper is to provide an introduction and overview to these works. In doing so, we restrict our attention to those developments that address the intersection of probabilistic logic le  the key ideas and will not address any advanced issues such as efficiency concerns. 5.2.1 Model Theoretic Parameter estimations for probabilistic-logic programs [59], probabilistic relational models =-=[30; 31; 33]-=-, and Bayesian logic programs [51; 54] all follow the same principle: The given data and the current model induce a Bayesian network explaining each data cases. Then, the parameters of the induced Bay erent aggregate functions and where there is at most one clause defining a predicate (such as p/1). Structure learning with probabilistic relational models now occurs by refining this type of clauses =-=[30; 36; 34; 33; 31]-=- 11 . Aggregated literals such as τ1(q(X, Y)) can be added to or deleted from clauses. At this point, the graphical notation for probabilistic relational models (illustrated in Figure 10) is convenien   </text>
<query_num> 9230 </query_num>
<text>   racterization of logic than suitable representations for learning. Finally, in the past few years there have been some important attempts that address the full problem of probabilistic logic learning =-=[59; 30; 67; 31; 14; 53; 51; 92; 68]-=-. The goal of this paper is to provide an introduction and overview to these works. In doing so, we restrict our attention to those developments that address the intersection of probabilistic logic le , the improved estimates for each clause are obtained by dividing the clause’s expected counts by the sum of the expected counts of clauses for the same predicate. The EM algorithm for PRISM programs =-=[50; 91; 49; 92]-=- is similar in spirit. The main differences are that (1) no probability values are associated to (intensional) clauses, (2) failed derivations have a probability of zero, and (3) the disjoint represen   </text>
<query_num> 9231 </query_num>
<text>   racterization of logic than suitable representations for learning. Finally, in the past few years there have been some important attempts that address the full problem of probabilistic logic learning =-=[59; 30; 67; 31; 14; 53; 51; 92; 68]-=-. The goal of this paper is to provide an introduction and overview to these works. In doing so, we restrict our attention to those developments that address the intersection of probabilistic logic le at BLPs employ a minimal set of primitives such that BLPs have both definite clause logic (i.e. pure Prolog) and Bayesian networks as a direct special case. This in turn facilitates the learning, cf. =-=[53; 51; 54]-=-. Bayesian logic programs again view the atoms as the random variables. More precisely, the atoms of the (least) Herbrand model constitute the set of random variables of the induced Bayesian network.  e probabilistic relational model will be acyclic. Bayesian logic programs are represented by regular clauses, on which the typical refinement operators from inductive logic programming can be applied =-=[53; 54]-=-. However, in Bayesian logic programs, one must take into account the covers constraint. More formally, it is required that the examples are models of the Bayesian logic programs, i.e. cover(H, e) = t   </text>
<query_num> 9232 </query_num>
<text>   racterization of logic than suitable representations for learning. Finally, in the past few years there have been some important attempts that address the full problem of probabilistic logic learning =-=[59; 30; 67; 31; 14; 53; 51; 92; 68]-=-. The goal of this paper is to provide an introduction and overview to these works. In doing so, we restrict our attention to those developments that address the intersection of probabilistic logic le at BLPs employ a minimal set of primitives such that BLPs have both definite clause logic (i.e. pure Prolog) and Bayesian networks as a direct special case. This in turn facilitates the learning, cf. =-=[53; 51; 54]-=-. Bayesian logic programs again view the atoms as the random variables. More precisely, the atoms of the (least) Herbrand model constitute the set of random variables of the induced Bayesian network.  vanced issues such as efficiency concerns. 5.2.1 Model Theoretic Parameter estimations for probabilistic-logic programs [59], probabilistic relational models [30; 31; 33], and Bayesian logic programs =-=[51; 54]-=- all follow the same principle: The given data and the current model induce a Bayesian network explaining each data cases. Then, the parameters of the induced Bayesian network are estimated using stan n the evidence but are introduced by the program, 7 Some approaches employ gradient-based approaches instead of EM, e.g., Muggleton [67; 68] for structural learning of SLPs, and Kersting and De Raedt =-=[51; 54]-=- developed both EM-based and gradient-based approaches for parameter estimation of BLPs. 8 This is only the underlying idea of the EM. More formally, the E-step consists of computing the expectation o arantee this for PRMs as the aggregated examples constitute the states of the nodes in the induced Bayesian network. Within PLPs and BLPs, uniqueness is enforced by using decomposable combining rules =-=[59; 51]-=-. The effect of a decomposable combining rule can be represented using extra nodes in the induced Bayesian network. Most combining rules commonly employed in Bayesian networks such as noisy or, noisy   </text>
<query_num> 9233 </query_num>
<text>   racterization of logic than suitable representations for learning. Finally, in the past few years there have been some important attempts that address the full problem of probabilistic logic learning =-=[59; 30; 67; 31; 14; 53; 51; 92; 68]-=-. The goal of this paper is to provide an introduction and overview to these works. In doing so, we restrict our attention to those developments that address the intersection of probabilistic logic le in Figure 9. All nodes of the Bayesian network, which do not occur in the evidence but are introduced by the program, 7 Some approaches employ gradient-based approaches instead of EM, e.g., Muggleton =-=[67; 68]-=- for structural learning of SLPs, and Kersting and De Raedt [51; 54] developed both EM-based and gradient-based approaches for parameter estimation of BLPs. 8 This is only the underlying idea of the E the related PRISMs) has not yet been addressed. So far, the only contributions to structure learning of stochastic logic programs are restricted to learning missing clauses for a single predicate. In =-=[70; 67]-=-, Muggleton introduced a two-phase approach that separates the structure learning aspects from the parameter estimation phase. In a more recent approach [68] Muggleton presents an initial attempt to i   </text>
<query_num> 9234 </query_num>
<text>   racterization of logic than suitable representations for learning. Finally, in the past few years there have been some important attempts that address the full problem of probabilistic logic learning =-=[59; 30; 67; 31; 14; 53; 51; 92; 68]-=-. The goal of this paper is to provide an introduction and overview to these works. In doing so, we restrict our attention to those developments that address the intersection of probabilistic logic le mc/2 is the sum of probabilities of all refutations of the most gen4 More general definitions of stochastic logic programs where e.g. the probability values do not sum to 1.0 are discussed by Cussens =-=[12; 14]-=-. Because learning of stochastic logic programs is considered only for ‘normalized’ programs we restrict our attention to them. SIGKDD Explorations. Volume 2, Issue 2 - page 8seral goal over mc(X, Y), Whereas model-based approaches complete the data in terms of a Bayesian network, proof-theoretic approaches complete the data based on refutations and failures. For stochastic logic programs, Cussens =-=[14]-=- introduced the failure-adjusted maximization (FAM) algorithm. In FAM, the logical part of the SLP is fixed and given, and the parameters have to be learned on the basis of the evidence. The evidence   </text>
<query_num> 9235 </query_num>
<text>   racterization of logic than suitable representations for learning. Finally, in the past few years there have been some important attempts that address the full problem of probabilistic logic learning =-=[59; 30; 67; 31; 14; 53; 51; 92; 68]-=-. The goal of this paper is to provide an introduction and overview to these works. In doing so, we restrict our attention to those developments that address the intersection of probabilistic logic le sake of simplicity, we will only state the key ideas and will not address any advanced issues such as efficiency concerns. 5.2.1 Model Theoretic Parameter estimations for probabilistic-logic programs =-=[59]-=-, probabilistic relational models [30; 31; 33], and Bayesian logic programs [51; 54] all follow the same principle: The given data and the current model induce a Bayesian network explaining each data  arantee this for PRMs as the aggregated examples constitute the states of the nodes in the induced Bayesian network. Within PLPs and BLPs, uniqueness is enforced by using decomposable combining rules =-=[59; 51]-=-. The effect of a decomposable combining rule can be represented using extra nodes in the induced Bayesian network. Most combining rules commonly employed in Bayesian networks such as noisy or, noisy   </text>
<query_num> 9236 </query_num>
<text>   regate functions to deal with the multiple instantiations problem and the use of a purely relational language, which does not allow for functors (needed for representing infinite structures), but see =-=[79]-=-. PRMs have been successfully applied to various problems such as relational clustering [98], hypertext classification [37], selectivity estimation within databases [38], and modelling identity uncert   </text>
<query_num> 9237 </query_num>
<text>   s volume and in [69; 23]. - Early work on incorporating probabilities into logic programs was done by Clark and McCabe [9] and by Shapiro [97]. Subsequently, researchers such as Nilsson [75], Halpern =-=[43]-=-, Ng and Subrahmanian [71], and Poole [82] have studied probabilistic logics from a knowledge representational perspective. The aim of this research is more a probabilistic characterization of logic t   </text>
<query_num> 9238 </query_num>
<text>   sent some of the best-known examples of scientific discovery by AI systems in the literature. Overviews of inductive logic learning and multi-relational data mining can be found in this volume and in =-=[69; 23]-=-. - Early work on incorporating probabilities into logic programs was done by Clark and McCabe [9] and by Shapiro [97]. Subsequently, researchers such as Nilsson [75], Halpern [43], Ng and Subrahmania Furthermore, these examples e should be logically entailed by the target program T , i.e., T |= e. This is akin to the traditional learning from entailment setting in inductive logic programming, cf. =-=[69]-=-. 5.1.3 Intermediate Representations Intermediate representations, such as (hidden) Markov models, are typically learned from examples that correspond to sequences of states that arose in particular s learning as well as in theory revision approaches [100] to inductive logic programming. Background knowledge is typically employed in multi-relational data mining and inductive logic programming, cf. =-=[69]-=-. Further distinctions could be made according to whether the data are given initially and fixed or whether they are gathered incrementally. Such distinctions and enhancements are quite similar to tho -the-art 10 At this point, note that we employ a simplified and idealized refinement operator. Some essential features, such as substitutions are not taken into account for reasons of simplicity, cf. =-=[69; 74]-=-. SIGKDD Explorations. Volume 2, Issue 2 - page 13slearning algorithm is the structural EM (SEM) [29]. It adapts the standard EM algorithm for structure learning. The key idea is that the expected cou   </text>
<query_num> 9239 </query_num>
<text>   t it would be impossible to specify that the probability is 0.70 that Jef is the father of Mary. Nevertheless, some partial solutions for dealing with probabilistic relations have been developed, cf. =-=[34; 31]-=-. The genetic example shows that – at the logical level – PRMs essentially define via single clauses probabilistic dependencies between the attributes of various entities. It also shows that probabili erent aggregate functions and where there is at most one clause defining a predicate (such as p/1). Structure learning with probabilistic relational models now occurs by refining this type of clauses =-=[30; 36; 34; 33; 31]-=- 11 . Aggregated literals such as τ1(q(X, Y)) can be added to or deleted from clauses. At this point, the graphical notation for probabilistic relational models (illustrated in Figure 10) is convenien   </text>
<query_num> 9240 </query_num>
<text>   the related PRISMs) has not yet been addressed. So far, the only contributions to structure learning of stochastic logic programs are restricted to learning missing clauses for a single predicate. In =-=[70; 67]-=-, Muggleton introduced a two-phase approach that separates the structure learning aspects from the parameter estimation phase. In a more recent approach [68] Muggleton presents an initial attempt to i   </text>
<query_num> 9241 </query_num>
<text>   tial features, such as substitutions are not taken into account for reasons of simplicity, cf. [69; 74]. SIGKDD Explorations. Volume 2, Issue 2 - page 13slearning algorithm is the structural EM (SEM) =-=[29]-=-. It adapts the standard EM algorithm for structure learning. The key idea is that the expected counts are not computed anew for every structure that is proposed, but only after several iterations. Th   </text>
</top>
<top>
<paper_num> 93 </paper_num>
<paper_title>   Algorithmic Mechanism Design.  </paper_title>
<paper_abstract>   Distributed algorithmic mechanism design (DAMD) is an approach to designing distributed systems that takes into account both the distributed-computational environment and the incentives of autonomous agents. In this dissertation, we study two problems, multicast cost sharing and interdomain routing. We also touch upon several issues important to DAMD in general, including approximation, compatibility with existing protocols, and hardness that results from the interplay of incentives and distributed computation.  </paper_abstract>
<query_num> 9301 </query_num>
<text>   . The multicast cost-sharing problem is a mechanism-design problem that was first in3stroduced by Feigenbaum et al. [FPS01] and has since been studied by many researchers, including Jain and Vazirani =-=[JV01]-=-, Adler and Rubenstein [AR02], Fiat et al. [FGHK02], and Mitchell and Teague [MT02]. The problem is described in detail in Chapter 3; here, we only outline it in brief. The problem is as follows: Ther o satisfy the SYM property. The egalitarian (EG) mechanism of Dutta and Ray [DR89] is another well studied GSP, budget-balanced mechanism that satisfies the four basic requirements. Jain and Vazirani =-=[JV01]-=- present a novel family of GSP, approximately budget-balanced mechanisms 3 that satisfy NPT, VP, and CS. Each mechanism in the family is defined by its underlying cost-sharing function, and the result  shows that, for a restricted class of algorithms (called “linear distributed algorithms”), there is an infinite set of instances with |P | = O(|N|) that require Ω(|N| 2 ) messages. Jain and Vazirani =-=[JV01]-=- give centralized, polynomial-time algorithms to compute the approximately budget-balanced mechanisms in the class JV. 3.5 Our Results In this chapter, we show that: • Any distributed algorithm, deter G mechanisms. • Any distributed algorithm, deterministic or randomized, that computes an approximately budget-balanced, GSP multicast cost-sharing mechanism must send Ω(log(|P |)) 3 The mechanisms in =-=[JV01]-=- actually satisfy a more stringent definition of approximate budget balance than we use; thus, our network-complexity lower bounds apply to them a fortiori. 30sbits over linearly many links in the wor nd so we need to compute the new cost shares x 1 i = gi(R 1 ) and iterate. This process ultimately converges, terminating with the receiver set R(u) and the final cost shares xi(u). Jain and Vazirani =-=[JV01] g-=-ive a geometric characterization of the space of cross-monotonic mechanisms for a given submodular cost-sharing problem. A mechanism Mg can be represented as a point in Rn2n−1 as follows: For each S   </text>
<query_num> 9302 </query_num>
<text>   d new light on the general issue of approximation in algorithmic mechanism design. It is known that approximating a mechanism’s output can destroy its strategic properties (see, e.g., Nisan and Ronen =-=[NR00]-=-). Our results show that, in some circumstances, it may be possible to preserve the strategic properties of a mechanism while trading off accuracy and communication complexity. The multicast cost-shar  NP-hard combinatorial optimization problem. This naturally leads us to ask whether an approximation algorithm can be used to yield a mechanism that is approximately optimal. However, Nisan and Ronen =-=[NR00]-=- showed that this is not as straightforward as it may appear: Replacing an exact solution by an approximate solution may destroy the game-theoretic properties (e.g., strategyproofness) of the mechanis wer bound on the network complexity of approximately budget-balanced, GSP mechanisms. First, we recall the definition of an “approximately budget-balanced” mechanism. As explained at length in, e.g., =-=[NR00]-=-, one cannot define an approximation of a cost-sharing mechanism (σ, x) simply as a pair (σ ′ , x ′ ) such that σ ′ and x ′ approximate σ and x, respectively, as functions. Such an approach may destro al other interesting notions of approximation have been put forth – see Section 5 of [FS02] for an overview. Here we mention only the work that is most closely related to the results in this chapter. =-=[NR00]-=- were the first to address the question of approximate computation in algorithmic mechanism design. They considered VCG mechanisms in which optimal outcomes are NP-hard to compute (as they are in comb he above discussion of how we should define “approximating the SH mechanism” and why approximating the pair of functions (σ, x) is not sufficient is based on the analogous observation in our context. =-=[NR00]-=- approach this problem by developing a notion of “feasible” strategyproofness and describing a broad class of situations in which NP-hard VCG mechanisms have feasibly strategyproof approximations. Thi   </text>
<query_num> 9303 </query_num>
<text>   design theory in the book by Mas-Colell, Whinston, and Green [MWG95, Ch. 13], and to the survey article by Jackson [Jac01]. A recent survey of distributed algorithmic mechanism design can be found in =-=[FS02]-=-. 2.1 Mechanism Design Framework Figure 2.1 depicts a simple mechanism setting: There are n agents; each agent i has some private information, called her private type ti. This type is drawn from a set lly faithful approximation. Approximation is an increasingly active area of algorithmic mechanism design, and several other interesting notions of approximation have been put forth – see Section 5 of =-=[FS02]-=- for an overview. Here we mention only the work that is most closely related to the results in this chapter. [NR00] were the first to address the question of approximate computation in algorithmic mec   </text>
<query_num> 9304 </query_num>
<text>   e Internet, we are seeing the design and deployment of large-scale distributed systems. The deployed systems include large-scale parallel computing projects such as SETI@home [ACK + 02] and factoring =-=[LM90]-=-, Internet services such as file-sharing and caching, and systems to support the basic network functionality, such as routing and congestion control. Usually, the primary focus of research on these di   </text>
<query_num> 9305 </query_num>
<text>   econd formulation, the policy routing mechanism-design problem, in which † Sections 4.2 to 4.7 describe joint work with Joan Feigenbaum, Christos Papadimitriou, and Scott Shenker, and was reported in =-=[FPSS02]-=-. Ramesh Govindan provided us with a recent AS graph, and taught us about some of the intricacies of BGP. We also thank Kunal Talwar for helpful discussions on the role of incentives in AS-graph forma   </text>
<query_num> 9306 </query_num>
<text>   ishes this. Our results contribute in several ways to the understanding of how incentives and computation affect each other in routing-protocol design. Nisan and Ronen [NR01] and Hershberger and Suri =-=[HS01]-=- considered the LCP mechanism-design problem, motivated in part by the desire to include incentive issues in Internet-route selection. The LCP mechanism studied in [NR01, HS01] takes as input a biconn s to the Vickrey-Clarke-Groves (VCG) family [Vic61, Cla71, Gro73]. It is in essence a node-centric, all-pairs extension of the LCP mechanism studied by Nisan and Ronen [NR01] and Hershberger and Suri =-=[HS01]-=-. There are several aspects of this result that are worth noting. First, although the payments could have taken any form and could have depended arbitrarily on the traffic matrix, it turns out the pay   </text>
<query_num> 9307 </query_num>
<text>   lues; this means that the hardness holds over a region of preference space with non-zero volume, instead of at isolated points. This is similar in spirit to the smoothed analysis of Spielman and Teng =-=[ST01]-=-. First, we observe that the LCP-mechanism satisfies these properties, provided the costs are similar to each other, not very skewed. We follow the notation in Section 4.6.3, using d to denote the len   </text>
<query_num> 9308 </query_num>
<text>   nsionality that has been developed in the economics literature [MR74, Wal77]. The parallels between the two approaches are demonstrated by Nisan and Segal, in their analysis of combinatorial auctions =-=[NS03]-=-. However, the concepts are different in that the message-space dimensionality only deals with the informational requirements at equilibrium, whereas the communication complexity of computing the mech   </text>
<query_num> 9309 </query_num>
<text>   pproximate minimum worst-case welfare loss. Finally, “competitive-ratio” analysis (a form of approximation) has been studied for a variety of strategyproof auctions (see, e.g., [FGHK02], [GHW01], and =-=[LN00]-=-). 3.11 Towards approximating the SH mechanism In this section, we develop a GSP mechanism that exhibits a trade-off between the other properties of the Shapley value: It can be computed by an algorit   </text>
<query_num> 9310 </query_num>
<text>   problem is a mechanism-design problem that was first in3stroduced by Feigenbaum et al. [FPS01] and has since been studied by many researchers, including Jain and Vazirani [JV01], Adler and Rubenstein =-=[AR02]-=-, Fiat et al. [FGHK02], and Mitchell and Teague [MT02]. The problem is described in detail in Chapter 3; here, we only outline it in brief. The problem is as follows: There is some digital content (sa   </text>
<query_num> 9311 </query_num>
<text>   ributed computational model would be more appropriate for the Internet, where agents are scattered across a network, and communication may be expensive or slow. Feigenbaum, Papadimitriou, and Shenker =-=[FPS01]-=- extended the Nisan-Ronen framework to include distributed computation, in order to encompass systems in which the agents are distributed across a network, and it is impractical to collect all the inp  is focused on two specific problems: multicast cost sharing and interdomain routing. The multicast cost-sharing problem is a mechanism-design problem that was first in3stroduced by Feigenbaum et al. =-=[FPS01]-=- and has since been studied by many researchers, including Jain and Vazirani [JV01], Adler and Rubenstein [AR02], Fiat et al. [FGHK02], and Mitchell and Teague [MT02]. The problem is described in deta  can collude to manipulate the mechanism to their advantage. (In addition, we assume that the mechanism satisfies certain other natural properties, which are detailed in Chapter 3.) Feigenbaum et al. =-=[FPS01]-=- studied one particularly attractive mechanism of this class, the Shapley-value mechanism (SH). They proved a lower bound on the communication required to compute this mechanism for a restricted class nd it may cause congestion near the centralized server. This led Feigenbaum, Papadimitriou, and Shenker to consider distributed computational models in their paper on multicast costsharing mechanisms =-=[FPS01]-=-; this started the study of distributed algorithmic mechanism design. Feigenbaum et al. pointed out that for a mechanism to be feasible in an Internet setting, it must be computable by a distributed a HK02, MT02] adopt the distributed algorithmic mechanism design approach, which augments a game-theoretic perspective with distributed computational concerns. In this chapter, we extend the results of =-=[FPS01]-=- by considering a more general computational model and approximate solutions. We also extend a classic impossibility [GL79] result by showing that no strategyproof mechanism can be both approximately   </text>
<query_num> 9312 </query_num>
<text>   rts of it were reported in [FKSS03, AFK + 03]. Aaron Archer provided valuable comments. 23sproblem of cost sharing for Internet multicast transmissions. In the first paper on the topic, Herzog et al. =-=[HSE97]-=- considered axiomatic and implementation aspects of the problem. Subsequently, Moulin and Shenker [MS01] studied the problem from a purely economic point of view. Several more recent papers [FPS01, AR   </text>
<query_num> 9313 </query_num>
<text>   sm-design problem that was first in3stroduced by Feigenbaum et al. [FPS01] and has since been studied by many researchers, including Jain and Vazirani [JV01], Adler and Rubenstein [AR02], Fiat et al. =-=[FGHK02]-=-, and Mitchell and Teague [MT02]. The problem is described in detail in Chapter 3; here, we only outline it in brief. The problem is as follows: There is some digital content (say, a movie) at one nod ned by a monopoly network operator, then one might expect the goal to be maximizing revenue. There have been some recent investigations of revenuemaximizing charging schemes for multicast (see, e.g., =-=[FGHK02]-=-), but here we assume, as in [HSE97, MS01, FPS01, AR02], that the charging mechanism is decided by society at large (e.g., through standards bodies) or through competition. Competing network providers are not attempting to approximate minimum worst-case welfare loss. Finally, “competitive-ratio” analysis (a form of approximation) has been studied for a variety of strategyproof auctions (see, e.g., =-=[FGHK02]-=-, [GHW01], and [LN00]). 3.11 Towards approximating the SH mechanism In this section, we develop a GSP mechanism that exhibits a trade-off between the other properties of the Shapley value: It can be c   </text>
<query_num> 9314 </query_num>
<text>   stributed algorithm that accomplishes this. Our results contribute in several ways to the understanding of how incentives and computation affect each other in routing-protocol design. Nisan and Ronen =-=[NR01]-=- and Hershberger and Suri [HS01] considered the LCP mechanism-design problem, motivated in part by the desire to include incentive issues in Internet-route selection. The LCP mechanism studied in [NR0 j)cr . r∈N This mechanism belongs to the Vickrey-Clarke-Groves (VCG) family [Vic61, Cla71, Gro73]. It is in essence a node-centric, all-pairs extension of the LCP mechanism studied by Nisan and Ronen =-=[NR01]-=- and Hershberger and Suri [HS01]. There are several aspects of this result that are worth noting. First, although the payments could have taken any form and could have depended arbitrarily on the traf   </text>
<query_num> 9315 </query_num>
<text>   tempting to approximate minimum worst-case welfare loss. Finally, “competitive-ratio” analysis (a form of approximation) has been studied for a variety of strategyproof auctions (see, e.g., [FGHK02], =-=[GHW01]-=-, and [LN00]). 3.11 Towards approximating the SH mechanism In this section, we develop a GSP mechanism that exhibits a trade-off between the other properties of the Shapley value: It can be computed b   </text>
<query_num> 9316 </query_num>
<text>   the current BGP algorithm, which is the repository of interdomain routing information, as the computational substrate. We adopt the abstract model of the BGP protocol described by Griffin and Wilfong =-=[GW99]-=-, which involves several simplifying assumptions. Specifically, we assume that there is at most one link between any two ASes, that the links are bidirectional, and that each AS can be treated as an a   </text>
<query_num> 9317 </query_num>
<text>   the literature in this vast field; we refer the readers to the chapter on mechanism design theory in the book by Mas-Colell, Whinston, and Green [MWG95, Ch. 13], and to the survey article by Jackson =-=[Jac01]-=-. A recent survey of distributed algorithmic mechanism design can be found in [FS02]. 2.1 Mechanism Design Framework Figure 2.1 depicts a simple mechanism setting: There are n agents; each agent i has t appears to be too stringent a requirement for many Internet mechanisms. Attempts to get around this problem of multiple equilibria have been made in the literature on implementation theory; Jackson =-=[Jac01]-=- surveys the literature in this field. There has also been a lot of research on refinements of Nash equilibria: If we make additional assumptions about which strategies are “reasonable” for an agent t   </text>
</top>
<top>
<paper_num> 94 </paper_num>
<paper_title>   Privacy-Preserving Top-K Queries.  </paper_title>
<paper_abstract>   Processing of top-k queries has been attracting considerable attention. Much of the work assumes distributed data, with each site holding a different set of attributes for the same set of entities. These methods assume that all sites are happy with revealing the local order or scores of the entities. Privacy/security concerns faced by the individual sites can prohibit such disclosure. We present a mechanism through which only the result of the top-k query (i.e., the top-k entities) is returned while minimizing disclosure of other information. 1.  </paper_abstract>
<query_num> 9401 </query_num>
<text>   F | rounds the correct value is identified. Algorithm 2 gives the detailed algorithm. 3.2 Set Union Secure Union has also been solved; we sketch the method from [16] (another solution can be found in =-=[3]-=-.) The problem can be stated as follows: There exist p sites P1, . . . , Pp. Party Pi has set Si where the elements of the set come from a common global universe G. Together they want to compute the u ion algorithms also exist, some of which may be more efficient in certain cases [2, 12]. Depending on the protocol used for set union, detailed complexity analysis can be found in the relevant papers =-=[16, 3]-=-. For now, we just give a rough estimate based on the protocol in [16]. Effectively, the set union protocol simply requires the same number of encryptions as the set intersection protocol as well as r   </text>
<query_num> 9402 </query_num>
<text>   This paper draws heavily on work in privacy-preserving data mining[17] and secure multiparty computation[13]. Of particular relevance is a method for determining the kth element in a distributed list=-=[1]-=-; this can be used for privacypreserving top-k queries in horizontally partitioned data. A key contribution of this paper is to demonstrate how components developed in privacy-preserving data mining a here are two parties A and B having n elements (a1, . . . , an and b1, . . . , bn respectively) such that ∀i, ai + bi = di (mod F). Aggarwal et. al solved the problem for horizontally partitioned data=-=[1]-=-, i.e., two parties A and B have na and nb items (a1, . . . , ana and b1, . . . , bnb respectively). Our method shares the same basic idea: a binary search for an appropriate threshold score. Use of s   </text>
<query_num> 9403 </query_num>
<text>   e candidates to identify the top-K results. This requires another x secure comparisons. The secure comparison [21] requires O(size) oblivious transfers. More efficient comparison protocols also exist =-=[11, 15]-=-. Ioannidis et. al [15] report that the computation cost for comparing two number of size 15 bits is 217.5ms on a Pentium III/450 MHz machine. The communication overhead is &amp;lt; 4ms. Using this data we c   </text>
<query_num> 9404 </query_num>
<text>   e grown until they have k items in common, then the union of the lists is guaranteed to contain the top-k. He used this observation to develop an algorithm which can work with O(k) communication cost =-=[7]-=-, although worst case could be O(n). 1 Top-k queries are particularly relevant when dealing with privacy-sensitive information. For example, profiling has potential anti-terrorism applications (e.g.,  o based on this assumption, so we are not adding any extra constraints on to the problem. There has been a lot of work on the top-k query/selection problem since Fagin first proposed his A0 algorithm =-=[7]-=-. Fagin also proposed other algorithms (TA and NRA) [10] that perform better than his original algorithm. However, because of its amenability to secure evaluation we start with the original A0 algorit   </text>
<query_num> 9405 </query_num>
<text>   ess to a separate database that give information about different features. Data is collected for the same set of entities. This is the data model assumed in several distributed best-match query papers=-=[9, 8, 5, 4]-=-. We include the added twist that none of the parties trust each other completely; privacy concerns prevent them from disclosing scores or local ordering of the entities. Our data distribution model a   </text>
<query_num> 9406 </query_num>
<text>   ess to a separate database that give information about different features. Data is collected for the same set of entities. This is the data model assumed in several distributed best-match query papers=-=[9, 8, 5, 4]-=-. We include the added twist that none of the parties trust each other completely; privacy concerns prevent them from disclosing scores or local ordering of the entities. Our data distribution model a ation we start with the original A0 algorithm. There has also been work on top-k queries that assumes all of the attributes of every object are readily available to the query processor [5, 6, 14]. In =-=[4]-=-, the authors introduce an algorithm for evaluating top-K selection queries over web accessible databases. This paper draws heavily on work in privacy-preserving data mining[17] and secure multiparty   </text>
<query_num> 9407 </query_num>
<text>   result, a Secure Multiparty Computation proof is possible. Refer to [16] for details. 3.3 Size of Set Intersection There are several solutions to securely finding the size of the intersection of sets=-=[20, 2, 12]-=-, we sketch the method of [20]. Consider several parties having their own sets of items from a common domain. The problem is to securely compute the cardinality/size of the intersection of these local ere are O(p 2 ) messages, O(p 2 x) bits and p rounds. Actual experimental cost can be found in [20]. Other set intersection algorithms also exist, some of which may be more efficient in certain cases =-=[2, 12]-=-. Depending on the protocol used for set union, detailed complexity analysis can be found in the relevant papers [16, 3]. For now, we just give a rough estimate based on the protocol in [16]. Effectiv   </text>
<query_num> 9408 </query_num>
<text>   result, a Secure Multiparty Computation proof is possible. Refer to [16] for details. 3.3 Size of Set Intersection There are several solutions to securely finding the size of the intersection of sets=-=[20, 2, 12]-=-, we sketch the method of [20]. Consider several parties having their own sets of items from a common domain. The problem is to securely compute the cardinality/size of the intersection of these local y party. None of the parties is able to know which of the items are present in the intersection set because of the encryption. The complete protocol is shown in algorithm 4. For details and proof see =-=[20]-=-. 4. Security Analysis The Secure Multiparty Computation community has developed a method for demonstrating that protocols do not leak information known as proof by simulation. The key idea is that if mber of candidates. Obviously this is an upper bound on the size of the candidate list generated by each site. A single invocation of the size of set intersection protocol will have the following cost=-=[20]-=-: For p parties and maximum set size x, there are p 2 ∗ x encryptions. Parallelization is possible and the real computation time cost is that of p ∗ x encryptions. In terms of communication cost, ther   </text>
<query_num> 9409 </query_num>
<text>   ted to be lower or higher. Within log |F | rounds the correct value is identified. Algorithm 2 gives the detailed algorithm. 3.2 Set Union Secure Union has also been solved; we sketch the method from =-=[16]-=- (another solution can be found in [3].) The problem can be stated as follows: There exist p sites P1, . . . , Pp. Party Pi has set Si where the elements of the set come from a common global universe  ize of intersections and the final result, nothing is revealed. By assuming that the count of duplicated items is part of the final result, a Secure Multiparty Computation proof is possible. Refer to =-=[16]-=- for details. 3.3 Size of Set Intersection There are several solutions to securely finding the size of the intersection of sets[20, 2, 12], we sketch the method of [20]. Consider several parties havin ion algorithms also exist, some of which may be more efficient in certain cases [2, 12]. Depending on the protocol used for set union, detailed complexity analysis can be found in the relevant papers =-=[16, 3]-=-. For now, we just give a rough estimate based on the protocol in [16]. Effectively, the set union protocol simply requires the same number of encryptions as the set intersection protocol as well as r   </text>
<query_num> 9410 </query_num>
<text>   the best k matches to the query, we can maintain k-anonymity while meeting the profiling goals. This paper presents an algorithm to solve this problem. The basic idea is to replicate Fagin’s algorithm=-=[9]-=-, while limiting disclosure to that inherently required to meet the efficiency goals of the algorithm. For example, if a site is required to test if an entity that is not locally close is in the top-k ess to a separate database that give information about different features. Data is collected for the same set of entities. This is the data model assumed in several distributed best-match query papers=-=[9, 8, 5, 4]-=-. We include the added twist that none of the parties trust each other completely; privacy concerns prevent them from disclosing scores or local ordering of the entities. Our data distribution model a t enough; they most be cleverly composed in ways that achieve both efficiency and enable proof that the complete algorithm is secure. 3. Algorithm We follow the basic structure of Fagin’s A0 algorithm=-=[9]-=-. We first give a brief description of Fagin’s algorithm (adapted from [9]) and then discuss our modifications. Please refer to the original paper [9] for more details. The A0 algorithm consists of 3  es). Thus, if we represent the local set at site i as Oi, | � i Oi| ≥ k. Under a monotone grading function, it is shown that the top-k objects are guaranteed to be within the union of all the objects =-=[9]-=-. Random access phase: For each object seen in the sorted access phase (i.e., ∀o ∈ � i Oi), random access is done to each subsystem to find the local score of that object. Computation phase: Compute t ems are accessed in the random access phase would reveal this information, so to prevent disclosing the candidates each site would need to access every item (losing the efficiency of the algorithm of =-=[9]-=-.) The information disclosed is discussed further in Section 4.1. The algorithm does place some restrictions on the scoring functions to ensure that they can be easily computedsin a distributed and se   </text>
<query_num> 9411 </query_num>
<text>   to secure evaluation we start with the original A0 algorithm. There has also been work on top-k queries that assumes all of the attributes of every object are readily available to the query processor =-=[5, 6, 14]-=-. In [4], the authors introduce an algorithm for evaluating top-K selection queries over web accessible databases. This paper draws heavily on work in privacy-preserving data mining[17] and secure mul   </text>
</top>
<top>
<paper_num> 95 </paper_num>
<paper_title>   Probability Product Kernels.  </paper_title>
<paper_abstract>   The advantages of discriminative learning algorithms and kernel machines are combined with generative modeling using a novel kernel between distributions. In the probability product kernel, data points in the input space are mapped to distributions over the sample space and a general inner product is then evaluated as the integral of the product of pairs of distributions. The kernel is straightforward to evaluate for all exponential family models such as multinomials and Gaussians and yields interesting nonlinear kernels. Furthermore, the kernel is computable in closed form for latent distributions such as mixture models, hidden Markov models and linear dynamical systems. For intractable models, such as switching linear dynamical systems, structured mean-field approximations can be brought to bear on the kernel evaluation. For general distributions, even if an analytic expression for the kernel is not feasible, we show a straightforward sampling method to evaluate it. Thus, the kernel permits discriminative learning methods, including support vector machines, to exploit the properties, metrics and invariances of the generative models we infer from each datum. Experiments are shown using multinomial models for text, hidden Markov models for biological data sets and linear dynamical systems for time series data.  </paper_abstract>
<query_num> 9501 </query_num>
<text>   (=-=Ong et al., 2002-=-), convolutional kernels (=-=Haussler, 1999; Collins and Duffy, 2002-=-), alignment kernels (=-=Watkins, 2000-=-), rational kernels (=-=Cortes et al., 2002-=-) and string kernels (=-=Leslie et al., 2002;Vishawanathan and Smola, 2002-=-). This paper is organized as follows. In Section 2, we introduce the probability product kernel and point out connections to other probabilistic kernels such as Fisher kernels and exponentiated Kullb   </text>
<query_num> 9502 </query_num>
<text>   bles with ˜k(p, p ′ ) (and at times omit the ∼ symbol when it is self-evident) since it readily accommodates efficient marginalization and propagation algorithms (such as the junction tree algorithm (=-=Jordan and Bishop, 2004-=-)). One open issue with latent variables is that the ˜k kernels that marginalize over them may produce different values if the underlying latent distributions have different joint distributions over h |q ′ t| (T+1) elementary kernel evaluations. However, we can take advantage of the factorization of the hidden Markov model by using graphical modeling algorithms such as the junction tree algorithm (=-=Jordan and Bishop, 2004-=-) to compute the kernel efficiently. First, we use the factorization of the HMM in the kernel as follows: ˜k(p, p ′ ) = ∑ T ∑ ∏ p(xt|qt) x0,...,xT q0...qT t=0 ρ p(qt|qt−1) ρ ∑ q ′ 0 ...q ′ T ∏ p t=0 T   </text>
<query_num> 9503 </query_num>
<text>   bsequently compute the kernel. In other words, we map points asx → p(x|x ) andx ′ → p(x|x ′ ). For the setting of ρ = 1/2 the kernel is none other than the classical Bhattacharyya similarity measure (=-=Kondor and Jebara, 2003-=-), the affinity corresponding to Hellinger divergence (=-=Jebara and Kondor, 2003-=-). 1 One goal of this article is to explore a contact point between discriminative learning (support vector machines and k bler (KL) divergence, and in fact is a bound on KL, as shown in (=-=Topsoe, 1999-=-), where relationships between several common divergences are discussed. The Bhattacharyya kernel was first introduced in (=-=Kondor and Jebara, 2003-=-) and has the important special property k(x ,x ) = 1. When ρ = 1, the kernel takes the form of the expectation of one distribution under the other: k(x ,x ′ � ) = p(x) p ′ (x) dx = Ep[p ′ (x)] = Ep ′   </text>
<query_num> 9504 </query_num>
<text>   cussion and additional references. 6.1 Factorial Hidden Markov Models One model that traditionally uses a structured mean field approximation for inference is the factorial hidden Markov model (FHMM)(=-=Ghahramani and Jordan, 1997-=-). Such a model is appropriate for modeling output sequences that are emitted by multiple independent hidden processes. The FHMM is given by p(x0,x1,x2,...,xT) = C ∏ c=1 � p(q c 0) T ∏ t=1 p(q c t |q   </text>
<query_num> 9505 </query_num>
<text>   is known that square root squashing functions on frequencies and count typically outperform logarithmic squashing functions (=-=Goldzmidt and Sahami, 1998; Cutting et al., 1992-=-). In (=-=Tsuda et al., 2002;Kashima et al., 2003-=-), another related probabilistic kernel is put forward, the so-called marginalized kernel. This kernel is again similar to the Fisher kernel as well as the probability product kernel. It involves marg   </text>
<query_num> 9506 </query_num>
<text>   ive learning paradigm. The proposed probability product kernel falls into this latter category. Previous efforts to build kernels that accommodate probability distributions include the Fisher kernel (=-=Jaakkola and Haussler, 1998-=-), the heat kernel (=-=Lafferty and Lebanon, 2002-=-) and kernels arising from exponentiating KullbackLeibler divergences (=-=Moreno et al., 2004-=-). We discuss, compare and contrast these approaches to the prob robability product kernel is straightforward to compute for a much wider range of distributions providing a practical alternative to heat kernels. Another probabilistic approach is the Fisher kernel (=-=Jaakkola and Haussler, 1998-=-) which approximates distances and affinities on the statistical manifold by using the local metric at the maximum likelihood estimate θ ∗ of the whole data set as shown in Figure 6. One caveat, howev   </text>
<query_num> 9507 </query_num>
<text>   luding super-kernels (=-=Ong et al., 2002-=-), convolutional kernels (=-=Haussler, 1999; Collins and Duffy, 2002-=-), alignment kernels (=-=Watkins, 2000-=-), rational kernels (=-=Cortes et al., 2002-=-) and string kernels (=-=Leslie et al., 2002; Vishawanathan and Smola, 2002-=-). This paper is organized as follows. In Section 2, we introduce the probability product kernel and point out connections to other probabilistic kernels such as Fisher   </text>
<query_num> 9508 </query_num>
<text>   nd an appropriate kernel, the problem becomes accessible to a whole array of non-parametric discriminative kernel based learning algorithms, such as support vector machines, Gaussian processes, etc. (=-=Schölkopf and Smola, 2002-=-). In contrast, generative models fit probability distributions p(x ) tox1,x2,...,xm and base their predictions on the likelihood under different models. When faced with a discriminative task, approac   </text>
<query_num> 9509 </query_num>
<text>   product kernel falls into this latter category. Previous efforts to build kernels that accommodate probability distributions include the Fisher kernel (=-=Jaakkola and Haussler, 1998-=-), the heat kernel (=-=Lafferty and Lebanon, 2002-=-) and kernels arising from exponentiating KullbackLeibler divergences (=-=Moreno et al., 2004-=-). We discuss, compare and contrast these approaches to the probability product kernel in Section 7. One compe Cutting et al., 1992). Lafferty and Lebanon also arrive at a similar kernel, but from a very different angle, looking at the diffusion kernel on the statistical manifold of multinomial distributions (=-=Lafferty and Lebanon, 2002-=-). 2.2 Frequentist and Bayesian Methods of Estimation Various statistical estimation methods can be used to fit the distributions p1, p2,..., pm to the examplesx1,x2,...,xm. Perhaps it is most immedia erage generative modeling, it does have advantages in terms of computational feasibility as well as in nonlinear flexibility. For instance, heat kernels or diffusion kernels on statistical manifolds (=-=Lafferty and Lebanon, 2002-=-) are a more elegant way to generate a kernel in a probabilistic setting, but computations involve finding geodesics on complicated manifolds and finding closed form expressions for the heat kernel, w  web pages and 1124 faculty web pages. The data for each class is further split into 4 universities and 1 miscellaneous category and we performed the usual training and testing split as described by (=-=Lafferty and Lebanon, 2002-=-; Joachims et al., 2001) where testing is performed on a held out university. The average error was computed from 20-fold cross-validation for the different kernels as a function of the support vector   </text>
<query_num> 9510 </query_num>
<text>   riminative learning algorithms and optimize performance on a particular task. Examples of such approaches include conditional learning (=-=Bengio and Frasconi, 1996-=-) or large margin generative modeling (=-=Jaakkola et al., 1999-=-). Another approach is to use kernels to integrate the generative models within a discriminative learning paradigm. The proposed probability product kernel falls into this latter category. Previous ef   </text>
<query_num> 9511 </query_num>
<text>   te probability distributions include the Fisher kernel (=-=Jaakkola and Haussler, 1998-=-), the heat kernel (=-=Lafferty and Lebanon, 2002-=-) and kernels arising from exponentiating KullbackLeibler divergences (=-=Moreno et al., 2004-=-). We discuss, compare and contrast these approaches to the probability product kernel in Section 7. One compelling feature of the new kernel is that it is straightforward and efficient to compute ove nvolves an inner product over both hidden variables and the input space x with a joint distribution. A more recently introduced kernel, the exponentiated symmetrized Kullback-Leibler (KL) divergence (=-=Moreno et al., 2004-=-) is also related to the probability product kernel. This kernel involves exponentiating the negated symmetrized KL-divergence between two distributions k(p, p ′ ) = exp(−αD(p�p ′ ) − αD(p ′ �p)+β) wh d intractable graphical models) is intractable and must be approximated from the outset using numerical or sampling procedures which further decreases the possibility of having a valid Mercer kernel (=-=Moreno et al., 2004-=-). 8. Experiments In this section we discuss three different learning problems: text classification, biological sequence classification and time series classification. For each problem, we select a ge   </text>
<query_num> 9512 </query_num>
<text>   the two is to estimate parameters in generative models with discriminative learning algorithms and optimize performance on a particular task. Examples of such approaches include conditional learning (=-=Bengio and Frasconi, 1996-=-) or large margin generative modeling (=-=Jaakkola et al., 1999-=-). Another approach is to use kernels to integrate the generative models within a discriminative learning paradigm. The proposed probability   </text>
<query_num> 9513 </query_num>
<text>   thus gives a generative modeling route to kernel design to complement the family of other kernel engineering methods including super-kernels (=-=Ong et al., 2002-=-), convolutional kernels (=-=Haussler, 1999; Collins and Duffy, 2002-=-), alignment kernels (=-=Watkins, 2000-=-), rational kernels (=-=Cortes et al., 2002-=-) and string kernels (=-=Leslie et al., 2002; Vishawanathan and Smola, 2002-=-). This paper is organized as follows. In Section 2,   </text>
<query_num> 9514 </query_num>
<text>   x ′ → p(x|x ′ ). For the setting of ρ = 1/2 the kernel is none other than the classical Bhattacharyya similarity measure (=-=Kondor and Jebara, 2003-=-), the affinity corresponding to Hellinger divergence (=-=Jebara and Kondor, 2003-=-). 1 One goal of this article is to explore a contact point between discriminative learning (support vector machines and kernels) and generative learning (distributions and graphical models). Discrimi   </text>
<query_num> 9515 </query_num>
<text>   y of other kernel engineering methods including super-kernels (=-=Ong et al., 2002-=-), convolutional kernels (=-=Haussler, 1999; Collins and Duffy, 2002-=-), alignment kernels (=-=Watkins, 2000-=-), rational kernels (=-=Cortes et al., 2002-=-) and string kernels (=-=Leslie et al., 2002; Vishawanathan and Smola, 2002-=-). This paper is organized as follows. In Section 2, we introduce the probability product kernel and point out connections to ot   </text>
</top>
<top>
<paper_num> 96 </paper_num>
<paper_title>   Controlling Access to XML Documents over XML Native and Relational Databases.  </paper_title>
<paper_abstract>   Abstract. In this paper we investigate the feasibility and efficiency of mapping XML data and access control policies onto relational and native XML databases for storage and querying. We developed a re-annotation algorithm that computes the XPath query which designates the XML nodes to be re-annotated when an update operation occurs. The algorithm uses XPath static analysis and our experimental results show that our re-annotation solution is on the average 7 times faster than annotating the entire document.  </paper_abstract>
<query_num> 9601 </query_num>
<text>   ated problem of optimizing security checks during query evaluation with respect to an annotated document was investigated in [5]. XML access control over relational databases has been also studied in =-=[23]-=-. Our work is different in that we use annotations (materialized approach), whereas Lee et al. check the accessibility of the document on-the-fly. [20] discusses a “function-based” model that translat   </text>
<query_num> 9602 </query_num>
<text>   ation (med) with value “celecoxib” or a bill (bill) with a value greater than 1000 respectively are accessible. This set of rules is associated with a conflict resolution policy and default semantics =-=[14,15,17]-=-. The former specify the accessibility of a node in the case in which it is in the scope of access control rules with opposite signs. The later determines the default accessibility of a node. In our e   </text>
<query_num> 9603 </query_num>
<text>   er relational databases has been also studied in [23]. Our work is different in that we use annotations (materialized approach), whereas Lee et al. check the accessibility of the document on-the-fly. =-=[20]-=- discusses a “function-based” model that translates policy rules to functions (e.g. Java methods) which are subsequently called to check the policy whenever a part of the document is accessed. Securit   </text>
<query_num> 9604 </query_num>
<text>   es a “function-based” model that translates policy rules to functions (e.g. Java methods) which are subsequently called to check the policy whenever a part of the document is accessed. Security views =-=[10, 16]-=- address the problem of information leaks in the presence of read-only security policies and queries. Security views contain just the information a user is allowed to read; queries to the view can be   </text>
<query_num> 9605 </query_num>
<text>   n XPath expression p is contained in an XPath expression p ′ (p ⊑ p ′ ) iff the set of nodes obtained by evaluating p on any XML tree T is a subset of the set of nodes obtained by evaluating p ′ on T =-=[18, 19, 22]-=-. Algorithm Redundancy-Elimination shown in Figure 4 takes as input a set of access control rules S and returns a subset S ′ of S which is free of redundant rules. The idea is the following: redundanc e combined to obtain a revised policy. We employ the containment algorithm of [13,18]. Containment for fragments of XPath such as XP(/, //, ∗, []) has been studied in [18] and for larger fragments in =-=[19]-=- (see [22] for a survey).Redundancy-Elimination(rules) Ensure: ∀r1∀r2 ∈ rules ⇒ ∄r1 ⊏ r2 1: for all r ∈ rules do 2: for all r ′ ∈ rules where r ′ ̸= r do 3: if r ⊏ r ′ then 4: rules ← rules − {r} 5:   </text>
<query_num> 9606 </query_num>
<text>   n XPath expression p is contained in an XPath expression p ′ (p ⊑ p ′ ) iff the set of nodes obtained by evaluating p on any XML tree T is a subset of the set of nodes obtained by evaluating p ′ on T =-=[18, 19, 22]-=-. Algorithm Redundancy-Elimination shown in Figure 4 takes as input a set of access control rules S and returns a subset S ′ of S which is free of redundant rules. The idea is the following: redundanc nation is performed for both sets of positive and negative rules (A and D). The resulting redundancy-free sets of rules are combined to obtain a revised policy. We employ the containment algorithm of =-=[13,18]-=-. Containment for fragments of XPath such as XP(/, //, ∗, []) has been studied in [18] and for larger fragments in [19] (see [22] for a survey).Redundancy-Elimination(rules) Ensure: ∀r1∀r2 ∈ rules ⇒   </text>
<query_num> 9607 </query_num>
<text>   nodes; these specify whether a node is accessible or not. We evaluate our approach using (i) a native XML storage system and (ii) a relational database where the XML documents are shredded à la ShreX =-=[8]-=-. Specifically we: – propose a method to annotate XML documents stored in a relational database and in an XML database;– discuss an optimization procedure based on XPath containment that removes redu abase, we first need to create the relational tables used to store the XML document, and in a second phase, produce a relational representation of the XML document using these tables. We employ ShreX =-=[1, 8]-=- to obtain the relational representation of XML documents. ShreX is a system that handles the translation of XML data into relational tables. This includes relational schema creation, document loading nting the policy semantics as described in Table 2. In therelational case, the resource part of an access control rule (XPath expression) is translated into an equivalent SQL query q using the ShreX =-=[1, 8]-=- translation. The resource part of the rules that grant (resp. deny) access are unioned using the relational UNION (XQuery union) operator. Depending on the default semantics and conflict resolution p   </text>
<query_num> 9608 </query_num>
<text>   ource; – effect specifies whether the rule grants (“+” sign) or denies (“−” sign) access to the resource and finally – scope which defines whether the rule applies to the node only, or to its subtree =-=[11]-=-. In this paper we assume that the requester and action parameters are fixed and concentrate on the resource and effect components. We define the scope of a rule to be the XML node itself (explicit ru   </text>
<query_num> 9609 </query_num>
<text>   r storing and querying annotations have been investigated [26, 27]. The related problem of optimizing security checks during query evaluation with respect to an annotated document was investigated in =-=[5]-=-. XML access control over relational databases has been also studied in [23]. Our work is different in that we use annotations (materialized approach), whereas Lee et al. check the accessibility of th   </text>
<query_num> 9610 </query_num>
<text>   s. In this paper we study how to control access to XML documents stored in a relational database and in a native XML store. Prior work proposed the use of RDBMS for storing and querying XML documents =-=[1]-=-, to combine the flexibility and the usability of XML with the efficiency and the robustness of a relational schema. In this paper we examine the feasibility and efficiency of using the above approach abase, we first need to create the relational tables used to store the XML document, and in a second phase, produce a relational representation of the XML document using these tables. We employ ShreX =-=[1, 8]-=- to obtain the relational representation of XML documents. ShreX is a system that handles the translation of XML data into relational tables. This includes relational schema creation, document loading nting the policy semantics as described in Table 2. In therelational case, the resource part of an access control rule (XPath expression) is translated into an equivalent SQL query q using the ShreX =-=[1, 8]-=- translation. The resource part of the rules that grant (resp. deny) access are unioned using the relational UNION (XQuery union) operator. Depending on the default semantics and conflict resolution p   </text>
<query_num> 9611 </query_num>
<text>   starting with “/”), and T an XML tree, we write [p](T ) to denote the set of nodes of T obtained from evaluating expression p on the root node of T . The semantics of XPath expressions are defined in =-=[2, 12, 25]-=-. We say that an XPath expression p is contained in another expression q (denoted by p ⊑ q), if for every XML tree T , [p](T ) ⊆ [q ](T ). We say that two XPath expressions are disjoint (denoted by p   </text>
<query_num> 9612 </query_num>
<text>   the first attempt to compare the use of relational and XML databases to store annotated (with accessibility information) XML documents. Annotationbased enforcement techniques have been considered in =-=[3, 7]-=- for rule-based policies. More sophisticated techniques for storing and querying annotations have been investigated [26, 27]. The related problem of optimizing security checks during query evaluation   </text>
</top>
<top>
<paper_num> 97 </paper_num>
<paper_title>   Comparing the Decompositions Produced by Software Clustering Algorithms Using Similarity Measurements.  </paper_title>
<paper_abstract>   Decomposing source code components and relations into subsystem clusters is an active area of research. Numerous clustering approaches have been proposed in the reverse engineering literature, each one using a different algorithm to identify subsystems. Since di#erent clustering techniques may not produce identical results when applied to the same system, mechanisms that can measure the extent of these di#erences are needed. Some work to measure the similarity between decompositions has been done, but this work considers the assignment of source code components to clusters as the only criterion for similarity. We argue that better similarity measurements can be designed if the relations between the components are considered.  In this paper we propose two similarity measurements that overcome certain problems in existing measurements. We also provide some suggestions on how to identify and deal with source code components that tend to contribute to poor similarity results. We conclude by presenting experimental results, and by highlighting some of the benefits of our similarity measurements. 1.  </paper_abstract>
<query_num> 9701 </query_num>
<text>   e different algorithms, and make different assumptions about how subsystems are formed. Most techniques use source code component similarity [8, 16, 6, 15], optimization [13, 7, 12], concept analysis =-=[11, 20, 1]-=-, or implementation information such as module, directory, and/or package names [2, 3] to determine the clusters. Now that many clustering techniques exist, some researchers have turned their attentio   </text>
<query_num> 9702 </query_num>
<text>   e, directory, and/or package names [2, 3] to determine the clusters. Now that many clustering techniques exist, some researchers have turned their attention to evaluating their relative effectiveness =-=[10, 18, 19, 2, 3]-=-. There are several reasons for this: • Many of the papers on software clustering formulate conclusions based on case studies, or by soliciting opinions from the authors of the systems presented in th   </text>
<query_num> 9703 </query_num>
<text>   e, directory, and/or package names [2, 3] to determine the clusters. Now that many clustering techniques exist, some researchers have turned their attention to evaluating their relative effectiveness =-=[10, 18, 19, 2, 3]-=-. There are several reasons for this: • Many of the papers on software clustering formulate conclusions based on case studies, or by soliciting opinions from the authors of the systems presented in th ecompositions. For example, Anquetil et al. developed a similarity measurement based on Precision and Recall [2, 3]. Tzerpos and Holt authored a paper on a similarity distance measurement called MoJo =-=[18]-=-. Both of these measurements treat each difference between two decompositions of the same system equally. In Section 3 we present both a similarity and a distance measurement that ranks the individual lustering algorithm, but rather of the tendency of some modules to be assigned to more than one subsystem across clustering results. Similarity measurements can also be used to evaluate the stability =-=[17, 18, 19]-=- of a clustering algorithm. The idea behind stability is that small changes to a software system’s structure should not result in large differences in how a clustering algorithm decomposes the system.   </text>
<query_num> 9704 </query_num>
<text>   e, directory, and/or package names [2, 3] to determine the clusters. Now that many clustering techniques exist, some researchers have turned their attention to evaluating their relative effectiveness =-=[10, 18, 19, 2, 3]-=-. There are several reasons for this: • Many of the papers on software clustering formulate conclusions based on case studies, or by soliciting opinions from the authors of the systems presented in th lustering algorithm, but rather of the tendency of some modules to be assigned to more than one subsystem across clustering results. Similarity measurements can also be used to evaluate the stability =-=[17, 18, 19]-=- of a clustering algorithm. The idea behind stability is that small changes to a software system’s structure should not result in large differences in how a clustering algorithm decomposes the system.   </text>
<query_num> 9705 </query_num>
<text>   f little value. However, if a system is clustered several times, using different clustering algorithms, and similar decompositions are produced, we gain confidence that the proposed clusters are good =-=[14]-=-. Because various clustering algorithms use different criteria to form clusters, we need to identify if large differences in similarity are due to the clustering algorithms, or if the differences are  our similarity measurements when comparing decompositions produced by clustering algorithms other then Bunch. We have already taken some of the initial steps towards this goal by creating a framework =-=[14]-=- for software evaluation. 7. Acknowledgements This research is sponsored by grants CCR-9733569 and CISE-9986105 from the National Science Foundation (NSF). Additional support was provided by the resea   </text>
<query_num> 9706 </query_num>
<text>   he various clustering tools use different algorithms, and make different assumptions about how subsystems are formed. Most techniques use source code component similarity [8, 16, 6, 15], optimization =-=[13, 7, 12]-=-, concept analysis [11, 20, 1], or implementation information such as module, directory, and/or package names [2, 3] to determine the clusters. Now that many clustering techniques exist, some research view the results of our comparative study of similarity measurements. We conclude, in Section 6, with a summary of the research contributions of this work. 2. Measuring Similarity In an earlier paper =-=[13]-=- we show that the number of unique ways to decompose a software system into non-overlapping clusters (subsystems) grows exponentially with respect to the number of modules/classes in the source code.  o illustrate the effectiveness of the similarity measurements presented in Section 3, and to validate our assumptions about removing special modules from the clustering process. We will use the Bunch =-=[13, 7, 12]-=- clustering tool because it has several unique features that lend themselves well to our study. In particular, Bunch uses randomization in its optimization approach to form clusters, therefore, becaus s. Bunch determines candidate clusters by moving nodes between the clusters, or in some cases creating new clusters in order to maximize an objective function that we call Modularization Quality (MQ) =-=[13]-=-. Our MQ function has the property of rewarding cohesive clusters, while penalizing excessive inter-cluster coupling. Bunch’s non-determinism also provides the basis for further study about why some s   </text>
<query_num> 9707 </query_num>
<text>   he various clustering tools use different algorithms, and make different assumptions about how subsystems are formed. Most techniques use source code component similarity [8, 16, 6, 15], optimization =-=[13, 7, 12]-=-, concept analysis [11, 20, 1], or implementation information such as module, directory, and/or package names [2, 3] to determine the clusters. Now that many clustering techniques exist, some research ystem). Without automation, the identification of omnipresent and library modules is tedious. To address this, we added a feature to identify these special modules to our clustering tool called Bunch =-=[12]-=-. Bunch can propose candidate omnipresent and library modules automatically, while allowing the user to add or remove these special modules if necessary. Once these modules are identified, the softwar o illustrate the effectiveness of the similarity measurements presented in Section 3, and to validate our assumptions about removing special modules from the clustering process. We will use the Bunch =-=[13, 7, 12]-=- clustering tool because it has several unique features that lend themselves well to our study. In particular, Bunch uses randomization in its optimization approach to form clusters, therefore, becaus   </text>
<query_num> 9708 </query_num>
<text>   ip between a pair of modules in the system ‡ . Graph G can be generated for most software systems using readily available source code analysis tools such as Acacia for C/C++ [5, 4] and Chava for Java =-=[9]-=-. Consider two partitions, A and B, of graph G. Let Ai, 1 ≤ i ≤ k, be the clusters of A, and Bj, 1 ≤ j ≤ l be the clusters of B. Each Ai and Bj are subsets of the vertices in G. Figure 4 shows a sampl   </text>
<query_num> 9709 </query_num>
<text>   lustering algorithm, but rather of the tendency of some modules to be assigned to more than one subsystem across clustering results. Similarity measurements can also be used to evaluate the stability =-=[17, 18, 19]-=- of a clustering algorithm. The idea behind stability is that small changes to a software system’s structure should not result in large differences in how a clustering algorithm decomposes the system.   </text>
<query_num> 9710 </query_num>
<text>   strength of the relationship between a pair of modules in the system ‡ . Graph G can be generated for most software systems using readily available source code analysis tools such as Acacia for C/C++ =-=[5, 4]-=- and Chava for Java [9]. Consider two partitions, A and B, of graph G. Let Ai, 1 ≤ i ≤ k, be the clusters of A, and Bj, 1 ≤ j ≤ l be the clusters of B. Each Ai and Bj are subsets of the vertices in G.   </text>
<query_num> 9711 </query_num>
<text>   t techniques use source code component similarity [8, 16, 6, 15], optimization [13, 7, 12], concept analysis [11, 20, 1], or implementation information such as module, directory, and/or package names =-=[2, 3]-=- to determine the clusters. Now that many clustering techniques exist, some researchers have turned their attention to evaluating their relative effectiveness [10, 18, 19, 2, 3]. There are several rea issues, researchers have begun formulating ways to measure the differences between system decompositions. For example, Anquetil et al. developed a similarity measurement based on Precision and Recall =-=[2, 3]-=-. Tzerpos and Holt authored a paper on a similarity distance measurement called MoJo [18]. Both of these measurements treat each difference between two decompositions of the same system equally. In Se /Recall of Figure 2). This mistakenly leads us to believe that A and B in Figure 2 are more similar than the partitions shown in Figure 3. 2.3. Interpreting Similarity Results Anquetil and Lethbridge =-=[3]-=- state that clustering techniques do not recover a software system’s decomposition, but rather impose one. One might ask: What does it mean when the results produced by two different clustering techni   </text>
</top>
<top>
<paper_num> 98 </paper_num>
<paper_title>   REDUS: finding reducible subspaces in high dimensional data.  </paper_title>
<paper_abstract>   Finding latent patterns in high dimensional data is an important research problem with numerous applications. The most well known approaches for high dimensional data analysis are feature selection and dimensionality reduction. Being widely used in many applications, these methods aim to capture global patterns and are typically performed in the full feature space. In many emerging applications, however, scientists are interested in the local latent patterns held by feature subspaces, which may be invisible via any global transformation. In this paper, we investigate the problem of finding strong linear and nonlinear correlations hidden in feature subspaces of high dimensional data. We formalize this problem as identifying reducible subspaces in the full dimensional space. Intuitively, a reducible subspace is a feature subspace whose intrinsic dimensionality is smaller than the number of features. We present an effective algorithm, REDUS, for finding the reducible subspaces. Two key components of our algorithm are finding the overall reducible subspace, and uncovering the individual reducible subspaces from the overall reducible subspace. A broad experimental evaluation demonstrates the effectiveness of our algorithm.  </paper_abstract>
<query_num> 9801 </query_num>
<text>   Therefore, its intrinsic dimensionality is 2. The concept of intrinsic dimensionality has many applications in the database and data mining communities, such as clustering [3, 15], outlier detection =-=[24]-=-, nearest neighbor queries [23], and spatial query selectivity estimation [5, 12]. Different definitions of intrinsic dimensionality can be found in the literature. For example, in linear cases, matri   </text>
<query_num> 9802 </query_num>
<text>   applied after such strongly correlated feature subspaces have been identified. Correlation Clustering The goal of correlation clustering methods is to find clusters hidden in projected feature spaces =-=[1, 7, 30]-=-. They can be viewed as combinations of clustering methods and dimensionality reduction methods. Both ORCLUS [1] and 4C [7] can be treated as PCA-lized clustering methods. To find the low dimensional   </text>
<query_num> 9803 </query_num>
<text>   applied after such strongly correlated feature subspaces have been identified. Correlation Clustering The goal of correlation clustering methods is to find clusters hidden in projected feature spaces =-=[1, 7, 30]-=-. They can be viewed as combinations of clustering methods and dimensionality reduction methods. Both ORCLUS [1] and 4C [7] can be treated as PCA-lized clustering methods. To find the low dimensional  ientations. Therefore, they do not touch the problem of finding reducible subspaces. Instead, they implicitly assume the full dimensional space is reducible for certain subsets of data points. CURLER =-=[30]-=- finds clusters having non-linear correlations in subspaces. It first applies EM clustering algorithm to find a large number of microclusters, and then merges clusters with large overlaps. Since in it   </text>
<query_num> 9804 </query_num>
<text>   dimensional space. Informally, a feature subspace is reducible if its intrinsic dimensionality is smaller than the number of features. Various intrinsic dimensionality estimators have been developed =-=[9, 14, 21]-=-. Our problem formalization does not depend on any particular method for estimating the intrinsic dimensionality. We show two necessary properties that any estimator should satisfy in order to be a ge  finding the reducible subspaces. Intrinsic Dimensionality Due to correlations among features, a high dimensional dataset may lie in a subspace with dimensionality smaller than the number of features =-=[9, 14, 21]-=-. The intrinsic dimensionality can be treated as the minimum number of free variables required to define the data without any significant information loss [13]. For example, as shown in Figure 1, in t   </text>
<query_num> 9805 </query_num>
<text>   he data mining task at hand, such as classification. The selected features generally have low correlation with each other but have strong correlation with the target feature. Dimensionality reduction =-=[4, 8, 18, 27, 29]-=- is widely used as a key component of many approaches in analyzing high dimensional data. The insight behind dimensionality reduction methods is that a high dimensional dataset may exhibit interesting ethods, such as Multi-Dimensional Scaling (MDS) [8] and Principal Component Analysis (PCA)[18], and non-linear methods, such as Local Linear Embedding (LLE) [27], ISOMAP [29], and Laplacian eigenmaps =-=[4]-=-. For high dimensional datasets, if there exist low dimensional subspaces or manifolds embedded in the full dimensional spaces, these methods are successful in identifying these low dimensional embedd   </text>
<query_num> 9806 </query_num>
<text>   he data mining task at hand, such as classification. The selected features generally have low correlation with each other but have strong correlation with the target feature. Dimensionality reduction =-=[4, 8, 18, 27, 29]-=- is widely used as a key component of many approaches in analyzing high dimensional data. The insight behind dimensionality reduction methods is that a high dimensional dataset may exhibit interesting on methods can be categorized into linear methods, such as Multi-Dimensional Scaling (MDS) [8] and Principal Component Analysis (PCA)[18], and non-linear methods, such as Local Linear Embedding (LLE) =-=[27]-=-, ISOMAP [29], and Laplacian eigenmaps [4]. For high dimensional datasets, if there exist low dimensional subspaces or manifolds embedded in the full dimensional spaces, these methods are successful i   </text>
<query_num> 9807 </query_num>
<text>   ionality has many applications in the database and data mining communities, such as clustering [3, 15], outlier detection [24], nearest neighbor queries [23], and spatial query selectivity estimation =-=[5, 12]-=-. Different definitions of intrinsic dimensionality can be found in the literature. For example, in linear cases, matrix rank [16] and PCA [18] can be used to estimate intrinsic dimensionality. For no   </text>
<query_num> 9808 </query_num>
<text>   ructures hidden in the large number of features. Two well known approaches in analyzing high dimensional data are feature selection and dimensionality reduction. The goal of feature selection methods =-=[6, 22, 31, 33]-=- is to find a single representative subset of features that are most relevant for the data mining task at hand, such as classification. The selected features generally have low correlation with each o f3 f3 10 5 0 -5 -10 -15 -25 -20 -15 -10 -5 0 5 10 f2 (c) f2 and f2 Figure 3: Pairwise correlations of the Swiss roll in the example dataset 2. RELATED WORK Feature Selection Feature selection methods =-=[6, 22, 31, 33]-=- try to find a subset of features that are most relevant for certain data mining task, such as classification. In order to find the relevant feature subset, these methods search through various subset   </text>
<query_num> 9809 </query_num>
<text>   sionality is 2. The concept of intrinsic dimensionality has many applications in the database and data mining communities, such as clustering [3, 15], outlier detection [24], nearest neighbor queries =-=[23]-=-, and spatial query selectivity estimation [5, 12]. Different definitions of intrinsic dimensionality can be found in the literature. For example, in linear cases, matrix rank [16] and PCA [18] can be   </text>
<query_num> 9810 </query_num>
<text>   y a 2-dimensional manifold. Therefore, its intrinsic dimensionality is 2. The concept of intrinsic dimensionality has many applications in the database and data mining communities, such as clustering =-=[3, 15]-=-, outlier detection [24], nearest neighbor queries [23], and spatial query selectivity estimation [5, 12]. Different definitions of intrinsic dimensionality can be found in the literature. For example   </text>
</top>
<top>
<paper_num> 99 </paper_num>
<paper_title>   An Abductive Semantics for Disjunctive Logic Programs and Its Proof Procedure.  </paper_title>
<paper_abstract>   While it is well-known how normal logic programs may be viewed as a form of abduction and argumentation, the problem of how disjunctive programs may be used for abductive reasoning is rarely discussed. In this paper we propose an abductive semantics for disjunctive logic programs with default negation and show that Eshghi and Kowalski&amp;apos;s abductive proof procedure for normal programs can be adopted to compute abductive solutions for disjunctive programs.  </paper_abstract>
<query_num> 9901 </query_num>
<text>   M is a stable model of P iff SM is a stable extension of P where SM = fnot OE : OE 62 Mg. 2 This result provides another way to understand the results by Eiter and Gottlob as well as by Marek et al. =-=[9, 14, 22]-=-: disjunctive default logic falls into the same complexity level as default logic. In Gottlob&amp;apos;s explanation of these complexity levels, this can be said rather plainly as: for disjunctive programs epi   </text>
<query_num> 9902 </query_num>
<text>   al models. Logic programming has been identified with a goal-oriented programming paradigm. An implementation of a nonmonotonic semantics could benefit from the techniques used in implementing Prolog =-=[4]-=-. Thus, the existence of a top down proof procedure is a significant feature for any semantics. It is important to mention the three remarkable facts about Eshghi and Kowalski&amp;apos;s procedure. First, it i   </text>
<query_num> 9903 </query_num>
<text>   as a form of abductive reasoning. In particular, default assumptions in logic programs have been treated as abductive hypotheses and a number of reasoning mechanisms and semantics have been proposed =-=[7, 10, 16, 18, 19]-=-. Chief among these is Eshghi and Kowalski&amp;apos;s formulation of an elegant abductive proof procedure for normal programs where default assumptions are viewed as abducibles. Kakas et al. presented a compre for some not �� 2 S 0 . A preferred extension E is a maximal assumption set that is ?-consistent, and for every not OE 2 E, not OE is acceptable w.r.t. E. In terms of argumentation (see, for example, =-=[2, 8, 16, 28]-=-), not OE is acceptable w.r.t. S if any assumption set S 0 that attacks not OE is counter-attacked by S itself. Note that the definition is precisely that of Dung&amp;apos;s if we use ` instead of ` d . In fac   </text>
<query_num> 9904 </query_num>
<text>   ation of an elegant abductive proof procedure for normal programs where default assumptions are viewed as abducibles. Kakas et al. presented a comprehensive exploration of abductive logic programming =-=[16, 17]-=-. A fundamental insight is that abductive reasoning embodies an argumentation approach to logic program semantics. Dung [8], as well as Bondarenko et al. [2], subsequently showed that nonmonotonic rea  credit, not pay by credit(bob; greg); which resolves with the above disjunction to yield pay cash(bob; greg). The approach taken here falls into the general frameworks of abduction and argumentation =-=[2, 8, 17]-=- in that a specific inference system (using the above inference rule) is adopted for the purpose of capturing the meaning of the epistemic disjunction initially formulated by Gelfond and Lifschitz. Th   </text>
<query_num> 9905 </query_num>
<text>   er the left or right hand is broken. 1 An explanation of this observation must include a prediction that it is the left hand that is broken. We note that the static semantics for disjunctive programs =-=[24]-=- is not designed to be capable of accommodating these kind of applications. Despite well understood results in relating normal programs with abduction, and in more general cases, relating more general   </text>
<query_num> 9906 </query_num>
<text>   for some not �� 2 S 0 . A preferred extension E is a maximal assumption set that is ?-consistent, and for every not OE 2 E, not OE is acceptable w.r.t. E. In terms of argumentation (see, for example, =-=[2, 8, 16, 28]-=-), not OE is acceptable w.r.t. S if any assumption set S 0 that attacks not OE is counter-attacked by S itself. Note that the definition is precisely that of Dung&amp;apos;s if we use ` instead of ` d . In fac   </text>
<query_num> 9907 </query_num>
<text>   insights into a different aspect of the problem; how to represent abductive programs by (extended) disjunctive programs. In [15], they proposed a transformation from the former to the latter, and in =-=[27]-=-, they showed, in general, an abductive program can be viewed as a disjunctive program with priorities. Our work is about a proposal of an abductive/argumentation framework for disjunctive programs. F   </text>
<query_num> 9908 </query_num>
<text>   it has come to our attention that, in an unpublished manuscript [6], Dung defined a procedure that combines Eshghi-Kowalski procedure with a form of linear resolution, SLI-resolution of Lobo et al.&amp;apos;s =-=[21]-=-. The relationship between this procedure and our semantics is currently being investigated. 6 Related Work and Final Remarks Work by Inoue and Sakama [15, 15] yields important insights into a differe   </text>
<query_num> 9909 </query_num>
<text>   it is now known that this procedure is sound, and complete as well with a mechanism of positive loop checking, for a number of independently proposed but equivalent semantics for normal programs (cf. =-=[7, 13, 20, 30]-=-). This includes the preferential semantics based on abduction [7], the regular semantics [29] and the partial stable semantics [25] using model-theoretic approachs, and a stronger version of the stab ctive programs epistemic disjunction does not present one more source of computational difficulty than classic disjunction. 4 A Fixpoint Characterization We show that some of the results presented in =-=[30]-=-, namely the results that regular models/preferred extensions are maximal normal alternating fixpoints of a suitable operator for normal programs, can be extended to the case of disjunctive programs.  tion that applies FP twice, denoted F 2 P , is monotonic. That is, S 1 ` S 2 ) F 2 P (S 1 ) ` F 2 P (S 2 ) A fixpoint of the function F 2 P is called an alternating fixpoint of FP (or simply, P ; cf. =-=[1, 11, 30]-=-). An alternating fixpoint S is called normal if S ` FP (S). For credulous reasoning, we are interested in maximal normal alternating fixpoints. On the other hand, Dung&amp;apos;s preferred extensions are, by   </text>
<query_num> 9910 </query_num>
<text>   oposed but equivalent semantics for normal programs (cf. [7, 13, 20, 30]). This includes the preferential semantics based on abduction [7], the regular semantics [29] and the partial stable semantics =-=[25]-=- using model-theoretic approachs, and a stronger version of the stable class 2 The procedure becomes complete when properly combined with a complete resolution procedure. semantics [1]. Perhaps more i   </text>
<query_num> 9911 </query_num>
<text>   ration of abductive logic programming [16, 17]. A fundamental insight is that abductive reasoning embodies an argumentation approach to logic program semantics. Dung [8], as well as Bondarenko et al. =-=[2]-=-, subsequently showed that nonmonotonic reasoning in general is a form of argumentation using default assumptions. There are important applications of abductive reasoning with disjunctive programs. Fo e kind of applications. Despite well understood results in relating normal programs with abduction, and in more general cases, relating more general inference systems with abduction and argumentation =-=[2]-=-, the problem of how disjunctive programs may be viewed as abduction is still open. For example, Dung showed in [5] that acyclic disjunctive programs can be interpreted as abductive programs in the se  credit, not pay by credit(bob; greg); which resolves with the above disjunction to yield pay cash(bob; greg). The approach taken here falls into the general frameworks of abduction and argumentation =-=[2, 8, 17]-=- in that a specific inference system (using the above inference rule) is adopted for the purpose of capturing the meaning of the epistemic disjunction initially formulated by Gelfond and Lifschitz. Th for some not �� 2 S 0 . A preferred extension E is a maximal assumption set that is ?-consistent, and for every not OE 2 E, not OE is acceptable w.r.t. E. In terms of argumentation (see, for example, =-=[2, 8, 16, 28]-=-), not OE is acceptable w.r.t. S if any assumption set S 0 that attacks not OE is counter-attacked by S itself. Note that the definition is precisely that of Dung&amp;apos;s if we use ` instead of ` d . In fac   </text>
<query_num> 9912 </query_num>
<text>   resolution, SLI-resolution of Lobo et al.&amp;apos;s [21]. The relationship between this procedure and our semantics is currently being investigated. 6 Related Work and Final Remarks Work by Inoue and Sakama =-=[15, 15]-=- yields important insights into a different aspect of the problem; how to represent abductive programs by (extended) disjunctive programs. In [15], they proposed a transformation from the former to th   </text>
<query_num> 9913 </query_num>
<text>   tion that applies FP twice, denoted F 2 P , is monotonic. That is, S 1 ` S 2 ) F 2 P (S 1 ) ` F 2 P (S 2 ) A fixpoint of the function F 2 P is called an alternating fixpoint of FP (or simply, P ; cf. =-=[1, 11, 30]-=-). An alternating fixpoint S is called normal if S ` FP (S). For credulous reasoning, we are interested in maximal normal alternating fixpoints. On the other hand, Dung&amp;apos;s preferred extensions are, by   </text>
</top>
<top>
<paper_num> 100 </paper_num>
<paper_title>   A Quantitative Assessment of Thread-Level Speculation Techniques.  </paper_title>
<paper_abstract>   Speculative thread-level parallelism has been recently proposed as an alternative source of parallelism that can boost the performance for applications where independent threads are hard to find. Several schemes to exploit thread level parallelism have been proposed and significant performance gains have been reported. However, the sources of the performance gains are poorly understood as well as the impact of some design choices. In this work, the advantages of different thread speculation techniques are analyzed as are the impact of some critical issues including the value predictor, the branch predictor, the thread initialization overhead and the connectivity among thread units.  </paper_abstract>
<query_num> 10001 </query_num>
<text>   aces is specially suited to maximize the load balance among the different thread units. Finally, several works on speculative thread-level parallelism on multiprocessor platforms have appeared ([7][8]=-=[18]-=- among others). In all cases, programs are split by the compiler and usually no mechanism for predicting dependent values among dependent threads is provided. 3. Dynamic Exploitation of Thread Level P   </text>
<query_num> 10002 </query_num>
<text>   cases significant benefits have been reported but the absolute figures are not comparable due to very different architectural assumptions. Some work comparing different spawning policies has been done=-=[14]-=-. However, their base architecture was an on-chip multiprocessor where each processing unit wassone-way and issued the instructions in program order. In such architecture, the effects of the ILP on th   </text>
<query_num> 10003 </query_num>
<text>   he program into threads whereas others rely only on hardware techniques. Examples of the latter group are the Dynamic Multithreaded Processor [1] and the Clustered Speculative Multithreaded Processor =-=[10]-=-[11]. These works have shown that speculative thread-level parallelism has significant potential to boost performance. However, most of them use different heuristics to partition a sequential instruct some other architectures try to exploit thread-level parallelism speculating on threads dynamically created by the processor without any compiler intervention. The Speculative Multithreaded Processor =-=[10]-=- and its successor, the Clustered Speculative Multithreaded Processor [11] identifies loops at run time and simultaneously executes iterations in different thread units. Also, this architecture provid   </text>
<query_num> 10004 </query_num>
<text>   hread creation which requires communication between any hardware context since it has a centralized implementation of hardware contexts in a similar way to the Simultaneous Multithreaded Architecture =-=[20]-=-. Trace Processors[15] also exploit certain kind of speculative thread-level parallelism. Its mechanism to split the sequential program into almost fixed-length traces is specially suited to maximize   </text>
<query_num> 10005 </query_num>
<text>   ism (ILP) that inherently exists in programs at run-time. The exploitation of ILP in superscalar processors is strongly constrained by several significant hurdles, such as the instruction window size =-=[22]-=- and the data dependences (RAW dependences) [22]. To exploit large amounts of ILP, dynamically-scheduled superscalar processors require a large instruction window filled with useful instructions so th itat Politècnica de Catalunya Jordi Girona, 1-3 Mòdul D6 08034 Barcelona, Spain instructions produce average performance of just several tens of instructions per cycle (IPC) for some integer programs =-=[22]-=-. Data value speculation has recently been proposed to relieve the penalties due to data dependences and minimize their impact on the performance of the processor by means of predicting the input/outp   </text>
<query_num> 10006 </query_num>
<text>   recently been proposed to relieve the penalties due to data dependences and minimize their impact on the performance of the processor by means of predicting the input/output operands of instructions (=-=[9]-=- among others). However, recent studies show that the performance impact of data value speculation for superscalar processors is moderate, and its potential improvement approaches a linear function of   </text>
<query_num> 10007 </query_num>
<text>   rogram into threads whereas others rely only on hardware techniques. Examples of the latter group are the Dynamic Multithreaded Processor [1] and the Clustered Speculative Multithreaded Processor [10]=-=[11]-=-. These works have shown that speculative thread-level parallelism has significant potential to boost performance. However, most of them use different heuristics to partition a sequential instruction  ng on threads dynamically created by the processor without any compiler intervention. The Speculative Multithreaded Processor [10] and its successor, the Clustered Speculative Multithreaded Processor =-=[11]-=- identifies loops at run time and simultaneously executes iterations in different thread units. Also, this architecture provides mechanisms for value prediction in order to execute the concurrent thre   </text>
<query_num> 10008 </query_num>
<text>   the stride [4][5] and the FCM context-based value predictors[16] are well-known and have been thoroughly studied for superscalar and VLIW architectures. We also evaluate the increment value predictor =-=[12]-=-. The increment value predictor is an output-value traceoriented predictor that predicts each output value of a thread to be the value of that register at the beginning of the thread plus an increment   </text>
<query_num> 10009 </query_num>
<text>   thread and consumed by another. They differ in how programs are split into threads. In several proposals such as the Multiscalar [3][17], the SPSM architecture [2] and the Superthreaded architecture =-=[19]-=-, the compiler splits the program into threads whereas others rely only on hardware techniques. Examples of the latter group are the Dynamic Multithreaded Processor [1] and the Clustered Speculative M  minimize the data dependences among active threads or to have a better load balance, among other compiler criterias [21]. Other architectures such as the SPSM [2] and the Superthreaded architectures =-=[19]-=- also rely on the compiler to split the program into threads, but in these cases, threads are assumed to be loop iterations instead of the more complex analysis of the Multiscalar compiler. On the oth   </text>
<query_num> 10010 </query_num>
<text>   traces is specially suited to maximize the load balance among the different thread units. Finally, several works on speculative thread-level parallelism on multiprocessor platforms have appeared ([7]=-=[8]-=-[18] among others). In all cases, programs are split by the compiler and usually no mechanism for predicting dependent values among dependent threads is provided. 3. Dynamic Exploitation of Thread Lev   </text>
</top>
<top>
<paper_num> 101 </paper_num>
<paper_title>   On the effectiveness of automatic patching.  </paper_title>
<paper_abstract>   We study the e#ectiveness of automatic patching and quantify the speed of patch dissemination required for worm containment. We focus on random scanning as this is representative of current generation worms, though smarter strategies exist. We find that even such &amp;quot;dumb&amp;quot; worms require very fast patching. Our primary focus is on how delays due to worm detection and patch generation and dissemination a#ect worm spread. Motivated by scalability and trust issues, we consider a hierarchical system where network hosts are partitioned into subnets, each containing a patch server (termed superhost). Patches are disseminated to superhosts through an overlay connecting them and, after verification, to end hosts within subnets. When patch dissemination delay on the overlay is negligible, we find that the number of hosts infected is exponential in the ratio of worm infection rate to patch rate. This implies strong constraints on the time to disseminate, verify and install patches in order for it to be e#ective. We also provide bounds that account for alert or patch dissemination delay. Finally, we evaluate the use of filtering in combination with patching and show that it can substantially improve worm containment. The results accommodate a variety of overlays by a novel abstraction of minimum broadcast curve. They demonstrate that e#ective automatic patching is feasible if combined with mechanisms to bound worm scan rate and with careful engineering of the patch dissemination. The results are obtained analytically and verified by simulations.  </paper_abstract>
<query_num> 10101 </query_num>
<text>   (8) d dt w(t) = µ w(t)2 − (µ + g(t)) w(t) + g(t). (9) The last differential equation is known as Ricatti’s differential equation and it can be solved explicitly. The interested reader is referred to =-=[2]-=- for the solution and limit points of w(t) and discussion. Comment. A somewhat similar host immunization process to (7)–(8) was considered by Wong et al [18]. They do not consider patch dissemination  ot be patched on the timescale of the automatic patching system. This may be a more plausible assumption as a smart worm may prevent patch installation on an infected host. Wenowpresentourresults; see=-=[2]-=-fordetailedderivations. Theorem 1. For the system of differential equations (1)– (5), it holds that i(+∞) + µ β � +∞ 0 w(u)d log i(u) = i(0) + s(0). (10) Corollary 1. Assume g(0) = 1, i.e. all superho .5 1 1.5 time (sec) 2 2.5 3 Figure 1: Empirical fraction of alerted superhosts achieved by flooding on Pastry overlay of 100 Pastry nodes and delayed-logistic minimum broadcast curve. See Figure 4 in =-=[2]-=- for additional plots. The figure suggests delayed-logistic curve a natural candidate for minimum broadcast curve of standard overlays. Assume that f1(t) ≥ f2(t) for all t ≥ 0, and that on any finite  he fraction of infected hosts in the actual system. In Figure 1, we compare empirical broadcast curve obtained by flooding in Pastry and a minimum broadcast curve taken to be a logistic function; see =-=[2]-=- for details. 4. PUSH-BASED PATCH DISSEMINATION We have so far discussed a pull mechanism for patch dissemination, motivated by currently deployed systems. We now explore push schemes for comparison.  ceeds only if neither superhost k nor superhost m are alerted. As earlier, each host under an alerted superhost installs a patch with rate µ. TheraceofwormandpatchcanbedescribedbyaMarkov process (see =-=[2]-=-). Here we directly proceed to the limit population dynamics under the many hosts and many superhosts assumptions. We first consider the subpopulation of hosts in non-alerted subnets. Denote by i 0 (t   </text>
<query_num> 10102 </query_num>
<text>   0 seconds for Slammer. In the case of a worm like Slammer, this suggests that we either need to patch most hosts within a timescale of about 1 minute or deploy additional mechanisms like rate capping =-=[17]-=- in order to slow down the worm. Inanidealizedscenario, whenalertdisseminationbetween superhosts is instantaneous, we obtain an exact expression relating the initial and final number of infected hosts   </text>
<query_num> 10103 </query_num>
<text>   Patch dissemination The overlay connecting superhosts can be arbitrary. For example, it could be a balanced multicast tree or any one of several standard structured overlays (e.g. Pastry [12], Chord =-=[16]-=- or CAN [11]) or an unstructured overlay (e.g. Gnutella). Our framework only requires knowing a lower bound on the number of alerted superhosts at any time after the start of patch dissemination. We i   </text>
<query_num> 10104 </query_num>
<text>   itative understanding of how the relative speeds of worm and patch influence the outcome of the epidemic is important. 1.3 Related work Much work has been done on studying worms and their containment =-=[1, 6, 9, 13, 15]-=-. We do not aim at addressing all related work, but only that which to our knowledge is closely related to automatic patching. There are several schemes for worm detection, e.g., (i) honeypots: these   </text>
<query_num> 10105 </query_num>
<text>   itative understanding of how the relative speeds of worm and patch influence the outcome of the epidemic is important. 1.3 Related work Much work has been done on studying worms and their containment =-=[1, 6, 9, 13, 15]-=-. We do not aim at addressing all related work, but only that which to our knowledge is closely related to automatic patching. There are several schemes for worm detection, e.g., (i) honeypots: these   using the automatic response system. A similar system was also proposed by Sidiroglou and Keromytis [14]. There has also beenworkonanalysingthecompetingprocessesofpatching, filtering and worm spread =-=[13, 15, 18]-=-. These works typically consider a ‘flat’ network for both the worm and patch processes, whereas we study a hierarchical model motivated by considerations of scalability and trust. In practice, Micros   </text>
<query_num> 10106 </query_num>
<text>   mination The overlay connecting superhosts can be arbitrary. For example, it could be a balanced multicast tree or any one of several standard structured overlays (e.g. Pastry [12], Chord [16] or CAN =-=[11]-=-) or an unstructured overlay (e.g. Gnutella). Our framework only requires knowing a lower bound on the number of alerted superhosts at any time after the start of patch dissemination. We introduce the   </text>
<query_num> 10107 </query_num>
<text>   ng for common patterns in network traffic or by analysing data and control flow of computer program executions. Automatic patch generation is addressed in Vigilante, which was proposed by Costa et al =-=[4]-=-, and motivates the work in this paper. Vigilante is an end-host based system for automatic worm containment. Its main feature is that when a host detects a worm, it generates a self-certifying alert   random scanning worm; see [7] for an analysis of the underlying causes. 2.2 Hosts, Superhosts, Subnets The use a centralized server to distribute patches to all clients is not scalable. Costa et al. =-=[4]-=- consider a fully decentralized peer-to-peer scheme where hosts are organized into a structured overlay over which alerts/patches are spread. It is not clear whether such a system will be universally  s work. Trust The requirements on trust relationships among hosts and superhosts are out of the scope of this work. Note, however, that this problem is largely eliminated in systems such as Vigilante =-=[4]-=-, as alerts are self-certifying. 2.4 Summary of results 2.4.1 Estimate of frequency of updates Simple analysis shows that in the absence of countermeasures, the characteristic timescale of epidemic sp d filter generation for some worms can be realized on rather fast timescales (in the order of tens of milliseconds), with a notably larger time to generate a filter for CodeRed worm (about 4 seconds) =-=[4]-=-. It remains a challenge to deploy and operate a real-world, planet-scale automatic patching system for heterogeneous hosts that will be sufficiently rapid to guarantee effective worm containment. The   </text>
<query_num> 10108 </query_num>
<text>   pective. 2.3 Patch dissemination The overlay connecting superhosts can be arbitrary. For example, it could be a balanced multicast tree or any one of several standard structured overlays (e.g. Pastry =-=[12]-=-, Chord [16] or CAN [11]) or an unstructured overlay (e.g. Gnutella). Our framework only requires knowing a lower bound on the number of alerted superhosts at any time after the start of patch dissemi ience to failures. The same order for broadcast time holds in probability for dissemination on overlays by random gossiping [10]. In our simulations, we used flooding on the structured overlay Pastry =-=[12]-=-, but we reiterate that our results hold more generally. Once a superhost is alerted, it starts sending the patch to hosts within its own subnet in addition to propagating the alert to other superhost It is given by (i) exponential function (truncated at 1) for a tree, and (ii) logistic function for hypercubes and for gossip-based dissemination. We also verify that when flooding alerts on a Pastry =-=[12]-=- overlay, the fraction of alerted superhosts over time is well approximated by a logistic function. We can use the minimum broadcast curve to obtain an upper bound on the fraction of hosts that eventu  in [2, Proposition 3]. 6. SIMULATION VALIDATION We verify some of our results by packet-level discreteevent simulations in SimPastry [3]. In our simulations, superhosts are nodes in a Pastry overlay =-=[12]-=-. The Pastry nodes are attached to J nodes chosen uniformly at random from a network topology that is input to the simulator. We used a transit-stub topology generated by the Georgia Tech topology gen   </text>
<query_num> 10109 </query_num>
<text>   roperty of alerts is important as it solves the problem of trust and the concomitant one of attacks using the automatic response system. A similar system was also proposed by Sidiroglou and Keromytis =-=[14]-=-. There has also beenworkonanalysingthecompetingprocessesofpatching, filtering and worm spread [13, 15, 18]. These works typically consider a ‘flat’ network for both the worm and patch processes, wher   </text>
<query_num> 10110 </query_num>
<text>   using the automatic response system. A similar system was also proposed by Sidiroglou and Keromytis [14]. There has also beenworkonanalysingthecompetingprocessesofpatching, filtering and worm spread =-=[13, 15, 18]-=-. These works typically consider a ‘flat’ network for both the worm and patch processes, whereas we study a hierarchical model motivated by considerations of scalability and trust. In practice, Micros citly. The interested reader is referred to [2] for the solution and limit points of w(t) and discussion. Comment. A somewhat similar host immunization process to (7)–(8) was considered by Wong et al =-=[18]-=-. They do not consider patch dissemination on an overlay, hence w(t) ≡ 1. The major distinction is that Wong et al (Section 6.1 [18]) assume a host can be patched in both infected and susceptible stat   </text>
</top>
<top>
<paper_num> 102 </paper_num>
<paper_title>   Extracting information from multimedia meeting collections.  </paper_title>
<paper_abstract>   Multimedia meeting collections, composed of unedited audio and video streams, handwritten notes, slides, and electronic documents that jointly constitute a raw record of complex human interaction processes in the workplace, have attracted interest due to the increasing feasibility of recording them in large quantities, by the opportunities for information access and retrieval applications derived from the automatic extraction of relevant meeting information, and by the challenges that the extraction of semantic information from real human activities entails. In this paper, we present a succint overview of recent approaches in this field, largely influenced by our own experiences. We first review some of the existing and potential needs for users of multimedia meeting information systems. We then summarize recent work on various research areas addressing some of these requirements. In more detail, we describe our work on automatic analysis of human interaction patterns from audio-visual sensors, discussing open issues in this domain.  </paper_abstract>
<query_num> 10201 </query_num>
<text>   Due to their conversational nature, speech in meetings often have low information content (compared to text documents), and many speech utterances relate to communication issues rather than to topics =-=[7]-=-. Various techniques have been adapted from text summarization in [45, 7, 29]. Other approaches conceptually related to summaries are those which attempt to detect meeting parts where participants are   </text>
<query_num> 10202 </query_num>
<text>   SL (audio-only) [8], ICSI (audioonly) [19], with a dialog act annotation extension [33], NIST (audio-visual) [35], M4 (audio-visual) [24], AMI (audio, video, slides, whiteboard and handwritten notes) =-=[9]-=-, and VACE (audio, video, and motion) [10]. These collections are at different stages of annotation and availability to the research community. 4. GROUP INTERACTION MODELING In this section, we briefl   </text>
<query_num> 10203 </query_num>
<text>   ], with a dialog act annotation extension [33], NIST (audio-visual) [35], M4 (audio-visual) [24], AMI (audio, video, slides, whiteboard and handwritten notes) [9], and VACE (audio, video, and motion) =-=[10]-=-. These collections are at different stages of annotation and availability to the research community. 4. GROUP INTERACTION MODELING In this section, we briefly describe our work on modeling of group i   </text>
<query_num> 10204 </query_num>
<text>   c decisions), context could be extracted both from text content (personal notes) and from information other than spoken words (audio and video). Context can take a number of forms, including location =-=[16]-=-, visual focus [37], addressee information [21], manifestations of emotional engagement like emphasis [22], and display of social signals like interest [15] and dominance [32]. The main challenge in a   </text>
<query_num> 10205 </query_num>
<text>   ext summarization in [45, 7, 29]. Other approaches conceptually related to summaries are those which attempt to detect meeting parts where participants are particularly engaged (called “hot-spots” in =-=[43]-=-). The relation between prosodic cues, dialog acts, and human-annotated hot-spots has been investigated, using speech utterances as basic units [43, 44]. We recently addressed a related task, namely t   </text>
<query_num> 10206 </query_num>
<text>   information content (compared to text documents), and many speech utterances relate to communication issues rather than to topics [7]. Various techniques have been adapted from text summarization in =-=[45, 7, 29]-=-. Other approaches conceptually related to summaries are those which attempt to detect meeting parts where participants are particularly engaged (called “hot-spots” in [43]). The relation between pros   </text>
<query_num> 10207 </query_num>
<text>   is second step could be designed to integrate a variable number of individual high-level data. From a more abstract modeling level, other problems are still open, and several of them are discussed in =-=[5]-=-. For instance, assuming each individual behavior is represented by a separate stream, and a single group model is used to incorporate all these streams, current modeling techniques, based on Markovia   </text>
<query_num> 10208 </query_num>
<text>   l in later sections, in the context of recent and current research projects on the subject [51, 50, 49]. A recent overview of meeting technologies, with different emphasis to the one here, appears in =-=[13]-=-. The paper is organized as follows. Section 2 reviews existing work on user requirements. Section 3 summarizes relevant tasks and work in the various directions of the field. Section 4 presents vario   </text>
<query_num> 10209 </query_num>
<text>   on, matching personal notes and audio-visual records, aligning references made in speech to documents, and creating richer text models combining text from written documents and speech for other tasks =-=[26, 30]-=-. 4. Context modeling. To reexamine and understand information about meeting key phases (e.g. discussions that led to specific decisions), context could be extracted both from text content (personal n   </text>
<query_num> 10210 </query_num>
<text>   peech in natural meetings is spontaneous and multiparty, containing disfluencies, no clear sentence boundaries, and significant overlapping, phenomena that constitute challenges for speech processing =-=[34]-=-, from automatic transcription (see [53] for the most recent NIST automatic speech recognition (ASR) evaluation on meeting data) to higherlevel tasks, like segmentation and classification of dialog ac   </text>
<query_num> 10211 </query_num>
<text>   some asynchrony exists for the defined group actions, and that such asynchrony is reasonably captured by the model. A recent comparison of the layered HMM and other models on the same task appears in =-=[1]-=-. 4.3 Modeling Influence During the course of meetings, some people seem particularly capable of driving the conversation and dominating its outcome. These people, skilled at establishing the leadersh   </text>
<query_num> 10212 </query_num>
<text>   the research community. 4. GROUP INTERACTION MODELING In this section, we briefly describe our work on modeling of group interest-level, group activities, and influence. More details can be found in =-=[15, 47, 48]-=-, respectively. 4.1 Modeling Group Interest-Level As discussed before, finding relevant segments in meetings is important for summarization, browsing, and retrieval purposes. In [15], we defined relev t make learning and inference intractable. This motivates the development of simplified models that at the same time retain representation power. We have recently proposed a two-level influence model =-=[48]-=-, which is a dynamic Bayesian network (DBN) with a two-level structure: the player level and the team level. The player level represents the actions of individual players, evolving based on their own   </text>
<query_num> 10213 </query_num>
<text>   the research community. 4. GROUP INTERACTION MODELING In this section, we briefly describe our work on modeling of group interest-level, group activities, and influence. More details can be found in =-=[15, 47, 48]-=-, respectively. 4.1 Modeling Group Interest-Level As discussed before, finding relevant segments in meetings is important for summarization, browsing, and retrieval purposes. In [15], we defined relev ting shares information, engages in discussions, and makes decisions, proceeding through diverse communication phases both in single meetings and during the course of long-term collaborative work. In =-=[47]-=-, we attempted to structure meetings into sequences of high-level items (dubbed multimodal speaker turns), using a multi-layer HMM framework (Figure 2). We defined two sets of actions: group actions,   </text>
<query_num> 10214 </query_num>
<text>   tions [23, 18, 42, 11]. Remote meetings are inserted in the large teleconferencing and computer-supported collaborative work (CSCW) domains, for which work on many aspects of user requirements exists =-=[14, 36]-=- but not discussed here due to lack of space. Finally, we are not aware of any comprehensive user requirements studies when users are professional external observers, probably due to the recent emerge   </text>
<query_num> 10215 </query_num>
<text>   trieval (IR), and human-computer interaction, can be tested and advanced [40, 28]. The third view considers meetings as a rich source of information with specific users and real needs to be satisfied =-=[23, 18, 42]-=-, and where technology for meeting analysis is relevant as long as it addresses and contributes to satisfy user needs. This paper starts from the last view, and aims at providing the unfamiliar reader eds with a multimedia recording system in mind (any combination of audio, video, handwritten notes, electronic documents, etc. in a single system) can be traced back to [41, 27], and more recently to =-=[23, 18, 42, 11]-=-. In this section, we discuss user requirements for meetings that occur in a workplace context, using two broad categories: user type and meeting type.suser meeting type use case user needs 1. previou design team), and people who, although might not regularly attend such meetings, have an interest in the group activities (e.g. a high-level manager monitoring the yearly progress of a specific team) =-=[23, 18, 42, 11]-=-. Briefly speaking, the users in this class have information needs related to information loss (“the failure to record important information, decisions and actions, and how this affects future actions ee of overlap (e.g. cases 1 and 3, 2 and 4, and 2 and 7). To further focus the discussion, in this paper we limit to review recent work on user requirement for local, pre-recorded meeting collections =-=[23, 18, 42, 11]-=-. Remote meetings are inserted in the large teleconferencing and computer-supported collaborative work (CSCW) domains, for which work on many aspects of user requirements exists [14, 36] but not discu  sections, (4) remember key people’s statements, (5) recall ideas not stored in public or personal records, and (6) verify cases in which memory and written records are inconsistent [18]. The work in =-=[42]-=- confirmed some of these findings, with an interesting distinction between public and personal records. Binding in nature, public records -written minutes- are mostly useful to track group progress, t   </text>
<query_num> 10216 </query_num>
<text>   trieval (IR), and human-computer interaction, can be tested and advanced [40, 28]. The third view considers meetings as a rich source of information with specific users and real needs to be satisfied =-=[23, 18, 42]-=-, and where technology for meeting analysis is relevant as long as it addresses and contributes to satisfy user needs. This paper starts from the last view, and aims at providing the unfamiliar reader eds with a multimedia recording system in mind (any combination of audio, video, handwritten notes, electronic documents, etc. in a single system) can be traced back to [41, 27], and more recently to =-=[23, 18, 42, 11]-=-. In this section, we discuss user requirements for meetings that occur in a workplace context, using two broad categories: user type and meeting type.suser meeting type use case user needs 1. previou design team), and people who, although might not regularly attend such meetings, have an interest in the group activities (e.g. a high-level manager monitoring the yearly progress of a specific team) =-=[23, 18, 42, 11]-=-. Briefly speaking, the users in this class have information needs related to information loss (“the failure to record important information, decisions and actions, and how this affects future actions ee of overlap (e.g. cases 1 and 3, 2 and 4, and 2 and 7). To further focus the discussion, in this paper we limit to review recent work on user requirement for local, pre-recorded meeting collections =-=[23, 18, 42, 11]-=-. Remote meetings are inserted in the large teleconferencing and computer-supported collaborative work (CSCW) domains, for which work on many aspects of user requirements exists [14, 36] but not discu pation in the actual meeting [42]. The study also higlighted the importance of looking at meetings not only from the singlemeeting view but also from the collection perspective. The approach taken in =-=[23]-=- used people placed in four scenarios -missed meeting, new employee, manager tracking project progress, and manager tracking employee performanceand generated queries for a hypothetical meeting retrie   </text>
<query_num> 10217 </query_num>
<text>   tural). Annotation at the individual action level was done by hand. Several configurations were compared, including audio-only, visual-only, early integration, multi-stream [12] and asynchronous HMMs =-=[3]-=-. Overall, the results suggest three main findings. First, the multi-layer HMM approach outperforms the single-layer one. Second, the use of AV features always outperforms the use of single modalities   </text>
<query_num> 10218 </query_num>
<text>   vailable information, adequate ways of interacting with media to browse and retrieve information from meetings are needed. A recent review discussing existing systems to access multimedia meetings is =-=[38]-=-. Resources. Part of the research summarized above has been conducted using a number of multimedia meeting collections, each of which varies with respect to the sensor setup, the type of recorded meet   </text>
<query_num> 10219 </query_num>
<text>   xt could be extracted both from text content (personal notes) and from information other than spoken words (audio and video). Context can take a number of forms, including location [16], visual focus =-=[37]-=-, addressee information [21], manifestations of emotional engagement like emphasis [22], and display of social signals like interest [15] and dominance [32]. The main challenge in all cases is modelin   </text>
<query_num> 10220 </query_num>
<text>   ype of existing annotations. Existing corpora include the ones by ISL (audio-only) [8], ICSI (audioonly) [19], with a dialog act annotation extension [33], NIST (audio-visual) [35], M4 (audio-visual) =-=[24]-=-, AMI (audio, video, slides, whiteboard and handwritten notes) [9], and VACE (audio, video, and motion) [10]. These collections are at different stages of annotation and availability to the research c o set empirical performance bounds based on human performance; and (3) a mechanism to merge the multiple annotator judgements into a single annotation. The annotation was carried out on the M4 corpus =-=[24]-=-, composed of 60 five-minute, four-participant meetings, which was recorded with three video cameras, a small circular 8-microphone array, and lapel microphones for each participant. We extracted a se rom the different annotators were merged, after observing that there was sufficient agreement among them. Regarding features, we extracted SRP-PHAT audio features to detect speaking turns in meetings =-=[24]-=-. Additionally,slanguage features were extracted from manual speech transcripts. We compared our model with a method based on the speaking length (the proportion of time during which each participant   </text>
</top>
<top>
<paper_num> 103 </paper_num>
<paper_title>   Semantic-based delivery of OLAP summary tables in wireless environments.  </paper_title>
<paper_abstract>   With the rapid growth in mobile and wireless technologies and the availability, pervasiveness and cost e#ectiveness of wireless networks, mobile computers are quickly becoming the normal front-end devices for accessing enterprise data. In this paper, we are addressing the issue of e#cient delivery of business decision support data in the form of summary tables to mobile clients equipped with OLAP front-end tools. Towards this, we propose a new on-demand scheduling algorithm, called SBS, that exploits both the derivation semantics among OLAP summary tables and the mobile clients&amp;apos; capabilities of executing simple SQL queries. It maximizes the aggregated data sharing between clients and reduces the broadcast length compared to the already existing techniques. The degree of aggregation can be tuned to control the tradeo# between access time and energy consumption. Further, the proposed scheme adapts well to different request rates, access patterns and data distributions. The algorithm e#ectiveness with respect to access time and power consumption is evaluated using simulation.  </paper_abstract>
<query_num> 10301 </query_num>
<text>   also makes an effective use of the low wireless bandwidth and clearly improves user perceived performance. Several scheduling algorithms have been proposed that attempt to achieve maximum aggregation =-=[2, 8, 21, 22]-=-. Considering the traditional OLAP server basic functionality, the broadcast pull or on-demand environment as shown in Figure 1, is the most suitable for supporting wireless OLAP query processing. Int   </text>
<query_num> 10302 </query_num>
<text>   e active mode. Power conservative indexing methods for single-attribute and multi-attribute based queries in broadcast push environments appeared in [14, 12, 3]. For hybrid broadcasts, the authors in =-=[7]-=- investigated two sets of broadcast protocols, where the requested data items are broadcast in batches and they used the techniques in [14] to index the data items on the broadcast cycle. The idea of   </text>
<query_num> 10303 </query_num>
<text>   e broadcast channel. If data is properly organized to cater to the needs of the clients, such a scheme makes an effective use of the low wireless bandwidth and is ideal to achieve maximal scalability =-=[1, 14, 12]-=-. In broadcast pull, the clients make explicit requests for data. If multiple clients request the same data at approximately the same time, the server may aggregate these requests, and only broadcast   </text>
<query_num> 10304 </query_num>
<text>   e broadcast channel. If data is properly organized to cater to the needs of the clients, such a scheme makes an effective use of the low wireless bandwidth and is ideal to achieve maximal scalability =-=[1, 14, 12]-=-. In broadcast pull, the clients make explicit requests for data. If multiple clients request the same data at approximately the same time, the server may aggregate these requests, and only broadcast  ming power orders of magnitude less than that in the active mode. Power conservative indexing methods for single-attribute and multi-attribute based queries in broadcast push environments appeared in =-=[14, 12, 3]-=-. For hybrid broadcasts, the authors in [7] investigated two sets of broadcast protocols, where the requested data items are broadcast in batches and they used the techniques in [14] to index the data ed to the non-preemptive one which does not consider heterogeneity. However, as mentioned above, the use of preemption deprives a broadcast policy from deploying an effective indexing technique as in =-=[14]-=- which is essential for energy saving. Of the preemptive scheduling algorithms, LTSF has a corresponding nonpreemptive version that retains its basic properties and can support selective tuning (in th   </text>
<query_num> 10305 </query_num>
<text>   ely used in materialized views selection. The objective is to select the appropriate set of tables for materialization so that to speed up future query processing, while meeting the space constraints =-=[11, 10]-=-. To facilitate the selection process, the search lattice was introduced in [11]. The search lattice is a directed graph representing the subcubes space and captures the subsumption property among sub   </text>
<query_num> 10306 </query_num>
<text>   ems on the broadcast cycle. The idea of merging queries with overlapping answers to reduce broadcast data dissemination cost has been introduced in the context of a multicast subscription environment =-=[6]-=-. In this approach, a post-filtering is needed at the client side to obtain the answer to the original query. A similar proposal appeared in [17], where a semantic description is attached to broadcast sed to broadcast the data chunks. This assumes that the server has been informed about the clients’ queries at the beginning of each broadcast cycle. Our work carries some similarity with the work in =-=[6, 17]-=-. However, we are modeling an on-demand broadcast environment, where the server has no prior knowledge of the arriving requests. Additionally, in our case, the requests are for summary tables (aggrega   </text>
<query_num> 10307 </query_num>
<text>   errors. We generated a synthesized lattice for an n-dimensional data cube. The values of n is in the range between 4 and 12, with n = 6 be the default. The sizes of lattice subcubes is computed as in =-=[15]-=-, where a subcube is given a binary code C. The binary code is similar to the bit encoding we used for identifying cubes on broadcast. Then the subcube size (number of tuples) is set to C 2 . The fina   </text>
<query_num> 10308 </query_num>
<text>   g [5]. The multidimensional data model abstracts data in the form of a data cube where dimensions are the subject of interests (aggregated attributes) and the cell values are the measures of interest =-=[11]-=-. An OLAP server may store multiple summary tables (subcubes) for efficient access by queries issued by OLAP tools at the client. An interesting property of summary tables which we call derivation dep ely used in materialized views selection. The objective is to select the appropriate set of tables for materialization so that to speed up future query processing, while meeting the space constraints =-=[11, 10]-=-. To facilitate the selection process, the search lattice was introduced in [11]. The search lattice is a directed graph representing the subcubes space and captures the subsumption property among sub mmary tables. We are assuming that all the lattice subcubes are ready at the server, which is a reasonable assumption, specially for relatively small size data marts. The Essbase system (according to =-=[11]-=-) is an example of commercial product that materialize all the possible summary tables. A client sends an uplink request for a table on the uplink channel. Then it listens to the downlink channel for  n the active mode. Picking a reasonable value for α will balance the trade-off between reducing the wait time (doze energy consumption) and increasing the tune time (active energy consumption). As in =-=[11]-=-, we are assuming a linear cost model for aggregate query processing, where a table scan is required to compute the result. Hence, when extra filtering and extraction is required, it can simply overla   </text>
<query_num> 10309 </query_num>
<text>   ovide poor access time for a broadcast pull environments [8] and Most Requests First (MRF) and Longest Wait First (LWF) were proposed as alternative efficient algorithms in [8, 22]. The RxW algorithm =-=[4]-=- combines the benefits of MRF and FCFS, where the intuition underlying RxW is that hot or popular data items are disseminated as soon as possible yet it avoids starvation of cold or less popular data   </text>
<query_num> 10310 </query_num>
<text>   y including it as part of the table descriptor information along with the cardinality of T bcast . A client can easily estimate the size of its own requested table T req using the simple formula from =-=[19]-=- which only requires knowledge of the number of distinct values for each dimension. This method is particularly efficient, especially considering that the growth rate of most OLAP dimensions is very l   </text>
</top>
<top>
<paper_num> 104 </paper_num>
<paper_title>   Locating faults through automated predicate switching.  </paper_title>
<paper_abstract>   Typically debugging begins when during a program execution a point is reached at which an obviously incorrect value is observed. A general and powerful approach to automated debugging can be based upon identifying modifications to the program state that will bring the execution to a successful conclusion. However, searching for arbitrary changes to the program state is difficult due to the extremely large search space. In this paper we demonstrate that by forcibly switching a predicate’s outcome at runtime and altering the control flow, the program state can not only be inexpensively modified, but in addition it is often possible to bring the program execution to a successful completion (i.e., program produces the desired output). By examining the switched predicate, also called the critical predicate, the cause of the bug can then be identified. Since the outcome of a branch can only be either true or false, the number of modified states resulting by predicate switching is far less than those possible through arbitrary state changes. Thus, it is possible to automatically search through modified states to find one that leads to the correct output. We have developed an implementation based upon dynamic instrumentation to perform this search through program re-execution – the program is executed from the beginning and a predicate’s outcome is switched to produce the desired change in control flow. To evaluate our approach, we tried our technique on several reported bugs for a number of UNIX utility programs. Our technique was found to be practical (i.e., acceptable in time taken) and effective (i.e., we were able to automatically identify critical predicates). Moreover we show that bidirectional dynamic slices of critical predicates capture the faulty code.  </paper_abstract>
<query_num> 10401 </query_num>
<text>   being explored by researchers. Some examples of such techniques include delta debugging [16, 7, 17, 15, 2], variants of backward dynamic slicing [13, 9, 1, 18, 19, 20, 21], and failure inducing chops =-=[3]-=-. Let us assume that on a given input we observe that the execution of a program fails. An aggressive and general approach to automated debugging is to run the program on this input again, interrupt t ing both dynamic data and control dependences during this run. Partitioning of predicates into high and low priority subsets is performed using a slicing based chopping algorithm that we presented in =-=[3]-=-. Here first we compute the backward dynamic slice (BS) of the erroneous output value. We also identify the failure inducing input using delta debugging technique [17] and compute the forward slice (F ducing the scope of potentially faulty code. The faulty code was captured by the bidirectional slice in all cases except fors-flex-v10. In prior work we introduced the notion of failure-inducing chop =-=[3]-=- which is obtained by intersecting the contents of the backward slice of an incorrect output value and the forward slice of the failureinducing input difference. This is another approximation of poten es can capture the faulty code, identifying the faulty code from the set of statements in the slice still requires nontrivial human effort. We further narrowed the scope of potentially faulty code in =-=[3]-=- by, for the first time, using forward dynamic slices of failure-inducing input difference. In contrast, in this paper, we have shown that bidirectional dynamic slices of critical predicates can furth   </text>
<query_num> 10402 </query_num>
<text>   ing input, critical predicate, and erroneous output. This information is useful to the programmer during debugging. Some additional works include the following. Xie et al. show that many redundancies =-=[14]-=- in programs correspond to hard program errors. Hangal et al. [4] identified the causes of some programming errors in Java programs by observing violations of program invariants. In [6], we developed   </text>
<query_num> 10403 </query_num>
<text>   ion is useful to the programmer during debugging. Some additional works include the following. Xie et al. show that many redundancies [14] in programs correspond to hard program errors. Hangal et al. =-=[4]-=- identified the causes of some programming errors in Java programs by observing violations of program invariants. In [6], we developed a technique that used a notion of path based weakest precondition   </text>
<query_num> 10404 </query_num>
<text>   nistic bug. For isolating nondeterministic bugs, they use statistical regression techniques to identify predicates that are highly correlated with the program failure. In contrast, Renieris and Reiss =-=[12]-=- focused on the difference between the failing run and a single passing run with similar spectra as a means to narrow down the search space for faulty code. Our work is complementary to the above work   </text>
<query_num> 10405 </query_num>
<text>   ogrammer. Therefore automated debugging techniques are being explored by researchers. Some examples of such techniques include delta debugging [16, 7, 17, 15, 2], variants of backward dynamic slicing =-=[13, 9, 1, 18, 19, 20, 21]-=-, and failure inducing chops [3]. Let us assume that on a given input we observe that the execution of a program fails. An aggressive and general approach to automated debugging is to run the program   </text>
<query_num> 10406 </query_num>
<text>   ogrammer. Therefore automated debugging techniques are being explored by researchers. Some examples of such techniques include delta debugging [16, 7, 17, 15, 2], variants of backward dynamic slicing =-=[13, 9, 1, 18, 19, 20, 21]-=-, and failure inducing chops [3]. Let us assume that on a given input we observe that the execution of a program fails. An aggressive and general approach to automated debugging is to run the program   a program which are contained in a specific statement such that fixing that statement fixes the program. 6. RELATED WORK Dynamic slicing was introduced as a aid to debugging [9, 1]. Our recent works =-=[19, 20]-=- have greatly reduced the space and time cost of dynamic slicing. In [21], we evaluated the effectiveness of backward dynamic slices in fault location. Our result showed that even though dynamic slice   </text>
<query_num> 10407 </query_num>
<text>   ogrammer. Therefore automated debugging techniques are being explored by researchers. Some examples of such techniques include delta debugging [16, 7, 17, 15, 2], variants of backward dynamic slicing =-=[13, 9, 1, 18, 19, 20, 21]-=-, and failure inducing chops [3]. Let us assume that on a given input we observe that the execution of a program fails. An aggressive and general approach to automated debugging is to run the program  ere can also be faults in a program which are contained in a specific statement such that fixing that statement fixes the program. 6. RELATED WORK Dynamic slicing was introduced as a aid to debugging =-=[9, 1]-=-. Our recent works [19, 20] have greatly reduced the space and time cost of dynamic slicing. In [21], we evaluated the effectiveness of backward dynamic slices in fault location. Our result showed tha   </text>
<query_num> 10408 </query_num>
<text>   redundancies [14] in programs correspond to hard program errors. Hangal et al. [4] identified the causes of some programming errors in Java programs by observing violations of program invariants. In =-=[6]-=-, we developed a technique that used a notion of path based weakest preconditions to automatically locate faulty code in a function when the precondition and postcondition of the function are availabl   </text>
<query_num> 10409 </query_num>
<text>   to make the task of finding bugs less tedious for the programmer. Therefore automated debugging techniques are being explored by researchers. Some examples of such techniques include delta debugging =-=[16, 7, 17, 15, 2]-=-, variants of backward dynamic slicing [13, 9, 1, 18, 19, 20, 21], and failure inducing chops [3]. Let us assume that on a given input we observe that the execution of a program fails. An aggressive a es [17, 16, 15], the delta debugging algorithm has been developed to automatically simplify or isolate a failureinducing input [17, 16], produce cause effect chains [15] and to link cause transitions =-=[2]-=- to the faulty code. In [2] delta debugging algorithm is used to analyze program state changes during the execution of a failed run to identify points of cause transitions. Codes-bash-2.05&amp;gt;cat input \ % ,&amp; 2 read.csexecuted at the points of cause transitions is expected to be relevant to the fault. Comparing and changing memory states of C program executions at a point is difficult due to pointers =-=[2]-=-. In addition, to identify points of cause transitions, the above state-based analysis has to be performed at a large number of points along the failed run. Therefore, program state based analysis is   </text>
<query_num> 10410 </query_num>
<text>   to make the task of finding bugs less tedious for the programmer. Therefore automated debugging techniques are being explored by researchers. Some examples of such techniques include delta debugging =-=[16, 7, 17, 15, 2]-=-, variants of backward dynamic slicing [13, 9, 1, 18, 19, 20, 21], and failure inducing chops [3]. Let us assume that on a given input we observe that the execution of a program fails. An aggressive a ng algorithm that we presented in [3]. Here first we compute the backward dynamic slice (BS) of the erroneous output value. We also identify the failure inducing input using delta debugging technique =-=[17]-=- and compute the forward slice (FS) of the failure inducing input. The predicate instances that belong to the intersection of the forward and backward slice (FS ∩ BS) form the subset that contains pre strated the use of postive evidence in form of correct portions of the outputs produced during a failing run to order and prune statements in the potentially faulty code [22]. In a series of articles =-=[17, 16, 15]-=-, the delta debugging algorithm has been developed to automatically simplify or isolate a failureinducing input [17, 16], produce cause effect chains [15] and to link cause transitions [2] to the faul   </text>
<query_num> 10411 </query_num>
<text>   to make the task of finding bugs less tedious for the programmer. Therefore automated debugging techniques are being explored by researchers. Some examples of such techniques include delta debugging =-=[16, 7, 17, 15, 2]-=-, variants of backward dynamic slicing [13, 9, 1, 18, 19, 20, 21], and failure inducing chops [3]. Let us assume that on a given input we observe that the execution of a program fails. An aggressive a strated the use of postive evidence in form of correct portions of the outputs produced during a failing run to order and prune statements in the potentially faulty code [22]. In a series of articles =-=[17, 16, 15]-=-, the delta debugging algorithm has been developed to automatically simplify or isolate a failureinducing input [17, 16], produce cause effect chains [15] and to link cause transitions [2] to the faul   </text>
</top>
<top>
<paper_num> 105 </paper_num>
<paper_title>   Labeling recursive workflow executions on-the-fly.  </paper_title>
<paper_abstract>   This paper presents a compact labeling scheme for answering reachability queries over workflow executions. In contrast to previous work, our scheme allows nodes (processes and data) in the execution graph to be labeled on-the-fly, i.e., in a dynamic fashion. In this way, reachability queries can be answered as soon as the relevant data is produced. We first show that, in general, for workflows that contain recursion, dynamic labeling of executions requires long (linearsize) labels. Fortunately, most real-life scientific workflows are linear recursive, and for this natural class we show that dynamic, yet compact (logarithmic-size) labeling is possible. Moreover, our scheme labels the executions in linear time, and answers any reachability query in constant time. We also show that linear recursive workflows are, in some sense, the largest class of workflows that allow compact, dynamic labeling schemes. Interestingly, the empirical evaluation, performed over both real and synthetic workflows, shows that our proposed dynamic scheme outperforms the state-of-the-art static scheme for large executions, and creates labels that are shorter by a factor of almost 3.  </paper_abstract>
<query_num> 10501 </query_num>
<text>   8, 12] was devoted to reduce the constant factor (2). The best known scheme [4] uses labels of log n + O( √ log n) bits, which is still separated from the known lower bound of log n+Ω(log log n) bits =-=[3]-=-. Motivated by the fact that XML trees are not deep, recent work [12] developed a scheme that uses labels of log n + 2 log d + O(1) bits, where d is the depth of the tree. (Workflow Runs) Workflow run xample 13. Returning to the running example, consider v5 and v16 in Figure 9. Since φg(v5) and φg(v16), shown in Example 12, share the first two common entries, moreover, φg(v5)[2].type = L (1) φg(v5)=-=[3]-=-.index = 1 &amp;lt; φg(v16)[3].index = 2 (2) by Algorithm 4, πg(φg(v5), φg(v16)) = true. Note that (1) implies that the least common ancestor of their context x7 and x12 is a special L node, and (2) implies   </text>
<query_num> 10502 </query_num>
<text>   cific permission and/or a fee. SIGMOD’11, June 12–16, 2011, Athens, Greece. Copyright 2011 ACM 978-1-4503-0661-4/11/06 ...$10.00. ity queries over large, graph-structured data, which can be expensive =-=[8]-=-. Reachability labels are an important tool for efficiently processing reachability queries on large graphs. The main idea is to assign each vertex a label such that, using only the label of any two v   </text>
<query_num> 10503 </query_num>
<text>   loops), in parallel (forked executions) or through recursion. Much research has been devoted recently to develop compact and efficient labeling schemes for workflow runs [6, 13] and graphs in general =-=[24, 16, 17, 15, 2, 9]-=-. A significant shortcoming, however, of the existing schemes is that they all need to examine the entire graph before labeling is performed. This may not be realistic in our setting since scientific  ability queries over large DAGs: Chain Decomposition [15], Tree Cover [2] and 2-Hop [9]. Other recent work includes Path-Tree [17] and 3-Hop [16] that combine the previous three approaches, and GRAIL =-=[24]-=- that is based on randomized interval labeling. 2 In particular, they are more general than series-parallel graphs. Dynamic. (Trees) The dynamic problem is harder than the static case; it was shown in   </text>
<query_num> 10504 </query_num>
<text>   loops), in parallel (forked executions) or through recursion. Much research has been devoted recently to develop compact and efficient labeling schemes for workflow runs [6, 13] and graphs in general =-=[24, 16, 17, 15, 2, 9]-=-. A significant shortcoming, however, of the existing schemes is that they all need to examine the entire graph before labeling is performed. This may not be realistic in our setting since scientific  s triggered several alternative approaches for efficiently answering reachability queries over large DAGs: Chain Decomposition [15], Tree Cover [2] and 2-Hop [9]. Other recent work includes Path-Tree =-=[17]-=- and 3-Hop [16] that combine the previous three approaches, and GRAIL [24] that is based on randomized interval labeling. 2 In particular, they are more general than series-parallel graphs. Dynamic. (   </text>
<query_num> 10505 </query_num>
<text>   loops), in parallel (forked executions) or through recursion. Much research has been devoted recently to develop compact and efficient labeling schemes for workflow runs [6, 13] and graphs in general =-=[24, 16, 17, 15, 2, 9]-=-. A significant shortcoming, however, of the existing schemes is that they all need to examine the entire graph before labeling is performed. This may not be realistic in our setting since scientific  the maximum label length is Ω(n) bits. This triggered several alternative approaches for efficiently answering reachability queries over large DAGs: Chain Decomposition [15], Tree Cover [2] and 2-Hop =-=[9]-=-. Other recent work includes Path-Tree [17] and 3-Hop [16] that combine the previous three approaches, and GRAIL [24] that is based on randomized interval labeling. 2 In particular, they are more gene   </text>
<query_num> 10506 </query_num>
<text>   ms Algorithms, Performance, Theory 1. INTRODUCTION Scientific workflow systems are now becoming“provenance aware” by automatically recording data and module dependency during execution (e.g., Taverna =-=[14]-=-, VisTrails [7] and Kepler [5]). By using such information, provenance queries such as “Was data item A (or Module M) used to produce data item B, either directly or indirectly?” are enabled. Answerin   </text>
<query_num> 10507 </query_num>
<text>   oduced, and cannot modify the labels subsequently. Our goal is thus to develop a dynamic labeling scheme for workflow runs. Dynamic labeling has been previously considered in the context of XML trees =-=[10, 20, 23]-=-, but workflow 1 We follow the standard assumption that any operation on two words (log n bits) can be done in constant time [6]. 1runs can have an arbitrarily more complex DAG structure 2 . Although f O(log n) bits. Other variant prefix-based schemes with similar bounds were also studied. e.g., ORDPATH [20], implemented in Microsoft SQL Server, supports frequent inserts in XML documents, and DDE =-=[23]-=- is tailored for both static and dynamic XML documents. (Workflow Runs and General DAGs) To our knowledge, the present work is the first to study dynamic labeling of workflow runs (and more generally   </text>
<query_num> 10508 </query_num>
<text>   oduced, and cannot modify the labels subsequently. Our goal is thus to develop a dynamic labeling scheme for workflow runs. Dynamic labeling has been previously considered in the context of XML trees =-=[10, 20, 23]-=-, but workflow 1 We follow the standard assumption that any operation on two words (log n bits) can be done in constant time [6]. 1runs can have an arbitrarily more complex DAG structure 2 . Although that is based on randomized interval labeling. 2 In particular, they are more general than series-parallel graphs. Dynamic. (Trees) The dynamic problem is harder than the static case; it was shown in =-=[10]-=- that labeling dynamic trees requires labels of Ω(n) bits. [10] also proposed a prefix-based scheme, which provides a matching upper bound of O(n) bits, and if the depth of the dynamic tree is bounded imum length of labels used by this scheme is n−1 bits, which matches the lower bound of Ω(n) bits in Theorem 1. In fact, this scheme can be used to label executions of arbitrary DAGs. It was shown in =-=[10]-=- that even labeling dynamic trees with n nodes requires labels of n−1 bits. Hence, we provide as a side benefit tight lower and upper bounds of n − 1 bits on the maximum label length for the general p   </text>
<query_num> 10509 </query_num>
<text>   orkflows, e.g., sequentially (loops), in parallel (forked executions) or through recursion. Much research has been devoted recently to develop compact and efficient labeling schemes for workflow runs =-=[6, 13]-=- and graphs in general [24, 16, 17, 15, 2, 9]. A significant shortcoming, however, of the existing schemes is that they all need to examine the entire graph before labeling is performed. This may not  ing has been previously considered in the context of XML trees [10, 20, 23], but workflow 1 We follow the standard assumption that any operation on two words (log n bits) can be done in constant time =-=[6]-=-. 1runs can have an arbitrarily more complex DAG structure 2 . Although there have been efficient dynamic algorithms [19, 11] for maintaining the transitive closure of DAGs, they all produce a linear t work [12] developed a scheme that uses labels of log n + 2 log d + O(1) bits, where d is the depth of the tree. (Workflow Runs) Workflow runs are modeled as DAGs derived from a given specification. =-=[6]-=- proposed a compact static scheme for labeling runs that uses labels of 3 log n + O(1) bits. However, it can only be applied to non-recursive workflows (with only loops and forks). [13] also proposed  luate the proposed dynamic labeling scheme over both real and synthetic workflows. Interestingly, our dynamic scheme creates even shorter labels for large runs than the state-of-the-art static scheme =-=[6]-=- by a factor of almost 3 (Section 7). 2. MODEL AND PROBLEM STATEMENT We start with notations and basic definitions over graphs in Section 2.1. An informal description of our workflow model is given in  of v5 and v11 with respect to x6. Hence, u4 ❀h3 v11 implies that v5 ❀g v11. 5. LABELING DYNAMIC WORKFLOWS WITH LINEAR RECURSION Our dynamic schemes are built upon a skeleton-based labeling framework =-=[6]-=-. As a preprocessing step, we label the workflow specification using any static reachability scheme, and then extend the reachability labels on the specification, called the skeleton labels, to label   </text>
<query_num> 10510 </query_num>
<text>   ory 1. INTRODUCTION Scientific workflow systems are now becoming“provenance aware” by automatically recording data and module dependency during execution (e.g., Taverna [14], VisTrails [7] and Kepler =-=[5]-=-). By using such information, provenance queries such as “Was data item A (or Module M) used to produce data item B, either directly or indirectly?” are enabled. Answering such queries entails evaluat   </text>
<query_num> 10511 </query_num>
<text>   th Static. (Trees) The earliest work for labeling static trees [22] proposed an interval-based scheme that uses labels of 2 log n bits, where n is the number of nodes in the tree. Considerable effort =-=[1, 4, 18, 12]-=- was devoted to reduce the constant factor (2). The best known scheme [4] uses labels of log n + O( √ log n) bits, which is still separated from the known lower bound of log n+Ω(log log n) bits [3]. M   </text>
<query_num> 10512 </query_num>
<text>   th Static. (Trees) The earliest work for labeling static trees [22] proposed an interval-based scheme that uses labels of 2 log n bits, where n is the number of nodes in the tree. Considerable effort =-=[1, 4, 18, 12]-=- was devoted to reduce the constant factor (2). The best known scheme [4] uses labels of log n + O( √ log n) bits, which is still separated from the known lower bound of log n+Ω(log log n) bits [3]. M  (if LCA(x, x ′ ) is a special R node or a non-special node) answer an equivalent query for their origin u and u ′ with respect to a small subgraph h. To encode Step (1), we use a prefix-based scheme =-=[18]-=- 3 to label t. To encode Step (2), we enrich a prefix-based label with skeleton labels as well as other necessary information (e.g., node types). The formal description of a reachability label is give rkflows are similar. Figure 20 reports the maximum label length. Observe that DRL creates shorter labels than SKL when the run size is larger than 1.5K. This is because DRL uses a prefix-based scheme =-=[18]-=- to label the explicit parse tree, while SKL uses an interval-based scheme [22]. The former performs better on 6 It turns out that the linear recursion in this workflow can be converted to a loop whic   </text>
<query_num> 10513 </query_num>
<text>   tion that any operation on two words (log n bits) can be done in constant time [6]. 1runs can have an arbitrarily more complex DAG structure 2 . Although there have been efficient dynamic algorithms =-=[19, 11]-=- for maintaining the transitive closure of DAGs, they all produce a linear-size index per vertex, which is unacceptable for large graphs. Nevertheless, we will show in this paper that the knowledge of   </text>
</top>
<top>
<paper_num> 106 </paper_num>
<paper_title>   Action Recognition from One Example.  </paper_title>
<paper_abstract>   Abstract—We present a novel action recognition method based on space-time locally adaptive regression kernels and the matrix cosine similarity measure. The proposed method uses a single example of an action as a query to find similar matches. It does not require prior knowledge about actions, foreground/background segmentation, or any motion estimation or tracking. Our method is based on the computation of novel space-time descriptors from the query video which measure the likeness of a voxel to its surroundings. Salient features are extracted from said descriptors and compared against analogous features from the target video. This comparison is done using a matrix generalization of the cosine similarity measure. The algorithm yields a scalar resemblance volume, with each voxel indicating the likelihood of similarity between the query video and all cubes in the target video. Using nonparametric significance tests by controlling the false discovery rate, we detect the presence and location of actions similar to the query video. High performance is demonstrated on challenging sets of action data containing fast motions, varied contexts, and complicated background. Further experiments on the Weizmann and KTH data sets demonstrate state-of-the-art performance in action categorization. Index Terms—Action recognition, space-time descriptor, correlation, regression analysis. Ç 1  </paper_abstract>
<query_num> 10601 </query_num>
<text>   many studies have attempted to tackle this problem and made impressive progress. Approaches can be categorized on the basis of action representation, namely, appearance-based representation [2], [3], =-=[4]-=-, [5], shape-based representation [6], [7], [8], [9], optical-flow-based representation [10], [11], [12], [13], interest-point-based representation [1], [14], [15], [16], [17], [18], and volume-based  tively, Cls computed from the local analysis window l are similar to one another in the motion-free region (see Fig. 6 [1]). On the other hand, in the region where motion exists (see Fig. 6 [2], [3], =-=[4]-=-, [5]), the kernel size and shape depend on both Cl and its space-time location xl in the local space-time window. Thus, if the pixel of interest (center pixel of kernel) is located in space-time edge   </text>
</top>
<top>
<paper_num> 107 </paper_num>
<paper_title>   Space-efficient construction of Lempel-Ziv compressed text indexes.  </paper_title>
<paper_abstract>   Abstract. A compressed full-text self-index is a data structure that replaces a text and in addition gives indexed access to it, while taking space proportional to the compressed text size. This is very important nowadays, since one can accommodate the index of very large texts entirely in main memory, avoiding the slower access to secondary storage. In particular, the LZ-index [G. Navarro, Journal of Discrete Algorithms, 2004] stands out for its good performance at extracting text passages and locating pattern occurrences. Given a text T[1..u] over an alphabet of size σ, the LZ-index requires 4|LZ|(1 + o(1)) bits of space, where |LZ | is the size of the LZ78-compression of T. This can be bounded by |LZ |  = uHk(T) + o(u log σ), where Hk(T) is the k-th order empirical entropy of T, for any k = o(log σ u). The LZ-index is built in O(u log σ) time, yet requiring O(u log u) bits of main memory in the worst case. In practice, the LZ-index occupies 1.0-1.5 times the text size (and replaces the text), but its construction requires around 5 times the text size. This limits its applicability to medium-sized texts. In this paper we present a space-efficient algorithm to construct the LZ-index in O(u(log σ + log log u)) time and requiring 4|LZ|(1 + o(1)) bits of main memory, that is, asymptotically the same space of the final index. We also adapt our algorithm to construct more recent reduced versions of the LZ-index, which occupy from 1 to 3 times |LZ|(1 + o(1)) bits, and show that these can also be built using asymptotically the same space of the final index. Finally, we study an alternative model in which we are given only a limited amount of main memory to carry out  </paper_abstract>
<query_num> 10701 </query_num>
<text>   , we have the representation of Geary et al. [22], the one of Sadakane and Navarro [67], and the one by Navarro [55, Section 6.1]. The latter has shown to be very effective for representing LZindexes =-=[56, 3]-=-. DFUDS Tree Representation To get this representation, named after Depth-First Unary Degree Sequence [8], we perform a preorder traversal on the tree, and for every node reached we write its degree i   </text>
<query_num> 10702 </query_num>
<text>   .n]. This structure can be represented with n log n + O(n) bits of space. We explain the simpler case, which is the one that arises in our work, where the points represent a permutation of {1,... ,n} =-=[43]-=-, i.e., there is exactly one point with first coordinate i for any 1 � i � n, and one point with second coordinate j for any 1 � j � n. To construct Range, we sort the set by the second coordinate j,  uctures for constant-time rank and select [51], which are needed to support the search (as well as, given a node, finding the corresponding starting position within the level, see Mäkinen and Navarro =-=[43]-=- for more details). Thus, we only need O(log n) pointers to represent the levels of the tree, avoiding in this way the need to store the pointers that represent the balanced tree. The total o(n log n) s data structure supports counting the number of points that lie within a two-dimensional range in O(log n) time, as well as reporting the occ points inside the search range in O((1 + occ)log n) time =-=[43]-=-. RNode. In the practical implementation of the LZ-index [55, 56], the Range data structure is replaced by RNode, which is a mapping from phrase identifiers to their node in RevTrie. After we free the   </text>
<query_num> 10703 </query_num>
<text>   = O((log log u) 1−ǫ ), the construction time can be made optimal, O(u). However, the space requirement to construct the CSA is still bigger than that needed by the final index. The work of Hon et al. =-=[29, 30]-=- deal with the space (and time) efficient construction of the CSA. The former uses (2H0(T)+1+ǫ)u+o(ulog σ) bits of space to build the CSA, where ǫ is any positive constant. The construction time is O(  Genome, and used it to construct the Compressed Suffix Array. They used an IBM SP-2 (450MHz CPU) with 64GB of RAM to achieve 7 hours of indexing time. The indexing space was about 12GB. – Hon et al. =-=[29, 28]-=- indexed the Human Genome with the CSA in about 24 hours, using a Pentium IV processor at 1.7 GHz with 512 KB of L2 cache, and 4 GB of main memory, running Solaris 9 operating system. They also constr   </text>
<query_num> 10704 </query_num>
<text>   = O((log log u) 1−ǫ ), the construction time can be made optimal, O(u). However, the space requirement to construct the CSA is still bigger than that needed by the final index. The work of Hon et al. =-=[29, 30]-=- deal with the space (and time) efficient construction of the CSA. The former uses (2H0(T)+1+ǫ)u+o(ulog σ) bits of space to build the CSA, where ǫ is any positive constant. The construction time is O( xes. In all cases ǫ stands for any positive (and usually small) value. Index Indexing space (in bits) Indexing time Suffix Array (SA) [21] u log u O(u log u) SA [31] O(u log σ) (*) O(u log log σ) CSA =-=[30]-=- u(H0(T) + 2 + ǫ) + o(u log σ) (†) O(u log u) CSA [54] O(u log σ(logσ u) log3 2 ) (*) O(u) log σ log log u )) AF-FMI [24] uHk(T) + o(u log σ) (§) O(u log u(1 + LZ-index (original) [55, 56] O(u log u)   </text>
<query_num> 10705 </query_num>
<text>   a lower bound on the performance of statistical compressors based on predicting the next text symbol as a function of the k preceding ones. A separate track of indexes based on Lempel-Ziv compression =-=[72, 73]-=- was pursued in parallel to the research on compressing suffix arrays. These are generally called LZ-indexes [36, 55, 18, 64, 7]. Except for the first pioneering work [36], all the rest are self-index   </text>
<query_num> 10706 </query_num>
<text>   a lower bound on the performance of statistical compressors based on predicting the next text symbol as a function of the k preceding ones. A separate track of indexes based on Lempel-Ziv compression =-=[72, 73]-=- was pursued in parallel to the research on compressing suffix arrays. These are generally called LZ-indexes [36, 55, 18, 64, 7]. Except for the first pioneering work [36], all the rest are self-index ng any compressor that encodes each symbol considering only the context of k symbols that precede it in T . 2.3 Lempel-Ziv Compression The Lempel-Ziv compression algorithm of 1978 (usually named LZ78 =-=[73]-=-) is based on a dictionary of phrases, in which we add every new phrase computed. At the beginning of the compression, the dictionary contains a single phrase b0 of length 0 (i.e., the empty string).  m is O(u) time in the worst case and efficient in practice provided we use the LZTrie, which allows rapid searching of the new text prefix (for each symbol of T we move once in the trie). Property 3 (=-=[73]-=-). It holds that √ u � n � u logσ u . This implies log n = Θ(log u) and n log u � ulog σ. We shall use the following result of Kosaraju and Manzini [38] to bound the size of the output of the LZ78 par   </text>
<query_num> 10707 </query_num>
<text>   an be constructed within the same space required by the final indexes. In all cases ǫ stands for any positive (and usually small) value. Index Indexing space (in bits) Indexing time Suffix Array (SA) =-=[21]-=- u log u O(u log u) SA [31] O(u log σ) (*) O(u log log σ) CSA [30] u(H0(T) + 2 + ǫ) + o(u log σ) (†) O(u log u) CSA [54] O(u log σ(logσ u) log3 2 ) (*) O(u) log σ log log u )) AF-FMI [24] uHk(T) + o(u   </text>
<query_num> 10708 </query_num>
<text>   as a significant impact on the running time of an application, as accesses to secondary memory are orders of magnitude slower. Several attempts to reduce the space of the suffix arrays have been made =-=[41, 26, 65, 18, 25, 42, 19]-=-. They aim at compressed indexing, which takes advantage of the regularities of the text to operate in space proportional to that of the compressed text (i.e., c times the size of the text compressed  important breakthrough. The main families of self-indexes based on suffix arrays [57] are the Compressed Suffix Arrays (CSAs for short) [65, 25] and FM-indexes (for “Full-text index in Minute space”) =-=[18, 42, 19]-=-. The latter compress suffix arrays via the Burrows-Wheeler Transform [10]. The compressibility in both families is usually measured in terms of the k-th order empirical entropy, Hk, which is a lower   </text>
<query_num> 10709 </query_num>
<text>   as a significant impact on the running time of an application, as accesses to secondary memory are orders of magnitude slower. Several attempts to reduce the space of the suffix arrays have been made =-=[41, 26, 65, 18, 25, 42, 19]-=-. They aim at compressed indexing, which takes advantage of the regularities of the text to operate in space proportional to that of the compressed text (i.e., c times the size of the text compressed  lf-indexes do not consider the space-efficient construction of the indexes. Yet, this aspect becomes crucial when implementing the index in practice. For example, the original construction of the CSA =-=[26, 65]-=- and FM-index [18] involves building first the suffix array of the text, using for example the algorithm of Larsson and Sadakane [40] or the one by Manzini and Ferragina [48]. Similarly, Navarro’s LZ- also sample ǫn values of R in such a way that the computation of R[i] (by means of ϕ) takes O(1/ǫ) time in the worst case. Function ϕ has the same properties as function Ψ of Compressed Suffix Arrays =-=[26, 65]-=-, thus this can be also compressed to n log σ + O(n log log σ) bits of space (in this paper we show how to compress it to n log σ + O(n) bits and still compute any entry in constant time). The computa equirement raises to (1 + 3ǫ)n log n + 4n log σ + O(n) bits. 8. We use the approach of Chan et al. [11] to construct ϕ, which is originally defined for building function Ψ of Compressed Suffix Arrays =-=[26, 65]-=- requiring only O(ulog σ) bits of space. In our case we compute ϕ[i] = R−1 (parentlz(R[i])) for consecutive i values, each in time O(1/ǫ) as we have R stored in plain form and R−1 represented with the   </text>
<query_num> 10710 </query_num>
<text>   as a significant impact on the running time of an application, as accesses to secondary memory are orders of magnitude slower. Several attempts to reduce the space of the suffix arrays have been made =-=[41, 26, 65, 18, 25, 42, 19]-=-. They aim at compressed indexing, which takes advantage of the regularities of the text to operate in space proportional to that of the compressed text (i.e., c times the size of the text compressed  placing it, and providing efficient indexed access to it, is an important breakthrough. The main families of self-indexes based on suffix arrays [57] are the Compressed Suffix Arrays (CSAs for short) =-=[65, 25]-=- and FM-indexes (for “Full-text index in Minute space”) [18, 42, 19]. The latter compress suffix arrays via the Burrows-Wheeler Transform [10]. The compressibility in both families is usually measured   </text>
<query_num> 10711 </query_num>
<text>   ebuilt from scratch upon insertions, we represent each block by using dynamic data structures, which can be updated in time less than linear in the block size. We adapt the approach used by Arroyuelo =-=[2]-=- to represent succinct dynamic σ-ary trees: We first reduce the size of the problem by dividing the trie into small blocks, and then represent every block (i.e., smaller trie) with a dynamic data stru because the new node is inserted with no related inter-block pointer) at the corresponding position (given by preorderp). This data structure adds n + o(n) extra bits to our representation. Arroyuelo =-=[2]-=- gives a more involved representation for Fp, requiring o(n) bits, yet the one we are using here is simpler and still adequate for our purposes. Representation of the Symbols, lettsp We represent the  the efficient search of this position. The required functionality is easily achieved by regarding the vector of N 5 The space requirement of the trie topology can be reduced to 2n + o(n) bits overall =-=[2, 67]-=-. However, O(n) bits is sufficient for our purposes.idsp values, each of width t, as a bitmap of length tN. The dynamic data structure for bitmaps of Mäkinen and Navarro [44] would easily permit inse . Instead, we will look for z in advance to overflows, by looking for possible candidates in the insertion path of new nodes. To quickly select node z, we maintain in each block p a candidate list Cp =-=[2]-=-, storing the local preorders of the nodes that can be copied to a new child block p ′ upon block overflow. With selectnode we can obtain the candidate node corresponding to such a preorder. A subtree   </text>
<query_num> 10712 </query_num>
<text>   econdary memory for the construction is nowadays the most practical alternative [15]. Another research path is to try building the suffix array directly in compressed space in main memory. Hon et al. =-=[31]-=- present an algorithm to construct suffix arrays (and also suffix trees) using O(ulog σ) bits of storage, in O(ulog log σ) = o(ulog u) time for suffix arrays, and O(u(log ǫ u + log σ)) time for suffix e same space required by the final indexes. In all cases ǫ stands for any positive (and usually small) value. Index Indexing space (in bits) Indexing time Suffix Array (SA) [21] u log u O(u log u) SA =-=[31]-=- O(u log σ) (*) O(u log log σ) CSA [30] u(H0(T) + 2 + ǫ) + o(u log σ) (†) O(u log u) CSA [54] O(u log σ(logσ u) log3 2 ) (*) O(u) log σ log log u )) AF-FMI [24] uHk(T) + o(u log σ) (§) O(u log u(1 + L   </text>
<query_num> 10713 </query_num>
<text>   edge symbols), plus an array that, for any phrase identifier i, stores the preorder of the corresponding LZTrie node. Using our notation, the latter is just array ids−1 . (log log u) 2 Jansson et al. =-=[33]-=- propose an algorithm to construct the parsing in O( u logσ u log log log u ) time and requiring uHk(T) + o(ulog σ) bits of space. The algorithm, however, needs two passes over the text, which involve compute ids −1 in place, using the algorithm of Lemma 3, and this way we complete the representation for the LZ78 parsing of text T . As seen, we exchange the |T | bits of extra I/O of Jansson et al. =-=[33]-=- by 2|LZ|. This can be much better in the case of large compressible texts. The total time is O(u(log σ + log log u)), and the maximum mainmemory space used is |LZ|(1 + o(1)) bits. We think that our m   </text>
<query_num> 10714 </query_num>
<text>   ets the size of the subtree of node x, including x itself), degree(x) (which gets the degree, i.e., the number of children, of node x), childrank(x) (which gets the rank of node x within its siblings =-=[34]-=-), and ancestor(x,y) (which tells us whether node x is an ancestor of node y), all in O(1) time. If we assume that par represents the DFUDS sequence of the tree, then we have: parent(x) ≡ select )(par ose(par,select )(par,rank )(par,x) + 1) − i) + 1Operation depth(x) (which gets the depth of node x in the tree) can also be computed in constant time on DFUDS by using the approach of Jansson et al. =-=[34]-=-, requiring o(n) extra bits. Given a node in this representation, say at position i, its preorder position can be computed by counting the number of closing parentheses before position i; in other wor he first |X| parent operations can be executed with a single operation called level-ancestor, which can be executed in constant time using o(n) extra bits on top of the LZTrie topology representation =-=[34]-=-. Thus the overall time is O(|Y | − |X|) = O(ℓ). Since the total amount of skips traversed along the construction process is u, computing the skips in this log σ log σ way adds O(u(1+ log log u )) to   </text>
<query_num> 10715 </query_num>
<text>   h 0 (i-th 1) in B. We assume that select0(B,0) always equals 0 (similarly for select1). These operations can be supported in constant time and requiring n + o(n) bits [51], or even nH0(B) + o(n) bits =-=[62]-=-. The o(n) overhead can be made as small as O(n/log c n) for any constant c [61]. There exist a number of practical data structures supporting rank and select, like the one by González et al. [23], Ki elo et al. [7] add a bit vector TPos marking the phrase beginnings, which is then represented with a data structure for rank and select and requiring n log u n + O(n) + o(u) = o(ulog σ) bits of space =-=[62]-=-. A more practical approach [5] consists in sampling the starting positions of some phrases, and then representing the starting position of every other phrase as an offset from the previous sampled ph   </text>
<query_num> 10716 </query_num>
<text>   h problem plays a fundamental role in modern computer applications. Text Compression and Indexing. Despite that there has been some work on space-efficient inverted indexes for natural language texts =-=[71, 58]-=- (able to find whole words and phrases), until one decade ago it was ⋆ A preliminary partial version of this paper appeared in Proc. ISAAC 2005, pp. 1143–1152. ⋆⋆ Funded by CONICYT PhD Fellowship Prog   </text>
<query_num> 10717 </query_num>
<text>   in practice. For example, the original construction of the CSA [26, 65] and FM-index [18] involves building first the suffix array of the text, using for example the algorithm of Larsson and Sadakane =-=[40]-=- or the one by Manzini and Ferragina [48]. Similarly, Navarro’s LZ-index is constructed over a non-compressed intermediate representation [55]. In both cases one needs in practice about 5 times the te   </text>
<query_num> 10718 </query_num>
<text>   log u)), and the maximum mainmemory space used is |LZ|(1 + o(1)) bits. We think that our methods could be extended to build related LZ-indexes [18, 64] within limited space. Finally, recent advances =-=[27, 59]-=- (not all refereed yet) seem to indicate that it is possible to handle log n all the classical operations on a tree of n nodes within 2n + o(n) bits and O( ) time; and that a log log n dynamic sequenc   </text>
<query_num> 10719 </query_num>
<text>   ng a sequence of balanced parentheses: the succinct representation of general trees, with the so-called BP representation. Among the practical alternatives, we have the representation of Geary et al. =-=[22]-=-, the one of Sadakane and Navarro [67], and the one by Navarro [55, Section 6.1]. The latter has shown to be very effective for representing LZindexes [56, 3]. DFUDS Tree Representation To get this re   </text>
<query_num> 10720 </query_num>
<text>   nly in its cycle: π −1 (j) is just the value “pointing” to j in the diagram. To compute π −1 (13) in our example, we start at position 13, then move to position π(13) = 7, then to π(7) = 12, then to π=-=(12)-=- = 2, then to π(2) = 17, and as π(17) = 13 we conclude that π −1 (13) = 17. Since there are no bounds for the size of a cycle, this takes O(n) time in the worst case. Yet, it can be improved for a mor   </text>
<query_num> 10721 </query_num>
<text>   nstruction of the CSA [26, 65] and FM-index [18] involves building first the suffix array of the text, using for example the algorithm of Larsson and Sadakane [40] or the one by Manzini and Ferragina =-=[48]-=-. Similarly, Navarro’s LZ-index is constructed over a non-compressed intermediate representation [55]. In both cases one needs in practice about 5 times the text size (in the case of the CSA and the F imes the text size for English text, and 3.4–3.7 times the text size for DNA. As a comparison, the construction of a plain suffix array without any extra data structure requires 5 times the text size =-=[48]-=-. 3.4 Reduced Space Versions of the LZ-index New versions of the LZ-index have been introduced recently [6, 7, 5], which require less space than the original LZ-index, in some cases also improving its   </text>
<query_num> 10722 </query_num>
<text>   ough. Defining the Block Layout Each block p of N nodes consists of: – The representation Tp of the topology of the block, using any suitable tree representation. In particular, we will use the DFUDS =-=[8]-=-, which is particularly well suited for our goals. – A bit-vector Fp[1..N] (the flags) such that Fp[j] = 1 iff the j-th node of Tp (in preorder) has an associated inter-block pointer. We shall represe  following way. Representation of the Trie Topology, Tp To represent the trie topology of block p we use the data structure for dynamic balanced parentheses of Chan et al. [11] to represent the DFUDS =-=[8]-=- of the block. The main idea of Chan et al. is to divide the original parentheses sequence into segments Si of O(log N) bits. Every segment Si is stored in the leaves of a balanced binary tree T ′ p , ost n ′/2 � n, by not storing the skips of the leaf nodes. As we see soon, these will not be necessary. The topology representation rpar allows one to count the number of leaves to the left of a node =-=[8]-=-, so that we can index into the reduced array skips. Note that each skip may be as large as u. However, as they are at most n and add up to at most u, we can set up a bitmap S[1..u] where we write eac   </text>
<query_num> 10723 </query_num>
<text>   perated entirely in RAM on a modest desktop computer), but 15 GB of main memory are needed to build the indexes! Using secondary memory for the construction is nowadays the most practical alternative =-=[15]-=-. Another research path is to try building the suffix array directly in compressed space in main memory. Hon et al. [31] present an algorithm to construct suffix arrays (and also suffix trees) using O pace required by the uncompressed genome (assuming the base pairs are represented by bytes), and also in less than 5 hours. This is competitive with the best current algorithms to build suffix arrays =-=[15]-=-. 2 Table Preliminary 1 summarizes Concepts the results obtained in this paper and compares with existing approaches. 2.1 Model of Computation We assume the standard word RAM model of computation, in  , 100 and 500 files. We used the same construction parameters as in the original article [32]. The program was run in our machine. – The algorithm for constructing suffix arrays from Dementiev et al. =-=[15]-=-. Most of the work of this algorithm is carried out on secondary storage, using just a constant amount of main memory. Therefore its performance depends basically on the speed of the disk used, wherea her than our memory usage. Hence, the indexing space can be reduced to approach ours, yet at the price of degrading much the indexing time. Using computer (i) above, the algorithm of Dementiev et al. =-=[15]-=- indexes the Human Genome in about 8.52 hours, using secondary storage and just a constant amount of main memory. By using computer (ii), on the other hand, the indexing times are reduced to 5.11 hour ] 4.98 hours 2,038 MB Run-length Compressed Suffix Arrays – 100 files [32] 6.33 hours 1,904 MB Run-length Compressed Suffix Arrays – 500 files [32] 18.79 hours 1,799 MB Suffix array – on computer (i) =-=[15]-=- 8.52 hours 1,024 MB Suffix array – on computer (ii) [15] 5.11 hours 1,024 MB Scheme 2 of LZ-index This paper 4.63 hours 2,847 MB Scheme 2 – reduced-memory model This paper 4.63 hours 1,597 MB As a hi   </text>
<query_num> 10724 </query_num>
<text>   quires to operate. This size can be compared with that of compressed suffix array via the (not always tight) upper bound |LZ| � uHk(T) + o(ulog σ). At the time of the preliminary version of this work =-=[4]-=-, this was the first construction algorithm for a compressed self-index requiring space proportional to Hk(T) instead of H0(T). Recently, however, a construction algorithm for the socalled Alphabet Fr ce, and O(ulog ulog σ) time [44], and even O(u ) [24]. Yet, the time obtained in the present paper is far better, and it also improves significantly upon the O(σu) worst-case time of our early result =-=[4]-=-. We show how the reduced-space versions of the LZ-index [6, 5, 7] can similarly be constructed within asymptotically the space required by the final index. We also present an alternative model to con log u) CSA [54] O(u log σ(logσ u) log3 2 ) (*) O(u) log σ log log u )) AF-FMI [24] uHk(T) + o(u log σ) (§) O(u log u(1 + LZ-index (original) [55, 56] O(u log u) O(u log σ) LZ-index (our early result) =-=[4]-=- (4 + ǫ)uHk(T) + o(u log σ) (‡) O(σu) LZ-index (this paper) 4uHk(T) + o(u log σ) (‡) O(u(log σ + log log u)) Reduced LZ-index a (this paper) (1 + ǫ)uHk(T) + o(u log σ) (‡) O(u(log σ + log log u)) Redu  and dynamic sequences [44] to create the tries. However, the construction time would become at best O(ulog n(1 + log σ log log n )).Our early space-efficient construction algorithm for the LZ-index =-=[4]-=- partitions the tree into moderatelysized connected components, which are updated in naive form. As a result, it has a construction time of the form O(σu), which is impractical for moderately-large al t of any (non-fictitious) node is stored in the same block of the node, but also that all its sibling nodes are stored in the same block. Rather than using a static representation for the trie blocks =-=[4]-=-, which are rebuilt from scratch upon insertions, we represent each block by using dynamic data structures, which can be updated in time less than linear in the block size. We adapt the approach used   </text>
<query_num> 10725 </query_num>
<text>   s of the text to operate in space proportional to that of the compressed text (i.e., c times the size of the text compressed under some model, for some constant c). Especially, in some of those works =-=[65, 18, 25, 55, 42, 19, 64, 7]-=- the indexes replace the text and, using little space (sometimes even less than that of the original text), provide indexed access. This feature is known as self-indexing, since the index allows one t  the k preceding ones. A separate track of indexes based on Lempel-Ziv compression [72, 73] was pursued in parallel to the research on compressing suffix arrays. These are generally called LZ-indexes =-=[36, 55, 18, 64, 7]-=-. Except for the first pioneering work [36], all the rest are self-indexes and based on the Lempel-Ziv compression algorithm of 1978 (LZ78) [73]. Their space performance is measured in terms of the ou  since they have shown to be very effective in practice for extracting text, displaying occurrence contexts, and locating many occurrences, outperforming suffixarray-based self-indexes at these tasks =-=[56, 64, 5, 17]-=-. In theory, only LZ-indexes achieve high-order entropy space together with O(log u) worst-case time per located occurrence. Moreover, in practice many pattern occurrences can be actually found in con ge compressible texts. The total time is O(u(log σ + log log u)), and the maximum mainmemory space used is |LZ|(1 + o(1)) bits. We think that our methods could be extended to build related LZ-indexes =-=[18, 64]-=- within limited space. Finally, recent advances [27, 59] (not all refereed yet) seem to indicate that it is possible to handle log n all the classical operations on a tree of n nodes within 2n + o(n)   </text>
<query_num> 10726 </query_num>
<text>   s of the text to operate in space proportional to that of the compressed text (i.e., c times the size of the text compressed under some model, for some constant c). Especially, in some of those works =-=[65, 18, 25, 55, 42, 19, 64, 7]-=- the indexes replace the text and, using little space (sometimes even less than that of the original text), provide indexed access. This feature is known as self-indexing, since the index allows one t  the k preceding ones. A separate track of indexes based on Lempel-Ziv compression [72, 73] was pursued in parallel to the research on compressing suffix arrays. These are generally called LZ-indexes =-=[36, 55, 18, 64, 7]-=-. Except for the first pioneering work [36], all the rest are self-indexes and based on the Lempel-Ziv compression algorithm of 1978 (LZ78) [73]. Their space performance is measured in terms of the ou  with O(log u) worst-case time per located occurrence. Moreover, in practice many pattern occurrences can be actually found in constant time. In particular, we will be interested in Navarro’s LZindex =-=[55, 56]-=- and its more recent variants [6, 7, 5]. Compressed Construction of Self-Indexes. Many works on compressed full-text self-indexes do not consider the space-efficient construction of the indexes. Yet,  , using for example the algorithm of Larsson and Sadakane [40] or the one by Manzini and Ferragina [48]. Similarly, Navarro’s LZ-index is constructed over a non-compressed intermediate representation =-=[55]-=-. In both cases one needs in practice about 5 times the text size (in the case of the CSA and the FM-index, by using the deep-shallow algorithm [48]). For example, the Human Genome 3 log x means ⌈log2 s very important for self-indexes), their spaceefficient construction is certainly an important issue. Our Contribution. We present a practical and efficient algorithm to construct Navarro’s LZ-index =-=[55, 56]-=- using little space. Our idea is to replace, at construction time, the (space-inefficient) intermediate representations of the tries that conform the index by space-efficient counterparts. Basically,   </text>
<query_num> 10727 </query_num>
<text>   s of the text to operate in space proportional to that of the compressed text (i.e., c times the size of the text compressed under some model, for some constant c). Especially, in some of those works =-=[65, 18, 25, 55, 42, 19, 64, 7]-=- the indexes replace the text and, using little space (sometimes even less than that of the original text), provide indexed access. This feature is known as self-indexing, since the index allows one t  the k preceding ones. A separate track of indexes based on Lempel-Ziv compression [72, 73] was pursued in parallel to the research on compressing suffix arrays. These are generally called LZ-indexes =-=[36, 55, 18, 64, 7]-=-. Except for the first pioneering work [36], all the rest are self-indexes and based on the Lempel-Ziv compression algorithm of 1978 (LZ78) [73]. Their space performance is measured in terms of the ou ated occurrence. Moreover, in practice many pattern occurrences can be actually found in constant time. In particular, we will be interested in Navarro’s LZindex [55, 56] and its more recent variants =-=[6, 7, 5]-=-. Compressed Construction of Self-Indexes. Many works on compressed full-text self-indexes do not consider the space-efficient construction of the indexes. Yet, this aspect becomes crucial when implem the time obtained in the present paper is far better, and it also improves significantly upon the O(σu) worst-case time of our early result [4]. We show how the reduced-space versions of the LZ-index =-=[6, 5, 7]-=- can similarly be constructed within asymptotically the space required by the final index. We also present an alternative model to construct the indexes, in which we assume that the available main mem aining index components. Then we build the final representation of the topology of LZTrie, bitmap par, using the parentheses representation of Munro and Raman [53], yet newer versions of the LZ-index =-=[7]-=- use the DFUDS representation [8]. We also create the array ids[1..n], storing the LZ78 phrase identifiers in preorder, and letts[1..n], storing the symbols that label the trie edges, in preorder. The   </text>
<query_num> 10728 </query_num>
<text>   sk. Thus, the maximum main-memory space requirement to construct Range is n log n + O(n) bits. At the end we have only O(n) bits of main memory in use. The amount of I/O is 4n log n + O(n) bits. Step =-=(6)-=- We build Node from ids, by traversing LZTrie in preorder. In this way, array ids is sequentially traversed, while Node is randomly accessed. Thus, we allocate n log 2n bits of space for Node, and mai   </text>
<query_num> 10729 </query_num>
<text>   the Patricia-tree skips, as their space consumption is problematic. Instead, we use the following procedure to find out in O(ℓ) time the skip value ℓ of the edge leading to a node y from its parent x =-=[50]-=-. Let X and Y be the strings labeling the paths from the root of the reverse trie to nodes x and y, respectively, then ℓ = |Y | − |X|. We find the leftmost and rightmost leaves v1 r and v2 r descendin   </text>
<query_num> 10730 </query_num>
<text>   the construction. 2.2 Empirical Entropy A concept related to text compression is that of the k-th order empirical entropy of a sequence of symbols T[1..u] over an alphabet of size σ, denoted by Hk(T) =-=[47]-=-. The value uHk(T) provides a lower bound to the number of bits needed to compress T using any compressor that encodes each symbol considering only the context of k symbols that precede it in T . 2.3   </text>
<query_num> 10731 </query_num>
<text>   the succinct representation of general trees, with the so-called BP representation. Among the practical alternatives, we have the representation of Geary et al. [22], the one of Sadakane and Navarro =-=[67]-=-, and the one by Navarro [55, Section 6.1]. The latter has shown to be very effective for representing LZindexes [56, 3]. DFUDS Tree Representation To get this representation, named after Depth-First  the efficient search of this position. The required functionality is easily achieved by regarding the vector of N 5 The space requirement of the trie topology can be reduced to 2n + o(n) bits overall =-=[2, 67]-=-. However, O(n) bits is sufficient for our purposes.idsp values, each of width t, as a bitmap of length tN. The dynamic data structure for bitmaps of Mäkinen and Navarro [44] would easily permit inse   </text>
<query_num> 10732 </query_num>
<text>   uction algorithm for the socalled Alphabet Friendly FM-index (AF-FMI) [19] has appeared, requiring uHk(T) + o(ulog σ) bits of log u log σ log log u space, and O(ulog ulog σ) time [44], and even O(u ) =-=[24]-=-. Yet, the time obtained in the present paper is far better, and it also improves significantly upon the O(σu) worst-case time of our early result [4]. We show how the reduced-space versions of the LZ x Array (SA) [21] u log u O(u log u) SA [31] O(u log σ) (*) O(u log log σ) CSA [30] u(H0(T) + 2 + ǫ) + o(u log σ) (†) O(u log u) CSA [54] O(u log σ(logσ u) log3 2 ) (*) O(u) log σ log log u )) AF-FMI =-=[24]-=- uHk(T) + o(u log σ) (§) O(u log u(1 + LZ-index (original) [55, 56] O(u log u) O(u log σ) LZ-index (our early result) [4] (4 + ǫ)uHk(T) + o(u log σ) (‡) O(σu) LZ-index (this paper) 4uHk(T) + o(u log σ + o(ulog σ) bits. Mäkinen and Navarro [44] showed how to handle in addition insertions and deletions on bitmaps and sequences, achieving O(log ulog σ) time for all operations. This was later improved =-=[24]-=- to O(log u(1 + )), always within the same space bounds. log σ log log u Data Structures for Searchable Partial Sums Given an array A[1..n] of n integers of k ′ bits each, a data structure for searcha . The space and time bounds are valid in the standard model MB of memory allocation. Finally, the LZ-index of Lemma 6 adds the Alphabet-Friendly FM-index [19], which according to González and Navarro =-=[24]-=- can be constructed with uHk(T) + o(ulog σ) bits of space in O(ulog u(1 + )) time. Then, we have: log σ log log u Corollary 2. There exists an algorithm to construct the LZ-index of Lemma 6 for a text   </text>
<query_num> 10733 </query_num>
<text>   w of: – The space-efficient algorithm from Sirén [32] to build the Burrows-Wheeler transform of a text collection. In particular, the algorithm is used to build the Run-length Compressed Suffix Array =-=[45]-=- (RL-CSA for short). We divided the Human Genome into several equal-size files. To obtain different space/time trade-offs, we used 25 (which was the value tested by Sirén [32]), 50, 100 and 500 files.   </text>
</top>
<top>
<paper_num> 108 </paper_num>
<paper_title>   Distributed Planning in Hierarchical Factored MDPs  </paper_title>
<paper_abstract>   We present a principled and efficient planning algorithm for collaborative multiagent dynamical systems. All computation, during both the planning and the execution phases, is distributed among the agents; each agent only needs to model and plan for a small part of the system. Each of these local subsystems is small, but once they are combined they can represent an exponentially larger problem. The subsystems are connected through a subsystem hierarchy. Coordination and communication between the agents is not imposed, but derived directly from the structure of this hierarchy. A globally consistent plan is achieved by a message passing algorithm, where messages correspond to natural local reward functions and are computed by local linear programs; another message passing algorithm allows us to execute the resulting policy. When two portions of the hierarchy share the same structure, our algorithm can reuse plans and messages to speed up computation. 1  </paper_abstract>
<query_num> 10801 </query_num>
<text>   Chen were the first to apply Dantzig-Wolfe decomposition to Markov decision processes, while Dean and Lin combined decomposition with state abstraction. Hierarchical planning algorithms include MAXQ =-=[7]-=-, hierarchies of abstract machines [16], and planning with macro-operators [22, 9]. By contrast, in a parallel decomposition, multiple subproblems can be active at the same time, and the combined stat   </text>
<query_num> 10802 </query_num>
<text>   Wolfe decomposition to Markov decision processes, while Dean and Lin combined decomposition with state abstraction. Hierarchical planning algorithms include MAXQ [7], hierarchies of abstract machines =-=[16]-=-, and planning with macro-operators [22, 9]. By contrast, in a parallel decomposition, multiple subproblems can be active at the same time, and the combined state space is the cross product of the sub   </text>
<query_num> 10803 </query_num>
<text>   l of the subproblems is approximately equal to the size of the combined problem. Serial decomposition planners in the literature include Kushner and Chen’s algorithm [12] and Dean and Lin’s algorithm =-=[6]-=-, as well as a variety of hierarchical planning algorithms. Kushner and Chen were the first to apply Dantzig-Wolfe decomposition to Markov decision processes, while Dean and Lin combined decomposition   </text>
<query_num> 10804 </query_num>
<text>   les. The relationship between factored MDPs and the hierarchical decomposition described in this paper is analogous to the one between standard Bayesian networks and Object-Oriented Bayesian networks =-=[11]-=-. In terms of representational power, hierarchical multiagent factored MDPs are equivalent to factored MDPs with factored linear value functions. That is, a factored MDP associated with some choice of   </text>
<query_num> 10805 </query_num>
<text>   model. Guestrin et al. [8] used factored MDPs for multiagent planning. They presented a planning algorithm which approximates the value function of a factored MDP with factored linear value functions =-=[10]-=-. These value functions are a weighted linear combination of basis functions where each basis function is restricted to depend only on a small subset of state variables. The relationship between facto oice of basis functions can be easily transformed into a subsystem tree with a particular choice of subsystems and vice-versa. 1 This transformation involves the backprojection of the basis functions =-=[10]-=- and the triangulation of the 1 For some basis function choices, the transformation from factored MDPs to subsystem trees may also imply an approximate solution of the basic subsystem MDPs.sresulting   </text>
<query_num> 10806 </query_num>
<text>   rocesses, while Dean and Lin combined decomposition with state abstraction. Hierarchical planning algorithms include MAXQ [7], hierarchies of abstract machines [16], and planning with macro-operators =-=[22, 9]-=-. By contrast, in a parallel decomposition, multiple subproblems can be active at the same time, and the combined state space is the cross product of the subproblem state spaces. The size of the combi   </text>
<query_num> 10807 </query_num>
<text>   small number of message variables which allow us to reduce the global coordination problem to the problem of finding an appropriate reward-sharing plan. Our algorithm is also linked to reward shaping =-=[18]-=-. In reinforcement learning, it is common to add fictitious shap—Ø� ¨���� 1. Initialization: — For all subsystemsÅ�,¨���,Ì���,Ä���, and�Ë��. — For allÅ�and separating setsË�touchingÅ�, 2. For each sub   </text>
<query_num> 10808 </query_num>
<text>   stitute our approximate value function representation into the Bellman LP (1): (3) There is, in general, no guarantee on the quality of the approximationÎ�ÀÛ, but recent work of de Farias and Van Roy =-=[4]-=- provides some analysis of the error relative to that of the best possible approximation in the subspace, and some guidance as to selecting the state relevance weights« so as to improve the quality of   </text>
<query_num> 10809 </query_num>
<text>   twork of planners. One of this paper’s contributions is the method for transformation and decomposition. Our transformation is based on the factorized planning algorithm of Guestrin, Koller, and Parr =-=[8]-=-. Their algorithm uses a central planner, but allows distributed execution of plans. We extend that result to allow planning to be distributed as well, while still guaranteeing that we reach the same  ur method is the first to handle parallel decompositions of planning problems. Another contribution of our new hierarchical representation and planning algorithm over the algorithm of Guestrin et al. =-=[8]-=- is reuse. As we describe in Sec. 9, our approach can reuse plans and messages for parts of the hierarchy that share the same structure. functionÊ��¢���Ê 3 Markov Decision Processes The Markov Decisio  9 for details). 4.3 Relationship to factored MDPs Factored MDPs [2] allow the representation of large structured MDPs by using a dynamic Bayesian network [5] as the transition model. Guestrin et al. =-=[8]-=- used factored MDPs for multiagent planning. They presented a planning algorithm which approximates the value function of a factored MDP with factored linear value functions [10]. These value function ÒØ�ÖÒ�Ð�Å℄, we now have one variable for each assignment toÁÒØ�ÖÒ�Ð�Å�℄for each�. The number of constraints, however, remains the same: one for each assignment toËÓÔ��Å℄. Fortunately, Guestrin et al. =-=[8]-=- proposed an algorithm that reduces the number of constraints in a factored MDP by a method analogous to variable elimination. Their algorithm introduces an extra variableË�Þ� (called a message variab ing them. Unfortunately, our action space is exponentially large, making this É�Ü�����Ê�Ü���� approach infeasible. However, we can exploit the subsystem tree structure to select an action efficiently =-=[8]-=-. Recall that we are interested in finding the greedy action which maximizes theÉfunction. Our value function is decomposed as the sum of local value functions over subsystems. This decomposition also   </text>
<query_num> 10810 </query_num>
<text>   ubproblem of this task might be to plan a path for one robot using only a compact summary of the plans for the other robots. Parallel decomposition planners in the literature include Singh and Cohn’s =-=[21]-=- and Meuleau et al.’s [15] algorithms. Singh and Cohn’s planner builds the combined state space explicitly, using subproblem solutions to initialize the global search. So, while it may require fewer p   </text>
</top>
<top>
<paper_num> 109 </paper_num>
<paper_title>   Accurate Parsing of the Proposition Bank.  </paper_title>
<paper_abstract>   We integrate PropBank semantic role labels to an existing statistical parsing model producing richer output. We show conclusive results on joint learning and inference of syntactic and semantic representations. 1  </paper_abstract>
<query_num> 10901 </query_num>
<text>   and inference of syntactic and semantic representations. 1 Introduction Recent successes in statistical syntactic parsing based on supervised techniques trained on a large corpus of syntactic trees (=-=Collins, 1999; Charniak, 2000; Henderson, 2003-=-) have brought the hope that the same approach could be applied to the more ambitious goal of recovering the propositional content and the frame semantics of a sentenc ative comparison between our PropBank SSN parser and other PropBank parsers. However, state-of-theart semantic role labelling systems (=-=CoNLL, 2005-=-) use parse trees output by state-of-the-art parsers (=-=Collins, 1999;Charniak, 2000-=-), both for training and testing, and return partial trees annotated with semantic role labels. An indirect way of comparing our parser with semantic role labellers suggests itself. 2   </text>
<query_num> 10902 </query_num>
<text>   ight reservation system processing the sentence I want to book a flight from Geneva to New York will need to know that from Geneva indicates the origin of the flight and to New York the destination. (=-=Gildea and Jurafsky, 2002-=-) define this shallow semantic task as a classification problem where the semantic role to be assigned to each constituent is inferred on the basis of probability distributions of syntactic features e   </text>
<query_num> 10903 </query_num>
<text>   joint learning of parse tree and semantic role labels. PropBank encodes propositional information by adding a layer of argument structure annotation to the syntactic structures of the Penn Treebank (=-=Marcus et al., 1993-=-). Verbal predicates in the Penn Treebank (PTB) receive a label REL and their arguments are annotated with abstract semantic role labels A0A5 or AA for those complements of the predicative verb that a   </text>
<query_num> 10904 </query_num>
<text>   of syntactic and semantic representations. 1 Introduction Recent successes in statistical syntactic parsing based on supervised techniques trained on a large corpus of syntactic trees (=-=Collins, 1999;Charniak, 2000; Henderson, 2003-=-) have brought the hope that the same approach could be applied to the more ambitious goal of recovering the propositional content and the frame semantics of a sentence. Moving toward n between our PropBank SSN parser and other PropBank parsers. However, state-of-theart semantic role labelling systems (=-=CoNLL, 2005-=-) use parse trees output by state-of-the-art parsers (=-=Collins, 1999;Charniak, 2000-=-), both for training and testing, and return partial trees annotated with semantic role labels. An indirect way of comparing our parser with semantic role labellers suggests itself. 2 We merge the par ic role labelling systems (=-=Punyakanok et al., 2005b; Haghighi et al., 2005; Pradhan et al., 2005; Marquez et al., 2005; Surdeanu and Turmo, 2005-=-) in the CoNLL 2005 shared task. These systems all use (=-=Charniak, 2000-=-)’s parse trees both for training and testing, as well as various other information sources including sets of n-best parse trees, chunks, or named entities. Thus, the partial trees output by these sys   </text>
<query_num> 10905 </query_num>
<text>   ormance measures on the resulting parse trees. The third line, PropBank column of Table 1 reports such measures summarised for the five best semantic role labelling systems (=-=Punyakanok et al., 2005b; Haghighi et al., 2005; Pradhan et al., 2005; Marquez et al., 2005; Surdeanu and Turmo, 2005-=-) in the CoNLL 2005 shared task. These systems all use (=-=Charniak, 2000-=-)’s parse trees both for training and testing, as well as va   </text>
<query_num> 10906 </query_num>
<text>   representation of the sentence above. The indirectness of the relation is also confirmed by the difficulty in exploiting semantic information for parsing. Previous attempts have not been successful. (=-=Klein and Manning, 2003-=-) report a reduction in parsing accuracy of an unlexicalised PCFG from 77.8% to 72.9% in using Penn Treebank function labels in training. The two existing systems that use function labels sucessfully, ploit the intuition that semantic role labels are predictive of syntactic structure, we must provide semantic role information as early as possible to the parser. Extending a technique presented in (=-=Klein and Manning, 2003-=-) and adopted in (=-=Merlo and Musillo, 2005-=-) for function labels with stateof-the-art results, we split some part-of-speech tags into tags marked with AM-X semantic role labels. As a result, 240 new POS   </text>
<query_num> 10907 </query_num>
<text>   s such as phrase type, position, voice, and parse tree path. Consider, for example, a sentence such as The authority dropped at midnight Tuesday to $ 2.80 trillion (taken from section 00 of PropBank (=-=Palmer et al., 2005-=-)). The fact that to $ 2.80 trillion receives a direction semantic label Paola Merlo Department of Linguistics University of Geneva 2 Rue de Candolle 1211 Geneva 4 Switzerland merlo@lettres.unige.ch i   </text>
</top>
<top>
<paper_num> 110 </paper_num>
<paper_title>   OctoPocus: a dynamic guide for learning gesture-based command sets.  </paper_title>
<paper_abstract>   We describe OctoPocus, an example of a dynamic guide that combines on-screen feedforward and feedback to help users learn, execute and remember gesture sets. OctoPocus can be applied to a wide range of single-stroke gestures and recognition algorithms and helps users progress smoothly from novice to expert performance. We provide an analysis of the design space and describe the results of two experiments that show that OctoPocus is significantly faster and improves learning of arbitrary gestures, compared to conventional Help menus. It can also be adapted to a markbased gesture set, significantly improving input time compared to a two-level, four-item Hierarchical Marking menu. ACM Classification: D.2.2 [Software Engineering]: Design  </paper_abstract>
<query_num> 11001 </query_num>
<text>   [15] to recognize shapes. Another way to improve gesture recognition is through the gesture sets themselves. Some researchers create general tools for designing gesture-based interfaces, e.g., SATIN =-=[13]-=-. Others develop specific gesture sets that emphasize easy text input,such as, Graffiti [5] and the Unistroke [10] alphabets. Still others focus on quantitative models for predicting perceived simila   </text>
<query_num> 11002 </query_num>
<text>   accessible to users, independent of their expertise? Some research focuses on improving gesture recognition algorithms, e.g., Rubine’s example-based recognition [29] and the more recent $1 Recognizer =-=[31]-=-. Others focus on helping to design effective gesture sets, e.g., Cao et al.’s [11] pen-gesture model. We are interested in the complementary problem: helping users to learn, execute and remember new  gesture input. Technical perspective Gesture recognition algorithms include Rubine’s popular gesture classifier [29], which requires initial training by drawing sample gestures, and the $1 Recognizer =-=[31]-=-, which provides a simple and efficient algorithm intended to support rapid prototyping of gesture-based interfaces. Other algorithms use symbol fragmentation [14] or turning angle representation [15]   </text>
<query_num> 11003 </query_num>
<text>   aches provide recognition results during input. Most focus on shape beautification i.e. modifying the user’s hand-drawn input to illustrate a perfect instance of a given gesture class. Fluid Sketches =-=[2]-=- morph gesture input into simple shapes, such as circles. Users must draw a significant portion of the final gesture before obtaining useful feedback and, although Fluid Sketches have been integrated   </text>
<query_num> 11004 </query_num>
<text>   cute; the latter are easier to remember. Some researchers have explored gesture recognition in specific applications, including flick gestures in web browsers [27], interfaces for air traffic control =-=[8]-=- and drawing applications [17]. However most graphical user interfaces continue to use standard buttons and pull-down menus. Technically, gesture recognition demands robust and accurate algorithms tha   </text>
<query_num> 11005 </query_num>
<text>   d. The former are simple and quick to execute; the latter are easier to remember. Some researchers have explored gesture recognition in specific applications, including flick gestures in web browsers =-=[27]-=-, interfaces for air traffic control [8] and drawing applications [17]. However most graphical user interfaces continue to use standard buttons and pull-down menus. Technically, gesture recognition de   </text>
<query_num> 11006 </query_num>
<text>   diation strategies: repetition and choice. They also describe OOPS, a “toolkit that supports resolution of input ambiguity through mediation”. They use Igarashi&amp;apos;s interactive beautification technique =-=[16]-=- as an example of post-input mediation, since it shows users how the gesture was interpreted and lets them choose among “perfect” alternatives. Other feedback approaches provide recognition results du   </text>
<query_num> 11007 </query_num>
<text>   e Unistroke [10] alphabets. Still others focus on quantitative models for predicting perceived similarities among gestures [23,24] or models of human performance to analyze single-stroke pen gestures =-=[7]-=-. While these are all important, our focus here is on the complementary problem of improving gesture-based interfaces from the user’s perspective. Userʼs perspective Existing systems provide two basic   </text>
<query_num> 11008 </query_num>
<text>   fic gesture sets that emphasize easy text input,such as, Graffiti [5] and the Unistroke [10] alphabets. Still others focus on quantitative models for predicting perceived similarities among gestures =-=[23,24]-=- or models of human performance to analyze single-stroke pen gestures [7]. While these are all important, our focus here is on the complementary problem of improving gesture-based interfaces from the   </text>
<query_num> 11009 </query_num>
<text>   fic gesture sets that emphasize easy text input,such as, Graffiti [5] and the Unistroke [10] alphabets. Still others focus on quantitative models for predicting perceived similarities among gestures =-=[23,24]-=- or models of human performance to analyze single-stroke pen gestures [7]. While these are all important, our focus here is on the complementary problem of improving gesture-based interfaces from the  Some gesture-set design tools provide gradual information about recognition state, which helps gesture-set designers discover and fix recognition problems. For example, Long et al.‘s GDT class window =-=[23]-=- displays the state of Rubine’s algorithm after a given test input, indicating not only whether a particular gesture was recognized but also a numeric value quantifying how well it was recognized. Thi   </text>
<query_num> 11010 </query_num>
<text>   gestures next to their corresponding commands. They provide “animated, annotated demonstrations” to demonstrate each gesture and help users visualize how gestures should be performed. Avrahami et al. =-=[4]-=- Paper PDA is a paper-electronic interface with templates that guide simple, single-stroke input. The approach does not scale to on-screen interaction, because the templates include the whole gesture   </text>
<query_num> 11011 </query_num>
<text>   how best to interact with the system while simultaneously modifying their behavior to reduce errors and improve overall recognition accuracy. Applications such as Eisenstein &amp; Mackay’s Object Tracker =-=[9]-=- use the same basic principle to improve computer vision recognition of human movement gestures. However, we believe that this notion of a continuous, interactive guide, in which the user receives dyn   </text>
<query_num> 11012 </query_num>
<text>   leagues have further improved them by increasing menu breadth with zone and polygon menus [33] and have improved the efficiency of Hierarchical Marking menus by converting zigzags into single strokes =-=[32]-=-. Their key limitation relates to the gesture set itself. Kurtenbach notes that “the mark set is not particularly expressive” and that the Marking menu is not adapted to complex gestures. One of our g   </text>
<query_num> 11013 </query_num>
<text>   mple gestures, and the $1 Recognizer [31], which provides a simple and efficient algorithm intended to support rapid prototyping of gesture-based interfaces. Other algorithms use symbol fragmentation =-=[14]-=- or turning angle representation [15] to recognize shapes. Another way to improve gesture recognition is through the gesture sets themselves. Some researchers create general tools for designing gestur   </text>
<query_num> 11014 </query_num>
<text>   shapes, such as circles. Users must draw a significant portion of the final gesture before obtaining useful feedback and, although Fluid Sketches have been integrated into a graph-drawing application =-=[3]-=-, they remain limited to simple gesture sets. Li et al&amp;apos;s [21] Incremental Intention Extraction provides feedback that can be seen as a discrete version of Fluid Sketches. If a part of a user’s drawing   </text>
<query_num> 11015 </query_num>
<text>   sture and the cheat sheet. Also, displaying a complete set of gestures and associated commands takes a large amount of screen space and risks occluding large sections of the screen. Kurtenbach et al. =-=[20]-=- combined crib-sheets and contextual animation to help users learn which gestures are currently available. A pop-up cheat sheet displays the relevant subset of the gesture vocabulary available dependi   </text>
<query_num> 11016 </query_num>
<text>   the user has begun making a gesture. Feedback may consist of displaying the recognized command or provide incremental information as to the current state of the recognition algorithm. Mankoff et al. =-=[26]-=- focus on post-input mediation, in which the recognizer reveals how it interpreted the input. They survey existing error-correction techniques and identify two mediation strategies: repetition and cho   </text>
<query_num> 11017 </query_num>
<text>   ulldown menus [18] and have been adapted for text entry [28] and multiple command entry [12]. Zhao and his colleagues have further improved them by increasing menu breadth with zone and polygon menus =-=[33]-=- and have improved the efficiency of Hierarchical Marking menus by converting zigzags into single strokes [32]. Their key limitation relates to the gesture set itself. Kurtenbach notes that “the mark   </text>
<query_num> 11018 </query_num>
<text>   ving gesture recognition algorithms, e.g., Rubine’s example-based recognition [29] and the more recent $1 Recognizer [31]. Others focus on helping to design effective gesture sets, e.g., Cao et al.’s =-=[11]-=- pen-gesture model. We are interested in the complementary problem: helping users to learn, execute and remember new gesture sets. Our goal is to help users move from novice- to expert-level performan de simple, single-stroke input. The approach does not scale to on-screen interaction, because the templates include the whole gesture set and would occlude major sections of the screen. Hover Widgets =-=[11]-=- offer an alternative, providing a guide based on a dedicated algorithm that recognizes simple gestures. The user’s pen ‘hovers’ over a graphics tablet to obtain a feed-forward display of all possible   </text>
<query_num> 11019 </query_num>
<text>   y, no longer needing the Marking menu, and thus significantly increase overall performance. Marking menus are three times faster than ordinary pulldown menus [18] and have been adapted for text entry =-=[28]-=- and multiple command entry [12]. Zhao and his colleagues have further improved them by increasing menu breadth with zone and polygon menus [33] and have improved the efficiency of Hierarchical Markin   </text>
</top>
<top>
<paper_num> 111 </paper_num>
<paper_title>   Timing-Accurate Storage Emulation.  </paper_title>
<paper_abstract>   Timing-accurate storage emulation fills an important gap in the set of common performance evaluation techniques for proposed storage designs: it allows a researcher to experiment with not-yet-existing storage components in the context of real systems executing real applications. As its name suggests, a timing-accurate storage emulator appears to the system to be a real storage component with service times matching a simulation model of that component. This paper promotes timing-accurate storage emulation by describing its unique features, demonstrating its feasibility, and illustrating its value. A prototype, called the Memulator, is described and shown to produce service times within 2 % of those computed by its component simulator for over 99 % of requests. Two sets of measurements enabled by the Memulator illustrate its power: (1) application performance on a modern Linux system equipped with a MEMS-based storage device (no such device exists at this time), and (2) application performance on a modern Linux system equipped with a disk whose firmware has been modified (we have no access to firmware source code). 1  </paper_abstract>
<query_num> 11101 </query_num>
<text>   Although some details differ, this is similar to the Memulator’s design. A less direct example is the common practice of emulating nonvolatile RAM by simply pretending that normal RAM is non-volatile =-=[5, 8]-=-. Although this is unacceptable for a production system, such pretending is fine for performance experiments. A central purpose of this paper is to promote timingaccurate storage emulation as a first-   </text>
<query_num> 11102 </query_num>
<text>   and disk array controllers. Section 6 explores concrete examples of both types of experiments. We are aware of only one other technique offering a similar mix of features: complete machine simulation =-=[3, 17, 19]-=-. Under this technique, the hardware of a computer system is simulated in enough detail to boot a real OS and run applications. If the simulation progresses according to timing-accurate models of the   </text>
<query_num> 11103 </query_num>
<text>   be supported, but often the implementations of both sides must change to truly exploit a new interface’s potential. Two examples of this arise from recently-proposed mechanisms: freeblock scheduling =-=[16]-=- and eager writing [25]. Freeblock scheduling consists of replacing the rotational latency delays of high-priority disk requests with background media transfers. Since the high-priority data will rota   </text>
<query_num> 11104 </query_num>
<text>   he same types of experiments as storage emulation. Further, by manipulating simulator parameters, the effects of new storage devices on hypothetical machines (e.g., with 10 GHz CPUs) can be evaluated =-=[20, 22]-=-. Unfortunately, substantial effort is required to build and maintain a complete machine simulator, both in terms of correctly executing programs and correctly accounting for time. For example, the Si   </text>
<query_num> 11105 </query_num>
<text>   he same types of experiments as storage emulation. Further, by manipulating simulator parameters, the effects of new storage devices on hypothetical machines (e.g., with 10 GHz CPUs) can be evaluated =-=[20, 22]-=-. Unfortunately, substantial effort is required to build and maintain a complete machine simulator, both in terms of correctly executing programs and correctly accounting for time. For example, the Si  run time 74�523s 51�420s 31.0% std. dev. 0.618 1.678 — though the devices themselves are several years from production. We configured the Memulator to use the G2 device described by Schlosser et al. =-=[22]-=-. Table 7 compares the performance of PostMark running on the Cheetah X15 and on MEMS-based storage. Although the average response time of the disk was five times greater than the MEMS-based storage d   </text>
<query_num> 11106 </query_num>
<text>   product designs. However, these examples represent only the “storage emulation” half of timing-accurate storage emulation. The “timing-accurate” half has been much utilized by networking researchers =-=[1, 6, 18]-=-. Timingaccurate network emulation parallels our description of timing-accurate storage emulation: real hosts interconnected by the emulated network observe normal packet send/receive semantics and pe   </text>
<query_num> 11107 </query_num>
<text>   sible. We are aware of only a few previous cases of timingaccurate storage emulation being used for performance evaluation. The most relevant example is the evaluation of eager writing by Wang et al. =-=[25]-=-. Under eager writing, data is written to a disk location that is close to the disk head’s current location. To evaluate the benefits of having disk firmware support for eager writing, Wang et al. emb it returns the computed service time. After the appropriate real-time delay, the timing loop tells the storage interface component to report completion. The emulator-based evaulation of eager writing =-=[25]-=- used a disk simulator by Kotz et al. [15] in this way. Although it is straightforward, this first approach often does not properly handle concurrent requests. For example, a new request arrival may a n the implementations of both sides must change to truly exploit a new interface’s potential. Two examples of this arise from recently-proposed mechanisms: freeblock scheduling [16] and eager writing =-=[25]-=-. Freeblock scheduling consists of replacing the rotational latency delays of high-priority disk requests with background media transfers. Since the high-priority data will rotate around to the disk h   </text>
<query_num> 11108 </query_num>
<text>   unavailable storage components in the context of real systems. Such experimentation is important because complex system characteristics can hide or reduce predicted benefits of new storage components =-=[9]-=-. Further, some new storage architectures and interfaces require both OS modifications and new (or modified) storage components—until the new components are available, only emulation allows such colla mponents, overall system, or workload. In particular, representative workloads are rarely used, since synthetic workload generation is still an open problem, I/O traces ignore system feedback effects =-=[9]-=-, and available traces are often out-of-date—in fact, many storage researchers still rely on the decadeold “HP traces” from 1992 [21]. As a different example, experimenting with prototypes allows one   </text>
</top>
<top>
<paper_num> 112 </paper_num>
<paper_title>   Non-uniform ACC Circuit Lower Bounds.  </paper_title>
<paper_abstract>   The class ACC consists of circuit families with constant depth over unbounded fan-in AND, OR, NOT, and MODm gates, where m&amp;gt; 1 is an arbitrary constant. We prove: • NTIME[2 n] does not have non-uniform ACC circuits of polynomial size. The size lower bound can be slightly strengthened to quasi-polynomials and other less natural functions. • ENP, the class of languages recognized in 2O(n) time with an NP oracle, doesn’t have non-uniform ACC circuits of 2no(1) size. The lower bound gives an exponential size-depth tradeoff: for every d there is a δ&amp;gt; 0 such that ENP doesn’t have depth-d ACC circuits of size 2nδ. Previously, it was not known whether EXP NP had depth-3 polynomial size circuits made out of only MOD6 gates. The high-level strategy is to design faster algorithms for the circuit satisfiability problem over ACC circuits, then prove that such algorithms entail the above lower bounds. The algorithm combines known properties of ACC with fast rectangular matrix multiplication and dynamic programming, while the second step requires a subtle strengthening of the author’s prior work [STOC’10]. Supported by the Josef Raviv Memorial Fellowship.  </paper_abstract>
<query_num> 11201 </query_num>
<text>   C circuits of depth d and size 2nδ. Recall that the lowest complexity class for which we know exponential (unrestricted) circuit lower bounds is ∆ EXP 3 , the third level of the exponential hierarchy =-=[MVW99]-=-. Extending the approach of this paper to settle the second frontier question may be difficult, but this prospect does not look as implausible as it did before. If polynomial unrestricted circuits cou known to not have unrestricted polynomial size circuits is MAEXP [BFT98]. Later it was shown that the MAEXP lower bound can be improved to half-exponential size functions f which satisfy f(f(n)) ≥ 2n =-=[MVW99]-=-. Kabanets and Impagliazzo [KI04] proved that NEXP RP either doesn’t have polynomial size Boolean circuits (over AND, OR, NOT), or it doesn’t have polynomial size arithmetic circuits (over the integer   </text>
<query_num> 11202 </query_num>
<text>   f 2 n · poly(n) size is satisfiable. Fact 3.1 follows from several prior works concerned with the complexity of the Cook-Levin theorem [Tou01, FLvMV05]: Theorem 3.3 (Tourlakis [Tou01], Fortnow et al. =-=[FLvMV05]-=-) There is a c &amp;gt; 0 such that for all L ∈ NTIME[n], L reduces to 3SAT in O(n(log n) c ) time. Moreover there is an algorithm (with random access to its input) that, given an instance of L with length n   </text>
<query_num> 11203 </query_num>
<text>   nomial constructible function S : N → N, the class NTIME[S(n)] NP does not have polynomial size circuits. Another somewhat small class known to not have unrestricted polynomial size circuits is MAEXP =-=[BFT98]-=-. Later it was shown that the MAEXP lower bound can be improved to half-exponential size functions f which satisfy f(f(n)) ≥ 2n [MVW99]. Kabanets and Impagliazzo [KI04] proved that NEXP RP either does   </text>
<query_num> 11204 </query_num>
<text>   of its inputs.) Then Smolensky [Smo87] proved exponential lower bounds for computing MODq with constant-depth circuits made up of AND, OR, NOT, and MODp gates, for distinct primes p and q. Barrington =-=[Bar89]-=- suggested the next step would be to prove lower bounds for the class ACC, which consists of constant-depth circuit families over the basis AND, OR, NOT, and MODm for arbitrary constant m &amp;gt; 1. 1 It is cuits is key to the ACC SAT algorithm, and it would be very interesting to find similar lemmas for TC 0 or NC 1 . It is plausible that the characterization of NC 1 as bounded-width branching programs =-=[Bar89]-=- could be applied to prove an analogous evaluation lemma for Boolean formulas, which would lead to nontrivial depth lower bounds for NEXP. It should be possible to extend the superpolynomial lower bou   </text>
<query_num> 11205 </query_num>
<text>   rm ACC circuit family can be simulated by subexponential uniform SYM + circuits. This was applied to show that the Permanent does not have uniform ACC circuits of subexponential size. Later, Allender =-=[All99]-=- improved the Permanent lower bound to polynomial size uniform TC 0 circuits. However, these proofs require uniformity, and the difference between uniformity and non-uniformity may well be vast (e.g.,   </text>
<query_num> 11206 </query_num>
<text>   s. However, these proofs require uniformity, and the difference between uniformity and non-uniformity may well be vast (e.g., it is clear that P ̸= NEXP, but open whether NEXP ⊆ P/poly). Green et al. =-=[GKRST95]-=- showed that the symmetric function can be assumed to be the specific function which returns the middle bit of the sum of its inputs. This representation may also be used in the lower bounds of this p   </text>
<query_num> 11207 </query_num>
<text>   stic SYM + circuit of s O(logc s) size, where c depends on the depth, and the ANDs have poly(log s) fan-in. Beigel and Tarui [BT94] showed how to remove the probabilistic condition. Allender and Gore =-=[AG94]-=- showed that every subexponential uniform ACC circuit family can be simulated by subexponential uniform SYM + circuits. This was applied to show that the Permanent does not have uniform ACC circuits o thod can be implemented using either fast rectangular matrix multiplication, or a dynamic programming approach. It follows from the work of Yao [Yao90], Beigel and Tarui [BT94], and Allender and Gore =-=[AG94]-=- that, given any ACC circuit of size s, one can produce a s O(logc s) size SYM + circuit in poly(s O(log c s) ) time that has equivalent functionality, and very special properties. (For more backgroun   </text>
</top>
<top>
<paper_num> 113 </paper_num>
<paper_title>   Privacy and Integrity are Possible in the Untrusted Cloud.  </paper_title>
<paper_abstract>   From word processing to online social networking, user-facing applications are increasingly being deployed in the cloud. These cloud services are attractive because they offer high scalability, availability, and reliability. But adopting them has so far forced users to cede control of their data to cloud providers, leaving the data vulnerable to misuse by the providers or theft by attackers. Thus, users have had to choose between trusting providers or forgoing cloud deployment’s benefits entirely. In this article, we show that it is possible to overcome this trade-off for many applications. We describe two of our recent systems, SPORC [13] and Frientegrity [12], that enable users to benefit from cloud deployment without having to trust providers for confidentiality or integrity. In both systems, the provider only observes encrypted data and cannot deviate from correct execution without detection. Moreover, for cases when the provider does misbehave, SPORC introduces a mechanism, also applicable to Frientegrity, that enables users to recover. SPORC is a framework that enables a wide variety of collaborative applications such as collaborative text editors and shared calendars with an untrusted provider. It allows concurrent, low-latency editing of shared state, permits disconnected operation, and supports dynamic access control even in the presence of concurrency. Frientegrity extends SPORC’s model to online social networking. It introduces novel mechanisms for verifying the provider’s correctness and access control that scale to hundreds of friends and tens of thousands of posts while still providing the same security guarantees as SPORC. By effectively returning control of users ’ data to the users themselves, these systems do much to mitigate the risks of cloud deployment. 1  </paper_abstract>
<query_num> 11301 </query_num>
<text>   Frientegrity users are only known to the provider by pseudonym. Nevertheless, a provider may be able to glean some information via traffic analysis and social network deanonymization techniques [1], =-=[27]-=-. A full mitigation of these attacks is beyond the scope of this work. Users and clients Both systems assume that users may also be malicious and colluding with a malicious provider. As a result, user   </text>
<query_num> 11302 </query_num>
<text>   and rekeying must be efficient. To meet these requirements, we represent ACLs with a tree-like data structure that is a novel combination of a persistent authenticated dictionary [7] and a key graph =-=[40]-=- in which each node is a “friend.” A given user’s membership proof is simply a path from the root to that user’s node and requires space and verification time that is logarithmic in the number of user   </text>
<query_num> 11303 </query_num>
<text>   at least partially support these capabilities would go a long way in spurring the adoption of systems like ours. These solution may well involve algorithms that operate on encrypted data (e.g., [33], =-=[4]-=-, [3]) even if fully homomorphic encryption [15] remains impractical. Acknowledgements We thank Andrew Appel, Matvey Arye, Christian Cachin, Jinyuan Li, Wyatt Lloyd, Siddhartha Sen, Alexander Shraer,   </text>
<query_num> 11304 </query_num>
<text>   crypted operations, one might wonder why they use a centralized provider at all. Indeed, many peer-to-peer group collaboration and social networking systems have been proposed (e.g., [36], [18], [2], =-=[37]-=-, [8]). But decentralized schemes have at least two major disadvantages. First, they leave an end user with an unenviable dilemma: either sacrifice availability, reliability, and convenience by storin   </text>
<query_num> 11305 </query_num>
<text>   d users typically are interested only in the most recent updates, not in the thousands that may have come before. In addition, many previous proposals for secure social networking systems (e.g., [2], =-=[24]-=-, [5]) required work that is linear in the number of friends, if not FoFs, to revoke a friend’s access (i.e., to “un-friend”). But in real social networks, users may have hundreds of friends and tens   </text>
<query_num> 11306 </query_num>
<text>   e extensive, and often unnecessary, dependencies between objects, thereby making it difficult to spread objects across multiple servers without resorting to expensive agreement protocols (e.g., Paxos =-=[21]-=-). Thus, for scalability, Frientegrity orders operations and enforces fork* consistency on each object independently. Weakening consistency across objects leads to new attacks, however. For example, e   </text>
<query_num> 11307 </query_num>
<text>   e large, ACL changes and rekeying must be efficient. To meet these requirements, we represent ACLs with a tree-like data structure that is a novel combination of a persistent authenticated dictionary =-=[7]-=- and a key graph [40] in which each node is a “friend.” A given user’s membership proof is simply a path from the root to that user’s node and requires space and verification time that is logarithmic   </text>
<query_num> 11308 </query_num>
<text>   ed a compact representation of its view of the history 5 in every operation it creates, and clients which subsequently read the operation can compare their views to the embedded one. 4 A history tree =-=[6]-=- is a growable Merkle hash tree that has been used previously for tamper-evident logging. 5 i.e., the history tree’s current root hash signed by the provider. 7Thus, when Alice reads the tail of Bob’   </text>
<query_num> 11309 </query_num>
<text>   g site, tried to disguise its censorship of a user’s posts by hiding them from the user’s followers but still showing them to the user [34]. This behavior is an example of provider equivocation [25], =-=[22]-=-, in which a malicious service presents different clients with divergent views of the system state. In sum, the emerging class of user-facing cloud services currently requires users to cede control of edian server throughput reached 1600 ops/sec. Notably, latency was dominated by the cost of clients’ 2048-bit RSA signatures, and thus it could be improved greatly by a faster algorithm such as ESIGN =-=[22]-=-. 3 3 Complete results can be found in the full paper [13]. 64 Frientegrity Our second system, Frientegrity, extends SPORC’s confidentiality and integrity guarantees to online social networking. It s   </text>
<query_num> 11310 </query_num>
<text>   logging site, tried to disguise its censorship of a user’s posts by hiding them from the user’s followers but still showing them to the user [34]. This behavior is an example of provider equivocation =-=[25]-=-, [22], in which a malicious service presents different clients with divergent views of the system state. In sum, the emerging class of user-facing cloud services currently requires users to cede cont change views of the history out-ofband, even a provider which forks the clients will not be able to cheat for long. 2 Fork* consistency is a weaker variant of an earlier model called fork consistency =-=[25]-=-. They differ in that under fork consistency, a pair of clients only needs to exchange one message to detect server equivocation, whereas under fork* consistency, they may need to exchange two. Our sy   </text>
<query_num> 11311 </query_num>
<text>   ng and storing encrypted operations, one might wonder why they use a centralized provider at all. Indeed, many peer-to-peer group collaboration and social networking systems have been proposed (e.g., =-=[36]-=-, [18], [2], [37], [8]). But decentralized schemes have at least two major disadvantages. First, they leave an end user with an unenviable dilemma: either sacrifice availability, reliability, and conv   </text>
<query_num> 11312 </query_num>
<text>   ng encrypted operations, one might wonder why they use a centralized provider at all. Indeed, many peer-to-peer group collaboration and social networking systems have been proposed (e.g., [36], [18], =-=[2]-=-, [37], [8]). But decentralized schemes have at least two major disadvantages. First, they leave an end user with an unenviable dilemma: either sacrifice availability, reliability, and convenience by  y, and users typically are interested only in the most recent updates, not in the thousands that may have come before. In addition, many previous proposals for secure social networking systems (e.g., =-=[2]-=-, [24], [5]) required work that is linear in the number of friends, if not FoFs, to revoke a friend’s access (i.e., to “un-friend”). But in real social networks, users may have hundreds of friends and   </text>
<query_num> 11313 </query_num>
<text>   ys to at least partially support these capabilities would go a long way in spurring the adoption of systems like ours. These solution may well involve algorithms that operate on encrypted data (e.g., =-=[33]-=-, [4], [3]) even if fully homomorphic encryption [15] remains impractical. Acknowledgements We thank Andrew Appel, Matvey Arye, Christian Cachin, Jinyuan Li, Wyatt Lloyd, Siddhartha Sen, Alexander Shr   </text>
</top>
<top>
<paper_num> 114 </paper_num>
<paper_title>   A quantitative analysis of aspects in the eCos kernel.  </paper_title>
<paper_abstract>   Nearly ten years after its first presentation and five years after its first application to operating systems, the suitability of Aspect-Oriented Programming (AOP) for the development of operating system kernels is still highly in dispute. While the AOP advocacy emphasizes the benefits of AOP towards better configurability and maintainability of system software, most kernel developers express a sound skepticism regarding the thereby induced runtime and memory costs: Operating system kernels have to be lean and efficient. We have analyzed the runtime and memory costs of aspects in general, on the level of µ-benchmarks, and by refactoring and extending the eCos operating system kernel using AspectC++, an AOP extension to the C++ language. Our results show that most AOP features do not induce a intrinsic overhead and that the actual overhead induced by AspectC++ is very low. We have also analyzed a test case with significant aspect-related costs. This example shows how the structure of the underlying kernel can have a negative impact on aspect implementations and how these costs can be avoided by an aspect-aware design. Based on this analysis, our conclusion is that AOP is suitable for the development of operating system kernels and other kinds of highly efficient infrastructure software.  </paper_abstract>
<query_num> 11401 </query_num>
<text>   6, Leuven, Belgium. Copyright 2006 ACM 1-59593-322-0/06/0004 ...$5.00. 1. INTRODUCTION Nearly ten years after its first presentation[22] and five years after its first application to operating systems=-=[4, 7]-=-, it is still a controversial question whether Aspect-Oriented Programming (AOP) provides a real benefit for the development of operating system kernels. Many OS developers have a sound skepticism reg   </text>
<query_num> 11402 </query_num>
<text>   6, Leuven, Belgium. Copyright 2006 ACM 1-59593-322-0/06/0004 ...$5.00. 1. INTRODUCTION Nearly ten years after its first presentation[22] and five years after its first application to operating systems=-=[4, 7]-=-, it is still a controversial question whether Aspect-Oriented Programming (AOP) provides a real benefit for the development of operating system kernels. Many OS developers have a sound skepticism reg l. retroactively evaluated the evolution of four scattered OS concerns (prefetching, disk quotas, blocking, and page daemon activation) in the FreeBSD kernel using the general-purpose AspectC language=-=[7, 6]-=-. It was shown that an aspect-oriented implementation would have led to significantly better evolvability. Due to missing tool support (namely an “aspect weaver”), her study did cover only a relativel ivided into work based on static weaving, and work based on dynamic weaving. An overview of related work regarding AOP in system software with static weaving was already presented in the introduction =-=[7, 6, 25, 27, 2, 16, 8, 31, 29]-=-. The main contribution of this paper over existing work is the in-depth cost-analysis, performed with system-specific and cost-critical concerns. Some related work regarding costs of AOP with static   </text>
<query_num> 11403 </query_num>
<text>   This is basically a question of tools, namely of AOP-aware editors and debuggers that visualize in the source how aspects affect the currently viewed or editeds192 EuroSys 2006 component. For AspectJ=-=[21]-=- such tool support already exists and has reached an industry-strength level of maturity[20]. For other languages, such as AspectC++[28], it is actively being developed (http://acdt.aspectc.org). Over modules, which also could be reused in other contexts without or with different synchronization schemes. Today, most AOP languages use the concepts and terminology that was first introduced by AspectJ=-=[21]-=-. In the remaining parts of this section, we will give a brief overview of the most common AOP language elements in general and the AspectC++ notion in particular, as required for understanding the re   </text>
<query_num> 11404 </query_num>
<text>   appropriate. The number of potential join-points could also be increased by a join-point model that supports join-points on the statement and expression level, such as loops and local variable access=-=[17]-=-. This clearly remains a topic for further research, even though we doubt that most of these join-points would offer enough semantics for robust aspect implementations.sEuroSys 2006 203 The Role of De   </text>
<query_num> 11405 </query_num>
<text>   as some of its fundamental concepts, such as late binding, induce a significant (and in many cases inevitable) overhead[11]. So far, only few studies analyze the costs of aspects. The existing studies=-=[12, 3]-=- were mainly conducted in the Java domain using AspectJ and are somewhat disappointing, as they show that AOP with AspectJ indeed induces some overhead[12]. It is, however, questionable if these resul , and even the number of aspects giving advice to the join-point (Table 1-b). This is noteworthy, as in AspectJ, for instance, around advice induces significantly higher costs than before/after advice=-=[12]-=-. The size of the text segment (code) remains also stable, the increase by 16 bytes in one case was mainly caused by linker alignment of the affected section. While advice for parameterless functions   concerns. Some related work regarding costs of AOP with static weaving has been conducted in the AspectJ domain. Dufour presented a benchmark suite to measure the dynamic behavior of AspectJ programs=-=[12]-=-. His work focuses on a novel measuring approach, however, it also shows that several AspectJ features induce significant overhead. Based on this work, Avgustinov suggested some improvements for the A context information through a generic interface. Furthermore, some aspect weavers require a runtime system and, in fact, earlier quantitative studies indicate that there is an overhead related to AOP =-=[12, 19]-=-. Our work shows that these earlier results do not mean that AOP imposes an overhead in general. Especially, aspect weaving in C and C++, which are the dominant languages in the domain of system softw   </text>
<query_num> 11406 </query_num>
<text>   as some of its fundamental concepts, such as late binding, induce a significant (and in many cases inevitable) overhead[11]. So far, only few studies analyze the costs of aspects. The existing studies=-=[12, 3]-=- were mainly conducted in the Java domain using AspectJ and are somewhat disappointing, as they show that AOP with AspectJ indeed induces some overhead[12]. It is, however, questionable if these resul a novel measuring approach, however, it also shows that several AspectJ features induce significant overhead. Based on this work, Avgustinov suggested some improvements for the AspectJ code generation=-=[3]-=- that would specifically reduce the overhead caused by cflow and around in AspectJ programs. Several papers also suggest to use dynamic weaving in system software. Engel presented an approach to dynam   </text>
<query_num> 11407 </query_num>
<text>   cost-critical concerns such as interrupt synchronization. In the C/C++ domain, the platform-dependence of most dynamic weaving approaches is another issue. A combination of static and dynamic weaving=-=[26]-=-, would be promising here, as it combines the best of both worlds. The costs of dynamically woven aspects was analyzed by Haupt for several Java-based dynamic weaving approaches[18]. According to the   </text>
<query_num> 11408 </query_num>
<text>   el from the AOP perspective would look like and how far we could go with separation of concerns. To answer this question is the goal of our ongoing and future work on the CiAO operating system family =-=[23, 24]-=-. The results of this paper are an important first step towards a truly aspect-oriented operating system. We are now convinced that the resource consump204 EuroSys 2006 tion of the CiAO system will e   </text>
<query_num> 11409 </query_num>
<text>   for the application of kernel patches[16]. Several other publication show that AOP provides benefits for the development of configurable infrastructure software in the broad sense, namely middleware =-=[8, 31, 19]-=- and databases[29] product lines. All these studies demonstrate that there are real cases for aspects in system software. From the separation of concerns viewpoint, which is the focus of the existing  context information through a generic interface. Furthermore, some aspect weavers require a runtime system and, in fact, earlier quantitative studies indicate that there is an overhead related to AOP =-=[12, 19]-=-. Our work shows that these earlier results do not mean that AOP imposes an overhead in general. Especially, aspect weaving in C and C++, which are the dominant languages in the domain of system softw   </text>
<query_num> 11410 </query_num>
<text>   for the application of kernel patches[16]. Several other publication show that AOP provides benefits for the development of configurable infrastructure software in the broad sense, namely middleware =-=[8, 31, 19]-=- and databases[29] product lines. All these studies demonstrate that there are real cases for aspects in system software. From the separation of concerns viewpoint, which is the focus of the existing  ivided into work based on static weaving, and work based on dynamic weaving. An overview of related work regarding AOP in system software with static weaving was already presented in the introduction =-=[7, 6, 25, 27, 2, 16, 8, 31, 29]-=-. The main contribution of this paper over existing work is the in-depth cost-analysis, performed with system-specific and cost-critical concerns. Some related work regarding costs of AOP with static   </text>
<query_num> 11411 </query_num>
<text>   grained design and implementation. This results in a high number of potential join-points with strong semantics. In monolithic systems with a high coupling among concern implementations, such as Linux=-=[30]-=-, the number of potential join-points is probably much lower and join-points are semantically ambiguous. In this case it might be more difficult or even impossible to implement concerns such as interr   </text>
<query_num> 11412 </query_num>
<text>   ion model[27] with aspects. Not a general-purpose AOP language, but an AOP-inspired language of temporal logic was used by Åberg et al. to integrate the Bossa scheduler framework into the Linux kernel=-=[2]-=-. C4 uses AOP concepts to implement a “semantic patch system” for the application of kernel patches[16]. Several other publication show that AOP provides benefits for the development of configurable i ntation by aspects without performing additional refactoring in the primary structure of the target systems. For such scenarios, special-purpose AOP approaches as suggested by Fiuczynski[16] and Åberg=-=[2]-=- might be more appropriate. The number of potential join-points could also be increased by a join-point model that supports join-points on the statement and expression level, such as loops and local v ivided into work based on static weaving, and work based on dynamic weaving. An overview of related work regarding AOP in system software with static weaving was already presented in the introduction =-=[7, 6, 25, 27, 2, 16, 8, 31, 29]-=-. The main contribution of this paper over existing work is the in-depth cost-analysis, performed with system-specific and cost-critical concerns. Some related work regarding costs of AOP with static   </text>
<query_num> 11413 </query_num>
<text>   lize in the source how aspects affect the currently viewed or editeds192 EuroSys 2006 component. For AspectJ[21] such tool support already exists and has reached an industry-strength level of maturity=-=[20]-=-. For other languages, such as AspectC++[28], it is actively being developed (http://acdt.aspectc.org). Overall, the comprehensibility issue is solvable and not a fundamental problem of AOP. Resource   </text>
<query_num> 11414 </query_num>
<text>   prior specific permission and/or a fee. EuroSys’06, April 18–21, 2006, Leuven, Belgium. Copyright 2006 ACM 1-59593-322-0/06/0004 ...$5.00. 1. INTRODUCTION Nearly ten years after its first presentation=-=[22]-=- and five years after its first application to operating systems[4, 7], it is still a controversial question whether Aspect-Oriented Programming (AOP) provides a real benefit for the development of op   </text>
<query_num> 11415 </query_num>
<text>   rrently viewed or editeds192 EuroSys 2006 component. For AspectJ[21] such tool support already exists and has reached an industry-strength level of maturity[20]. For other languages, such as AspectC++=-=[28]-=-, it is actively being developed (http://acdt.aspectc.org). Overall, the comprehensibility issue is solvable and not a fundamental problem of AOP. Resource Efficiency Efficiency in terms of runtime an nerated code without compromising on the expressiveness of AOP concepts. On the language level, this is achieved by integrating AOP with the C++ philosophy of static typing and compile-time genericity=-=[28]-=-. On the tool level, AspectC++ follows a source-to-source weaving approach with generation of code patterns that (1) do not use “expensive” C++ language elements (such as RTTI or exceptions), and (2)   </text>
<query_num> 11416 </query_num>
<text>   rt of the kernel code base and no heavily cross-cutting concerns such as tracing or kernel diagnostics. Our group conducted some experiments with AspectC++ in the PURE embedded operating system family=-=[5]-=- by implementing OS concerns like interrupt synchronization[25] and the driver execution model[27] with aspects. Not a general-purpose AOP language, but an AOP-inspired language of temporal logic was   </text>
<query_num> 11417 </query_num>
<text>   stics. Our group conducted some experiments with AspectC++ in the PURE embedded operating system family[5] by implementing OS concerns like interrupt synchronization[25] and the driver execution model=-=[27]-=- with aspects. Not a general-purpose AOP language, but an AOP-inspired language of temporal logic was used by Åberg et al. to integrate the Bossa scheduler framework into the Linux kernel[2]. C4 uses  ivided into work based on static weaving, and work based on dynamic weaving. An overview of related work regarding AOP in system software with static weaving was already presented in the introduction =-=[7, 6, 25, 27, 2, 16, 8, 31, 29]-=-. The main contribution of this paper over existing work is the in-depth cost-analysis, performed with system-specific and cost-critical concerns. Some related work regarding costs of AOP with static   </text>
<query_num> 11418 </query_num>
<text>   support or proper synchronization. Separation of concerns is clearly an important issue in the domain of operating systems. This was shown by retroactive studies in the context of the FreeBSD kernel =-=[6]-=-. It is also evident from the “#ifdef-hell” that can be found in the code of many current operating systems. Especially systems that strive for a high level of tailorability and configurability suffer l. retroactively evaluated the evolution of four scattered OS concerns (prefetching, disk quotas, blocking, and page daemon activation) in the FreeBSD kernel using the general-purpose AspectC language=-=[7, 6]-=-. It was shown that an aspect-oriented implementation would have led to significantly better evolvability. Due to missing tool support (namely an “aspect weaver”), her study did cover only a relativel ivided into work based on static weaving, and work based on dynamic weaving. An overview of related work regarding AOP in system software with static weaving was already presented in the introduction =-=[7, 6, 25, 27, 2, 16, 8, 31, 29]-=-. The main contribution of this paper over existing work is the in-depth cost-analysis, performed with system-specific and cost-critical concerns. Some related work regarding costs of AOP with static   </text>
</top>
<top>
<paper_num> 115 </paper_num>
<paper_title>   Heap data allocation to scratch-pad memory in embedded systems.  </paper_title>
<paper_abstract>   Abstract — This paper presents the first-ever compiletime method for allocating a portion of the heap data to scratch-pad memory. A scratch-pad is a fast directly addressed compiler-managed SRAM memory that replaces the hardware-managed cache. It is motivated by its better real-time guarantees vs cache and by its significantly lower overheads in access time, energy consumption, area and overall runtime. Existing compiler methods for allocating data to scratch-pad are able to place only global and stack data in scratch-pad memory; heap data is allocated entirely in DRAM, resulting in poor performance. Runtime methods based on software caching can place heap data in scratchpad, but because of their high overheads from software address translation, they have not been successful, especially for heap data. In this paper we present a dynamic yet compiler-directed allocation method for heap data that for the first time, (i) is able to place a portion of the heap data in scratch-pad; (ii) has no software-caching tags; (iii) requires no run-time per-access extra address translation; and (iv) is able to move heap data back and forth between scratch-pad and DRAM to better track the program’s locality characteristics. With our method, global, stack and heap variables can share the same scratch-pad. When compared to placing all heap variables in DRAM and only global and stack data in scratch-pad, our results show that our method reduces the average runtime of our benchmarks by 34.6%, and the average power consumption by 39.9%, for the same size of scratch-pad fixed at 5 % of total data size. Index Terms — heap allocation, scratch pad, SRAM, tightly coupled memory, TCM, dynamic allocation. I.  </paper_abstract>
<query_num> 11501 </query_num>
<text>   8], [19], [22] or to stack-like constructs called regions [17], [26] may help in allocating heap data to scratch-pad. Here is some background on these methods. These methods use escape analysis [45], =-=[60]-=- to try to prove that a heap data structure is never accessed outside a certain procedure. If so, the heap variable can be placed on the procedure’s stack frame, instead of the heap. The advantage of   </text>
<query_num> 11502 </query_num>
<text>   E USE HEAP TO STACK CONVERSION TO ALLOCATE TO SCRATCH PAD? At first glance, it seems that recent work on converting heap data to stack data [18], [19], [22] or to stack-like constructs called regions =-=[17]-=-, [26] may help in allocating heap data to scratch-pad. Here is some background on these methods. These methods use escape analysis [45], [60] to try to prove that a heap data structure is never acces iction with stack allocation is that it requires fixed-size heap variables, except in some cases when the data is on the frame on the top of the stack. Since this is restrictive, region-based schemes =-=[17]-=-, [26] have been proposed for when heap data is of unknown size. Regions, like stack frames, are associated with procedures but are physically allocated on the heap so that they can grow and shrink at   </text>
<query_num> 11503 </query_num>
<text>   HEAP TO STACK CONVERSION TO ALLOCATE TO SCRATCH PAD? At first glance, it seems that recent work on converting heap data to stack data [18], [19], [22] or to stack-like constructs called regions [17], =-=[26]-=- may help in allocating heap data to scratch-pad. Here is some background on these methods. These methods use escape analysis [45], [60] to try to prove that a heap data structure is never accessed ou  with stack allocation is that it requires fixed-size heap variables, except in some cases when the data is on the frame on the top of the stack. Since this is restrictive, region-based schemes [17], =-=[26]-=- have been proposed for when heap data is of unknown size. Regions, like stack frames, are associated with procedures but are physically allocated on the heap so that they can grow and shrink at runti   </text>
<query_num> 11504 </query_num>
<text>   ch-pad. Our proposed dynamic method promises to be the first to successfully place heap data in scratch-pad. Some software caching schemes have been proposed for desktops that use dynamic compilation =-=[32]-=- which changes the program at runtime in RAM. Most embedded systems, however, store the program in unchangeable ROM, and dynamic compilation cannot be used. Other software caching schemes have been pr   </text>
<query_num> 11505 </query_num>
<text>   e of these methods use greedy strategies to find a solution; others model the problem as a knapsack problem or an integer-linear programming problem to find a solution. Some static allocation methods =-=[6]-=-, [59] aim to allocate code to SPM rather than data. Other static methods [61], [53] can allocate both code and data to SPM. Their data allocation is restricted to global and stack data. The the A C B en for lack of space. These architectures are popular because SPMs are simple to design and verify, and provide better real-time guarantees for global and stack data [62], power consumption, and cost =-=[6]-=-, [53], [59], [11] compared to caches. For these architectures our method delivers runtime and energy reductions averaging 34.6% and 39.9%, respectively, compared to the best previous method. Neverthe 3], [10]. Second, other researchers have repeatedly demonstrated a significant energy and run-time savings for benchmarks containing only global and stack data for an SPM vs. a cache of the same area =-=[6]-=-, [53], [59], [11]. For these reasons, if an embedded task set contains some tasks of only global and stack data and other tasks having heap data as well, our method will enable the designer to use a   </text>
<query_num> 11506 </query_num>
<text>   e site into a single heap ”variable”. Additional techniques such as shape analysis have aimed to identify logical heap structures, such as trees. Finally, in languages with pointers, pointer analysis =-=[24]-=-, [52] is able to find all possible heap variables that a particular memory reference can access. Having understood heap variables, let us consider why heap data is difficult to allocate to scratch-pa   </text>
<query_num> 11507 </query_num>
<text>   em as a knapsack problem or an integer-linear programming problem to find a solution. Some static allocation methods [6], [59] aim to allocate code to SPM rather than data. Other static methods [61], =-=[53]-=- can allocate both code and data to SPM. Their data allocation is restricted to global and stack data. The the A C B D C E A 4sStep 1. Partition program into regions. /* Section V */ Step 2. Compute i r lack of space. These architectures are popular because SPMs are simple to design and verify, and provide better real-time guarantees for global and stack data [62], power consumption, and cost [6], =-=[53]-=-, [59], [11] compared to caches. For these architectures our method delivers runtime and energy reductions averaging 34.6% and 39.9%, respectively, compared to the best previous method. Nevertheless,  her advantages of SPMs over caches not apparent from the results above. First, it is widely known that for global and stack data, SPMs have significantly better real-time guarantees than caches [62], =-=[53]-=-, [10]. Second, other researchers have repeatedly demonstrated a significant energy and run-time savings for benchmarks containing only global and stack data for an SPM vs. a cache of the same area [6   </text>
<query_num> 11508 </query_num>
<text>   ever, store the program in unchangeable ROM, and dynamic compilation cannot be used. Other software caching schemes have been proposed with different goals and/or non-applicable platforms [58], [15], =-=[21]-=-, [47], [16], [34]. For lack of space, we do not discuss these further. Offline paging [14] derives an optimal page replacement strategy when future page references are known in advance. It cannot be   </text>
<query_num> 11509 </query_num>
<text>   into a single heap ”variable”. Additional techniques such as shape analysis have aimed to identify logical heap structures, such as trees. Finally, in languages with pointers, pointer analysis [24], =-=[52]-=- is able to find all possible heap variables that a particular memory reference can access. Having understood heap variables, let us consider why heap data is difficult to allocate to scratch-pad memo   </text>
<query_num> 11510 </query_num>
<text>   lasses of compiler methods for allocating global and stack variables to scratch-pad exist. First, static allocation methods are those in which the allocation does not change at runtime; these include =-=[10]-=-, [51], [29], [9], [50] and others not listed here. In such methods, the compiler places the most frequently used variables, as revealed by profiling, in scratch pad. Placing a portion of the stack va ion overhead in all embedded systems, but cannot be used to allocate most heap data to scratch-pad. III. RELATED WORK Static methods to allocate data to SPM include [50], [51], [11], [44], [29], [9], =-=[10]-=-. Static methods are those whose SPM allocation does not change at run-time. Some of these methods [50], [11], [44] are restricted to allocating global variables to SPM; others [51], [29], [9], [10] c vantages of SPMs over caches not apparent from the results above. First, it is widely known that for global and stack data, SPMs have significantly better real-time guarantees than caches [62], [53], =-=[10]-=-. Second, other researchers have repeatedly demonstrated a significant energy and run-time savings for benchmarks containing only global and stack data for an SPM vs. a cache of the same area [6], [53   </text>
<query_num> 11511 </query_num>
<text>   n DRAM. An M-core power simulator [13], [12], kindly donated by that group, is used to obtain energy estimates for instructions and SRAM. This is an instruction-level power simulator similar to [49], =-=[55]-=-; its instruction power numbers were measured using an ammeter connected to an M-core hardware board. DRAM power is estimated by a detailed DRAM power simulator we built into the M-core simulator. It   </text>
<query_num> 11512 </query_num>
<text>   ogram in unchangeable ROM, and dynamic compilation cannot be used. Other software caching schemes have been proposed with different goals and/or non-applicable platforms [58], [15], [21], [47], [16], =-=[34]-=-. For lack of space, we do not discuss these further. Offline paging [14] derives an optimal page replacement strategy when future page references are known in advance. It cannot be used for our purpo   </text>
<query_num> 11513 </query_num>
<text>   problem as a knapsack problem or an integer-linear programming problem to find a solution. Some static allocation methods [6], [59] aim to allocate code to SPM rather than data. Other static methods =-=[61]-=-, [53] can allocate both code and data to SPM. Their data allocation is restricted to global and stack data. The the A C B D C E A 4sStep 1. Partition program into regions. /* Section V */ Step 2. Com   </text>
<query_num> 11514 </query_num>
<text>   s, however, store the program in unchangeable ROM, and dynamic compilation cannot be used. Other software caching schemes have been proposed with different goals and/or non-applicable platforms [58], =-=[15]-=-, [21], [47], [16], [34]. For lack of space, we do not discuss these further. Offline paging [14] derives an optimal page replacement strategy when future page references are known in advance. It cann   </text>
<query_num> 11515 </query_num>
<text>   stacks, one for scratch-pad and one for DRAM. Second, recently proposed dynamic methods improve upon static methods by allowing variables to be moved at runtime; we know of two dynamic methods [56], =-=[37]-=-. Being able to move variables enables tailoring the allocation to each region in the program rather than having a fixed allocation as in a static method. Dynamic methods aim to keep variables that ar d just before a region in which they are expected to the frequently accessed. Other variables are evicted to DRAM by explicit copy out instructions to make space for incoming variables. The method in =-=[37]-=- is restricted to arrays accessed through affine functions; the method in [56] is fully general and can place all global and stack variables in scratch pad dynamically, even in the presence of unrestr its different banks in multibanked scratch-pads; and then to turn off (or send to a lower energy state) the banks that are not being actively accessed. Dynamic methods to allocate data to SPM include =-=[37]-=-, [56]; these are methods which can change the SPM allocation during run-time. The method in [37] can place global and stack arrays accessed through affine functions of enclosing loop induction variab   </text>
<query_num> 11516 </query_num>
<text>   store the program in unchangeable ROM, and dynamic compilation cannot be used. Other software caching schemes have been proposed with different goals and/or non-applicable platforms [58], [15], [21], =-=[47]-=-, [16], [34]. For lack of space, we do not discuss these further. Offline paging [14] derives an optimal page replacement strategy when future page references are known in advance. It cannot be used f   </text>
<query_num> 11517 </query_num>
<text>   systems, however, store the program in unchangeable ROM, and dynamic compilation cannot be used. Other software caching schemes have been proposed with different goals and/or non-applicable platforms =-=[58]-=-, [15], [21], [47], [16], [34]. For lack of space, we do not discuss these further. Offline paging [14] derives an optimal page replacement strategy when future page references are known in advance. I   </text>
<query_num> 11518 </query_num>
<text>   taining caches. Section XII concludes. II. CAN WE USE HEAP TO STACK CONVERSION TO ALLOCATE TO SCRATCH PAD? At first glance, it seems that recent work on converting heap data to stack data [18], [19], =-=[22]-=- or to stack-like constructs called regions [17], [26] may help in allocating heap data to scratch-pad. Here is some background on these methods. These methods use escape analysis [45], [60] to try to the fraction of heap data that is of fixed size, and can be converted to stack allocation using escape analysis, is small. Such data can be allocated to scratch pad using our stack/global method, but =-=[22]-=- reports that only 19% of the heap data in their benchmarks could be converted to fixed-size stack data. This low percentage is not surprising – most heap data is in dynamic data structures. Fixed-siz   </text>
<query_num> 11519 </query_num>
<text>   tectures containing caches. Section XII concludes. II. CAN WE USE HEAP TO STACK CONVERSION TO ALLOCATE TO SCRATCH PAD? At first glance, it seems that recent work on converting heap data to stack data =-=[18]-=-, [19], [22] or to stack-like constructs called regions [17], [26] may help in allocating heap data to scratch-pad. Here is some background on these methods. These methods use escape analysis [45], [6   </text>
<query_num> 11520 </query_num>
<text>   these methods use greedy strategies to find a solution; others model the problem as a knapsack problem or an integer-linear programming problem to find a solution. Some static allocation methods [6], =-=[59]-=- aim to allocate code to SPM rather than data. Other static methods [61], [53] can allocate both code and data to SPM. Their data allocation is restricted to global and stack data. The the A C B D C E  of space. These architectures are popular because SPMs are simple to design and verify, and provide better real-time guarantees for global and stack data [62], power consumption, and cost [6], [53], =-=[59]-=-, [11] compared to caches. For these architectures our method delivers runtime and energy reductions averaging 34.6% and 39.9%, respectively, compared to the best previous method. Nevertheless, it is  econd, other researchers have repeatedly demonstrated a significant energy and run-time savings for benchmarks containing only global and stack data for an SPM vs. a cache of the same area [6], [53], =-=[59]-=-, [11]. For these reasons, if an embedded task set contains some tasks of only global and stack data and other tasks having heap data as well, our method will enable the designer to use a SPM alone an   </text>
<query_num> 11521 </query_num>
<text>   ur a significant penalty in area cost, energy, hit latency and real-time guarantees. All of these other than hit latency are more important for embedded systems than desktops. A detailed recent study =-=[11]-=- compares caches with scratch pad. Their results are startling: a scratch pad has 34% smaller area and 40% lower power consumption than a cache of the same capacity. These savings are significant sinc nd energy consumption, a fraction that is increasing with time =-=[11]-=-. Even more surprising, the runtime cycle count they measured was 18% better with a scratch pad using a simple static knapsack-based [11] allocation algorithm, compared to a cache. Defying conventional wisdom, they found absolutely no advantage to using a cache, even in high-end embedded systems in which performance is important. With  bedded CPUs today (e.g., [20], [1], [40], [54], [39]), ahead of caches. Trends in recent embedded designs indicate that the dominance of scratchpad will likely consolidate further in the future [48], =-=[11]-=-, for regular as well as network processors. Although many embedded processors with scratch1spad exist, compiling program data to effectively use the scratch-pad has been a challenge. The challenge is location and de-allocation overhead in all embedded systems, but cannot be used to allocate most heap data to scratch-pad. III. RELATED WORK Static methods to allocate data to SPM include [50], [51], =-=[11]-=-, [44], [29], [9], [10]. Static methods are those whose SPM allocation does not change at run-time. Some of these methods [50], [11], [44] are restricted to allocating global variables to SPM; others  ace. These architectures are popular because SPMs are simple to design and verify, and provide better real-time guarantees for global and stack data [62], power consumption, and cost [6], [53], [59], =-=[11]-=- compared to caches. For these architectures our method delivers runtime and energy reductions averaging 34.6% and 39.9%, respectively, compared to the best previous method. Nevertheless, it is intere   </text>
</top>
<top>
<paper_num> 116 </paper_num>
<paper_title>   Static Analysis on x86 Executables for Preventing Automatic Mimicry Attacks.  </paper_title>
<paper_abstract>   Abstract. In 2005, Kruegel et al. proposed a variation of the traditional mimicry attack, to which we will refer to as automatic mimicry, which can defeat existing system call based HIDS models. We show how such an attack can be defeated by using information provided by the Interprocedural Control Flow Graph (ICFG). Roughly speaking, by exploiting the ICFG of a protected binary, we propose a strategy based on the use of static analysis techniques which is able to localize critical regions inside a program, which are segments of code that could be used for exploiting an automatic mimicry attack. Once the critical regions have been recognized, their code is instrumented in such a way that, during the executions of such regions, the integrity of the dangerous code pointers is monitored, and any unauthorized modification will be restored at once with the legal values. Moreover, our experiments shows that such a defensive mechanism presents a low run-time overhead. 1  </paper_abstract>
<query_num> 11601 </query_num>
<text>   The purpose of this phase is to determine the dangerous regions of p, usingp’s ICFG. In particular, we consider only nodes (basic blocks) that contain dangerous system calls, as defined by Xu. et al. =-=[32]-=-. Our method works as follows. Initially, we build p’s ICFG, then: – each node of the ICFG which contains a dangerous system call, is marked with u (i.e., we determine parameters’ use); – Let p1, ···,   </text>
<query_num> 11602 </query_num>
<text>   authors use the data-flow analysis in order to determine the values of the indirect calls so to improve the completeness of CFG. Instead, for the aliasing problem we can use the algorithm describe in =-=[3]-=-. This technique works on the x86 executable and has obtained good results. However, in future we think to work on the source code of the program in order to solve the problems that binary static anal   </text>
<query_num> 11603 </query_num>
<text>   hey provide a probabilistic defensive mechanism that, in general, cannot provide certainty in protecting from memory errors exploits. In this sense, quite recently, newer approaches have been devised =-=[9,4]-=- that make use of diversified process replicæto provide protection from a broad class of memory error attacks which mainly corrupt application’s code and data pointers. Even if the approach seems soun   </text>
<query_num> 11604 </query_num>
<text>   ich must be addressed when working on an executable binary: (1) the CFG’s completeness and (2) the aliasing problem. In order to improve the CFG’s completeness we can adopt the technique described in =-=[24]-=-. In this approach, the authors use the data-flow analysis in order to determine the values of the indirect calls so to improve the completeness of CFG. Instead, for the aliasing problem we can use th   </text>
<query_num> 11605 </query_num>
<text>   naries information, the tool disassembles the instructions contained inside the code section and converts them to an intermediate form. We used the well-known recursive traversal algorithm defined in =-=[22]-=- to disassemble the binary; – the tool computes the ICFG (§ 3), afterwards the program is converted into the SSA form using the standard Ferrant’s algorithm [8]; – finally, the tool uses the classic e   </text>
<query_num> 11606 </query_num>
<text>   odifying code, run-time code generation, and the unanticipated dynamic loading of code. Program shepherding, proposed by Kiriansky et al., monitors control flow transfers to enforce a security policy =-=[16]-=-. While CFI could be enforced by program shepherding, the approach proposed by Kiriansky et al. is more general. In fact, it prevents execution of data or modified code and ensures that libraries are   </text>
<query_num> 11607 </query_num>
<text>   researchers, who proposed several improvements over the original model, thus obtaining more efficient and more precise (i.e., which recognize broader classes of intrusions) anomaly detection HIDS. In =-=[31,30]-=- Wagner et al. observed that all the system call-based HIDS suffer a particular form of attack called mimicry. In its simplest form, to which we will refer to as traditional mimicry, it basically cons  are given in § 8. 2 Related Works Generally speaking, memory error exploits which corrupt code pointers aim at pursuing two main goals (or a combination of them), that is, (i) to perform IPE attacks =-=[30]-=- (to bypass security critical checks), and (ii) to execute arbitrary malicious code.Static Analysis on x86 Executables for Preventing Automatic Mimicry Attacks 215 Several strategies have been propos   </text>
<query_num> 11608 </query_num>
<text>   t lacks a more fine-grained randomization. Unfortunately, the approach is vulnerable to information leakage attacks or it has been proved to be not so effective on 32-bit Intel Architecture platforms =-=[14]-=-. Other address obfuscation techniques have been proposed in [21,20] by Bhatkar et al. as a particular form of program transformations to combat memory error exploits. Such approaches differ from the   </text>
<query_num> 11609 </query_num>
<text>   to as traditional mimicry, it basically consists of forcing a process to execute an attack vector by mimicking the system calls sequences and learnt by the HIDS. Subsequently, several strategies (see =-=[23,11,5]-=-) have been proposed for inhibiting traditional mimicry attacks. In a recent paper, Kruegel et al. [17] observed that even if the introduction of such techniques in anomaly-based HIDS [5,11,23] has si   </text>
<query_num> 11610 </query_num>
<text>   to as traditional mimicry, it basically consists of forcing a process to execute an attack vector by mimicking the system calls sequences and learnt by the HIDS. Subsequently, several strategies (see =-=[23,11,5]-=-) have been proposed for inhibiting traditional mimicry attacks. In a recent paper, Kruegel et al. [17] observed that even if the introduction of such techniques in anomaly-based HIDS [5,11,23] has si tuid(0); 30 if (execl(cmd, cmd, 0) &amp;lt; 0) { 31 perror(&amp;quot;error: execl&amp;quot;); exit(1); 32 } 33 } 34 } Fig. 2. The Liveness Area of the Parameter cmd system call, but, due to the system call coordinates checks =-=[11]-=-, he cannot neither invoke a system call from an illegal call site nor returning into different location after the system call-aware function termination, or (ii) set enable_logging, overwrite the pri   </text>
<query_num> 11611 </query_num>
<text>   uegel et al. [17] observed that even if the introduction of such techniques in anomaly-based HIDS [5,11,23] has significantly reduced the possibility to perform successful traditional mimicry attacks =-=[26,25,31]-=-, they do not impose any kind of restriction on the execution of arbitrary code which does not directly invoke system B. M. Hämmerli and R. Sommer (Eds.): DIMVA 2007, LNCS 4579, pp. 213–230, 2007. c○   </text>
<query_num> 11612 </query_num>
<text>   zed modification will be restored at once with the legal values. Moreover, our experiments shows that such a defensive mechanism presents a low run-time overhead. 1 Introduction In their seminal work =-=[13,12]-=- about anomaly-based Host Intrusion Detection System (HIDS), Forrest et al., introduced the idea that anomalous behavior of a process p can be detected by learning the sequences of system calls execut   </text>
</top>
<top>
<paper_num> 117 </paper_num>
<paper_title>   A Dependable Distributed Auction System: Architecture and an Implementation Framework.  </paper_title>
<paper_abstract>   The work presented here develops a distributed systems architecture and propose an implementation framework for conducting dependable Internet based on-line auctions, meeting the requirements of scalability and service integrity. Current auction services essentially rely on a central auction server. Given the increasing popularity and usage of electronic auctions, such a centralised approach is fundamentally restrictive with respect to scalability. Further, different national markets have different monetary regulations and may employ different procedures for payment settlements. Catering for local market autonomy means that decentralisation is an essential and practical requirement. With these design goals in mind, the paper develops an approach that permits an auction service to be mapped on to globally distributed auction servers. It then proposes a framework for a fault-tolerant implementation of the architecture. Faulttolerance is achieved through matured technologies: replication management and group paradigm.  </paper_abstract>
<query_num> 11701 </query_num>
<text>   , (ii) displaying the information about the auction details, such as the bidding deadline, the highest bid placed etc., and (iii) the core functionality of processing the bids. We refer the reader to =-=[15, 16]-=- for details on these server tasks and the functionalities they are designed to provide. With the increasing popularity and usage of Internet auctions, the centralised approach cannot clearly scale. F ent Bidder S4 Internet/Private Network S3 S1 S2 local market (a) (b) wired communication Figure 1. (a) Central Server S. (b) Interconnected Local Market Servers local market In this approach taken by =-=[16]-=-, there is a server for a given local market, which exercises policies best suited to local conditions and market mechanisms, e.g. setting bidding deadlines by taking into account of the reliability a bases. By thus letting a local server go global only on a need-to-go basis, scalability and autonomy concerns are addressed. In another model, known as the explicit multicast model (also supported in =-=[16]-=-), the seller of X simultaneously and explicitly requests, right at the start, not just the local server S1 but also a subset of remote markets of his choice, to conduct auctions on X in their respect t to the root of the tree to include the item in the next auction. We here note that selecting a set of markets for defining the auction market is similar to the explicit multicast model supported in =-=[16]-=-; also that we permit the number of selected markets to be unrestricted. In what follows, we assume that the tree has been set-up, containing many local markets. 3.3. System Architecture Information B   </text>
<query_num> 11702 </query_num>
<text>   A1 is essential for process replication and holds true in the context of bid processing; NA2 permits less than one half of the replicas to be faulty. (Without it, only less than a third can be faulty =-=[28]-=-). 4.3. An Implementation Framework We would adopt passive replication strategy to build reliable servers as it would enable a replicated server Si to provide fast responses in the absence of faults.   </text>
<query_num> 11703 </query_num>
<text>   E.offset[k]. After the computation of E, EFT is set to an empty table. Figure 4 gives an example EFT and the corresponding E. EFT: $110 $135 $150 2 1 3 E.base = $110; E.bidders = [2, 1, 3] E.offset = =-=[0, 25, 40]-=- Figure 4. Episode Frequency Table (EFT) and the corresponding E.. If the EFT is empty, the E computed is denoted as ⊥. We will model the process of computing E from EFT as an execution of function Φ: sseminates at most once (i) all bids accepted in the sub-tree rooted on itself, in its Gu; and, (ii) all bids accepted in the rest of the tree and by itself, in its Gd. A proof of this can be seen in =-=[25]-=-. Now, what needs to be shown is that the auction process does not terminate (prematurely) when a server has a non-empty EFT or its Episode message is in transit. Showing this will transform the above e last one will have {highest_bid, highest_bidders} ≡ {$Bmax, N}, where $Bmax is the highest bid amount placed by N bidders when bid placement ceases globally. To see how termination detection works (=-=[25]-=- has rigorous proofs), suppose that the root server decides on termination, with its pair {highest_bid, highest_bidders} being {$300, 5}. This means that none of its local bidders is willing to put a  ctively.) It then announces the winner and the losers. If S9 is the winner, it conducts a weighted draw, and this process of selection continues down the tree recursively. Full details can be seen in =-=[25]-=-. 4. Reliability Issues 4.1. Network Fault Model The distributed auction system described above has two subsystems: servers and the communication network that interconnects them. A server can fail, us   </text>
<query_num> 11704 </query_num>
<text>   bids get to the server before the deadline. Once the deadline is past, the item is sold to the highest bidder (as in English auctions). Some well-known central server systems are eBay and AuctionBot =-=[15]-=-, the latter having been developed as a research system to explore auction mechanisms. The server may be accessed by long- or short-distance bidders via Internet, or by mobile systems, or by agents ex , (ii) displaying the information about the auction details, such as the bidding deadline, the highest bid placed etc., and (iii) the core functionality of processing the bids. We refer the reader to =-=[15, 16]-=- for details on these server tasks and the functionalities they are designed to provide. With the increasing popularity and usage of Internet auctions, the centralised approach cannot clearly scale. F   </text>
<query_num> 11705 </query_num>
<text>   ection: Knowing that the global auction has/has not terminated is equivalent to solving termination detection problem in a distributed setting – which is not a trivial task though algorithms do exist =-=[18, 19]-=- but are typically message expensive. Hence, the system structuring we decide on, must assist solving this problem in economical ways. Market Shrinkage: Imagine that, in a particular local market, the   </text>
<query_num> 11706 </query_num>
<text>   eivers, first presented in [20], assumes that the IP-enabled routers are arranged in a tree (with the router attached to the message sender forming the root). Well-known scaleable transport protocols =-=[21,22,23]-=- use this tree structure to guarantee end-to-end reliability requirements. The analysis of [24] also favours that servers in a large scale setting be arranged in a tree for message efficiency. Assumin   </text>
<query_num> 11707 </query_num>
<text>   erver has both TC1 and TC2 to be true, it decides that bidding has ceased globally and the auction process has terminated. Observe that the Terminated message is similar to the marker message used in =-=[26]-=- for determining a consistent global state of a distributed computation; also that a non-root server may send more than one Terminated message; if so, only the last one will have {highest_bid, highest   </text>
<query_num> 11708 </query_num>
<text>   h other. Currently, we are implementing our architecture, for a system of 3 long-distance servers to start with. (They are situated in Pisa, Newcastle, and London.) Our NewTop group management system =-=[6]-=- (built in Java) provides the basic services which the framework proposed here assumes. We also intend to enhance the system design for multiple sellers, and other types of auctions including double a   </text>
<query_num> 11709 </query_num>
<text>   isting group management middleware systems [3-9] can readily provide. By exploiting the fact that a server is internally replicated (over a synchronous network), we circumvent the unsolveable problem =-=[10]-=- of accurate failure detection in an asynchronous network (e.g. the Internet) which the servers would use to communicate with each other. The paper is organised as follows. Next section describes vari  operating within S7) and joining G. S1 1 and S2 1 (the surviving members of G) should not be entrusted with failure detection, as accurate failure detection is impossible over an aynchronous network =-=[10]-=-. Join operations are usually costly and time-consuming; so, we construct G containing not just the primaries but also the secondaries. The composition of G is shown in figure 6. We assume a reliable  ment techniques which are well established both in theory and practice. By exploiting the fact that a server is internally replicated over a synchronous network, we circumvent the unsolveable problem =-=[10]-=- of accurate failure detection in an asynchronous network (e.g. the Internet) which the servers use to communicate with each other. Currently, we are implementing our architecture, for a system of 3 l   </text>
<query_num> 11710 </query_num>
<text>   server. Seeking tree-based structuring for reasons of scalability is frequently done in the literature. For example, the concept of IP-multicasting for a large number of receivers, first presented in =-=[20]-=-, assumes that the IP-enabled routers are arranged in a tree (with the router attached to the message sender forming the root). Well-known scaleable transport protocols [21,22,23] use this tree struct   </text>
</top>
<top>
<paper_num> 118 </paper_num>
<paper_title>   Towards a Fuzzy Description Logic for the Semantic Web (Preliminary Report).  </paper_title>
<paper_abstract>   In this paper we present a fuzzy version of SHOIN (D), the corresponding Description Logic of the ontology description language OWL DL. We show that the representation and reasoning capabilities of fuzzy SHOIN (D) go clearly beyond classical SHOIN (D). Interesting features are: (i) concept constructors are based on t-norm, t-conorm, negation and implication; (ii) concrete domains are fuzzy sets; (iii) fuzzy modifiers are allowed; and (iv) entailment and subsumption relationships may hold to some degree in the unit interval [0, 1].  </paper_abstract>
<query_num> 11801 </query_num>
<text>   , relative little work has been carried out in extending DLs towards the representation of imprecise concepts, notwithstanding DLs can be considered as a quite natural candidate for such an extension =-=[4, 12, 13, 25, 27, 29, 30, 32, 33]-=- (see also [7], Chapter 6). In this paper we consider a fuzzy extension of SHOIN (D), the corresponding DL of the ontology description language OWL DL, and present its syntax and semantics. The main f   </text>
<query_num> 11802 </query_num>
<text>   , relative little work has been carried out in extending DLs towards the representation of imprecise concepts, notwithstanding DLs can be considered as a quite natural candidate for such an extension =-=[4, 12, 13, 25, 27, 29, 30, 32, 33]-=- (see also [7], Chapter 6). In this paper we consider a fuzzy extension of SHOIN (D), the corresponding DL of the ontology description language OWL DL, and present its syntax and semantics. The main f  fuzzy ALC with concept modifiers of the form fm(x) = x β , where β &amp;gt; 0. A sound and complete reasoning algorithm for the graded subsumption problem, based on completion rules, is presented. Finally, =-=[25]-=- start addressing the issue of alternative semantics of quantifiers in fuzzy ALC (without the assertional component). No reasoning algorithm is given. Concerning [31], we already addressed it in the p uzzy SHOIN (D) with fuzzy quantifiers (see [19, 20] for an overview on fuzzy quantifiers), where the ∀ and ∃ quantifiers are replaced with fuzzy quantifiers like most, some, usually and the like (see =-=[25]-=- for a preliminary work in this direction). This allows to define concepts like TopCustomer = Customer ⊓ (Usually)buys.ExpensiveItem ExpensiveItem = Item ⊓ ∃price.High . Here, the fuzzy quantifier Usu   </text>
<query_num> 11803 </query_num>
<text>   , relative little work has been carried out in extending DLs towards the representation of imprecise concepts, notwithstanding DLs can be considered as a quite natural candidate for such an extension =-=[4, 12, 13, 25, 27, 29, 30, 32, 33]-=- (see also [7], Chapter 6). In this paper we consider a fuzzy extension of SHOIN (D), the corresponding DL of the ontology description language OWL DL, and present its syntax and semantics. The main f cation operator i. 3.2 Fuzzy SHOIN (D) In this section we give syntax and semantics of fuzzy SHOIN (D), using the fuzzy operators defined in the previous section. We generalize the semantics given in =-=[13, 29, 32]-=-. 3.2.1 Syntax We have seen that SHOIN (D) allows to reason with concrete data types, such as strings and integers using so-called concrete domains. In our fuzzy approach, concrete domains may be base embership function. Formally, a modifier is a function fm: [0, 1] → [0, 1]. For instance, we may define very(x) = x 2 , while define slightly(x) = √ x. Modifiers has been considered, for instance, in =-=[13, 32]-=-. From a syntax point of view, if M is a new alphabet for modifier symbols, m ∈ M is a modifier and C is a SHOIN (D) concept, then m(C) is fuzzy SHOIN (D) concept as well. For instance, by referring t  n〉 iff 〈T , R, A ∪ {〈α &amp;gt; n〉}〉 is not satisfiable . 3.3 Reasoning Unfortunately, from a computational point of view, no calculus exists yet checking satisfiability of fuzzy SHOIN (D) knowledge bases. =-=[13, 32]-=- report a calculus for the case of ALC [26] (with concept constructors ⊤, ⊥, ¬, ⊓, ⊔, ∀, ∃) with modifiers and simple TBox, with min, max and →KD connectives. No indication for the BDB problem is give  of ALC, FL − [5]. However, it already informally talks about the use of modifiers and concrete domains. Though, the unique reasoning facility, the subsumption test, is a crisp yes/no question. Tresp =-=[32]-=- considered fuzzy ALC extended with a special form of modifiers, which are a combination of two linear functions. min, max and 1−x membership functions has been considered 16sand a sound and complete   </text>
<query_num> 11804 </query_num>
<text>   , relative little work has been carried out in extending DLs towards the representation of imprecise concepts, notwithstanding DLs can be considered as a quite natural candidate for such an extension =-=[4, 12, 13, 25, 27, 29, 30, 32, 33]-=- (see also [7], Chapter 6). In this paper we consider a fuzzy extension of SHOIN (D), the corresponding DL of the ontology description language OWL DL, and present its syntax and semantics. The main f n b ≥ m. Indeed, from i(a, b) = max(1 − a, b) ≥ m, either 1 − a ≥ m or b ≥ m. But a ≥ n, so 1 − a ≥ m implies 1 − m ≥ a ≥ n &amp;gt; 1 − m, a contradiction. Therefore, b ≥ m must hold. This has been used in =-=[29]-=-. • under residuum based implication w.r.t. a t-norm t, we infer that b ≥ t(n, m). Indeed, from i(a, b) = sup{c: t(a, c) ≤ b} ≥ m and a ≥ n we have t(n, m) ≤ t(n, c) ≤ t(a, c) ≤ b. A (binary) fuzzy re cation operator i. 3.2 Fuzzy SHOIN (D) In this section we give syntax and semantics of fuzzy SHOIN (D), using the fuzzy operators defined in the previous section. We generalize the semantics given in =-=[13, 29, 32]-=-. 3.2.1 Syntax We have seen that SHOIN (D) allows to reason with concrete data types, such as strings and integers using so-called concrete domains. In our fuzzy approach, concrete domains may be base ⊔ C2 | ¬C | m(C) ∀R.C | ∃R.C | (≥ n S) | (≤ n S) | {a1, . . . , an} | (≥ n T ) | (≤ n T ) | ∀T1, . . . , Tn.D | ∃T1, . . . , Tn.D D −→ d | {c1, ..., cn} Concerning axioms and assertions, similarly to =-=[29]-=-, we introduce fuzzy axioms. Let be n ∈ (0, 1]. 11sFuzzy RBox. A fuzzy RBox R is a finite set of SHOIN (D) transitivity axioms trans(R) and fuzzy role inclusion axioms of the form 〈α ≥ n〉, 〈α ≤ n〉, 〈α een C and D is at least n. Fuzzy knowledge base. A SHOIN (D) fuzzy knowledge base K = 〈T , R, A〉 consists of a fuzzy TBox T , a fuzzy RBox R, and a fuzzy ABox A. 3.2.2 Semantics The semantics extends =-=[29]-=-. The main idea is that concepts and roles are interpreted as fuzzy subsets of an interpretation’s domain. Therefore, SHOIN (D) axioms, rather being satisfied (true) or unsatisfied (false) in an inter   </text>
<query_num> 11805 </query_num>
<text>   , relative little work has been carried out in extending DLs towards the representation of imprecise concepts, notwithstanding DLs can be considered as a quite natural candidate for such an extension =-=[4, 12, 13, 25, 27, 29, 30, 32, 33]-=- (see also [7], Chapter 6). In this paper we consider a fuzzy extension of SHOIN (D), the corresponding DL of the ontology description language OWL DL, and present its syntax and semantics. The main f port a calculus for the case of ALC [26] (with concept constructors ⊤, ⊥, ¬, ⊓, ⊔, ∀, ∃) with modifiers and simple TBox, with min, max and →KD connectives. No indication for the BDB problem is given. =-=[27, 29]-=- reports a calculus for ALC and simple TBox, with min, max and →KD connectives and addresses the BDB problem and, [30] shows how the satisfiability problem and the BDB problem can be reduced to classi lete reasoning algorithm testing the subsumption relationship has been presented. Similar to our approach, a linear programming oracle is needed. Assertional reasoning has been considered by Straccia =-=[27, 28, 29]-=-, where fuzzy assertion axioms have been allowed in fuzzy ALC (with min, max and 1 − x functions), concept modifiers are not allowed however ([28] reports a four-valued variant of fuzzy ALC). He also   </text>
<query_num> 11806 </query_num>
<text>   [16]. Although several XML and RDF syntaxes for OWL-DL exist, in this paper we use the traditional description logic notation. For explicating the relationship between OWL DL and DLs syntax, see e.g. =-=[15, 16]-=-. The purpose of this section is to make the paper self-contained. More importantly it helps in understanding 1 Taken from a text book on flowers. 2sthe differences between classical SHOIN (D) and fuz  is a model of each component T , R and A, respectively. Logical consequence. An axiom E is a logical consequence of a knowledge base K, denoted K |= E, iff every model of K satisfies E. According to =-=[15]-=-, the entailment and subsumption problem can be reduced to knowledge base satisfiability problem (e.g. 〈T , R, A〉 |= a:C iff 〈T , R, A∪{a:¬C}〉 unsatisfiable), for which decision procedures and reasoni   </text>
<query_num> 11807 </query_num>
<text>   capture the meaning of the most popular features of structured representation of knowledge. Nowadays, DLs have gained even more popularity due to their application in the context of the Semantic Web =-=[3, 16]-=-. Semantic Web has recently attracted much attention both from academia and industry, and is widely regarded as the next step in the evolution of the World Wide Web. It aims at enhancing content on th   </text>
<query_num> 11808 </query_num>
<text>   capture the meaning of the most popular features of structured representation of knowledge. Nowadays, DLs have gained even more popularity due to their application in the context of the Semantic Web =-=[3, 16]-=-. Semantic Web has recently attracted much attention both from academia and industry, and is widely regarded as the next step in the evolution of the World Wide Web. It aims at enhancing content on th s of it. Section 4 presents related work, while Section 5 concludes and presents some topics for further research. 2 Preliminaries The ontology language OWL DL is strictly related to the DL SHOIN (D) =-=[16]-=-. Although several XML and RDF syntaxes for OWL-DL exist, in this paper we use the traditional description logic notation. For explicating the relationship between OWL DL and DLs syntax, see e.g. [15,   </text>
<query_num> 11809 </query_num>
<text>   confident with the SHOIN (D) terminology may skip directly to Section 3. 2.1 Syntax SHOIN (D) allows to reason with concrete data types, such as strings and integers using so-called concrete domains =-=[2, 21, 22, 23]-=-. Concrete domains. A concrete domain D is a pair 〈∆D, ΦD〉, where ∆D is an interpretation domain and ΦD is the set of concrete domain predicates d with a predefined arity n and an interpretation d D ⊆   </text>
<query_num> 11810 </query_num>
<text>   e also introduced the BDB problem and provided a sound and complete reasoning algorithm based on completion rules ([30] provides a translation of fuzzy ALC into classical ALC). For an application see =-=[24]-=-. In the same spirit [13] extend Straccia’s fuzzy ALC with concept modifiers of the form fm(x) = x β , where β &amp;gt; 0. A sound and complete reasoning algorithm for the graded subsumption problem, based o   </text>
<query_num> 11811 </query_num>
<text>   e evolution of the World Wide Web. It aims at enhancing content on the World Wide Web with meta-data, enabling agents (machines or human users) to process, share and interpret Web content. Ontologies =-=[9]-=- play a key role in the Semantic Web and major effort has been put by the Semantic Web community into this issue. Informally, an ontology consists of a hierarchical description of important concepts i   </text>
<query_num> 11812 </query_num>
<text>   ith min, max and →KD connectives. No indication for the BDB problem is given. [27, 29] reports a calculus for ALC and simple TBox, with min, max and →KD connectives and addresses the BDB problem and, =-=[30]-=- shows how the satisfiability problem and the BDB problem can be reduced to classical ALC and, thus, can be resolved by means of a tools like FACT and RACER. However, despite these negative results, r odifiers are not allowed however ([28] reports a four-valued variant of fuzzy ALC). He also introduced the BDB problem and provided a sound and complete reasoning algorithm based on completion rules (=-=[30]-=- provides a translation of fuzzy ALC into classical ALC). For an application see [24]. In the same spirit [13] extend Straccia’s fuzzy ALC with concept modifiers of the form fm(x) = x β , where β &amp;gt; 0.   </text>
<query_num> 11813 </query_num>
<text>   tion problem can be reduced to knowledge base satisfiability problem (e.g. 〈T , R, A〉 |= a:C iff 〈T , R, A∪{a:¬C}〉 unsatisfiable), for which decision procedures and reasoning tools exists (e.g. RACER =-=[10]-=- and FACT [14]). Example 1 Let us consider the following excerpt of a simple ontology (TBox T ) about cars, with empty RBox (R = ∅): Car ⊑ (= 1 maker) ⊓ (= 1 passenger) ⊓ (= 1 speed) (= 1 maker) ⊑ Car   </text>
</top>
<top>
<paper_num> 119 </paper_num>
<paper_title>   Distributed Computing Column 38: Models for algorithm design in wireless networks.  </paper_title>
<paper_abstract>   The prosperity of research on wireless communication has not skipped the distributed computing community. Wireless networks provide unique and challenging platforms for distributed computation, inspiring researchers to develop many distributed algorithms for such environments. Any algorithmic work targeting wireless networks must begin by defining an appropriate model. The first research results in this vein employed simplified models, like the Unit Disk Graph (UDG), which facilitated algorithm design and were readily amenable to analysis. Recently, models that more accurately capture the physical nature of wireless networks were developed; such models are nowadays being gradually adopted. This column’s contribution, by Zvi (Zvika) Lotker and David Peleg focuses on the promising signalto-interference &amp; noise ratio (SINR) model, and studies it from an algorithmic perspective. It surveys the model’s structural properties as well as algorithms designed for this model. Along the way, some of the similarities and differences between the SINR model and the UDG model are highlighted. Many thanks to Zvika and David for their contribution! Call for contributions: I welcome suggestions for material to include in this column, including news, reviews, opinions, open problems, tutorials and surveys, either exposing the community to new and interesting topics, or providing new insight on well-studied topics by organizing them in new ways.  </paper_abstract>
<query_num> 11901 </query_num>
<text>   , referred to as paths, rather than a single pair of nodes. In the context of routing, the problem of constructing minimum delay end-to-end schedules for a given set of routing requests is studied in =-=[4]-=-. In this problem, each node is assigned a distinct power level, the paths for all requests are determined, and all message transmissions are scheduled to guarantee successful reception in the SINR mo   </text>
<query_num> 11902 </query_num>
<text>   a (b) Figure 5: As α → ∞, the SINR diagram converges (a) to the Voronoi diagram, when the background noise N = 0; (b) to the structure of alpha complexes, when N &amp;gt; 0. SINR diagrams were introduced in =-=[2]-=-, where it is shown that in uniform networks with α = 2 and β ≥ 1, every reception region in the SINR diagram is convex and fat 2 . This “niceness” of the reception regions may potentially make it eas hen 0 &amp;lt; β &amp;lt; 1, the reception regions may be nonconvex and even overlapping, as can be seen in figure 2(b). The characterization of the reception regions as convex and fat is subsequently exploited in =-=[2]-=- for developing an efficient structure for answering point location queries, by preprocessing the station locations (in polynomial time) and providing each station with an additional data structure of   </text>
<query_num> 11903 </query_num>
<text>   ecide also on the power assignment for the transmitters (again, possibly adhering to some specific policy). Schedules for basic network structures, namely, strongly connected networks, are studied in =-=[19, 21]-=-. It is shown that allowing arbitrary adjustments to the transmission power gives an exponential advantage over the uniform or linear power assignment schemes. (The latter require a transmitter s send   </text>
<query_num> 11904 </query_num>
<text>   ecide also on the power assignment for the transmitters (again, possibly adhering to some specific policy). Schedules for basic network structures, namely, strongly connected networks, are studied in =-=[19, 21]-=-. It is shown that allowing arbitrary adjustments to the transmission power gives an exponential advantage over the uniform or linear power assignment schemes. (The latter require a transmitter s send vable worst-case performance for the problem in this setting is presented in [4]. Another line of research, in which known results from the UDG model are analyzed under the SINR model, is explored in =-=[21]-=-, which studies the problem of topology control. The stations are assumed to be embedded in the Euclidean plane, and are required to simulate a given underlying graph topology. For each station v, let  length of the longest outgoing edge of v, and consider a ball of radius rv centered at v. Denote the number of nodes contained in this ball by Iin(v), and let Iin = maxv∈V Iin(v). The main result of =-=[21]-=- is that there exists a schedule that allows the communication to flow on the original topology and completes in time O(Iin log 2 (n)). SINR diagrams An issue of increasing significance is that of obt   </text>
<query_num> 11905 </query_num>
<text>   ed as reflective mirrors. The uncertainly involved in the dynamics of a radio channel can be modeled using a Markov process. For more information we refer the interested reader to Chapters 2 and 3 of =-=[9]-=-. For the purposes of the current discussion, we follow the approach of ignoring those complicating factors, and assuming a relatively clean abstract setting where the only players are the transmittin   </text>
<query_num> 11906 </query_num>
<text>   is means that the capacity increases when we allow transmission power to vary between the different sides of a communication channel. The problem of optimizing a single scheduling step was studied in =-=[1, 6]-=-. Given a collection of senderreceiver pairs (si, ti) in the plane, the goal is to assign each transmitter si a power (possibly 0) so as to maximize the number of successful pairs. Let ∆ = maxi d(si,  an appropriately defined game every Nash equilibrium has an expected number of successful connections that is within O(∆2α ) of optimal. The main result ACM SIGACT News 80 June 2010 Vol. 41, No. 2of =-=[6]-=- is that if all transmitters use no-regret algorithms to play the game defined in [1], then the average number of successful connections (over a sufficient number of rounds) will be within O(∆2α ) of   </text>
<query_num> 11907 </query_num>
<text>   ound on the resource; see [1, 3, 13, 22]. Capacity and scheduling in the SINR model Some recent studies aim at achieving a better understanding of the SINR model. In particular, in their seminal work =-=[12]-=-, Gupta and Kumar analyze the capacity of wireless networks in the physical and protocol models. Specifically, they consider a setting where n transmitter-receiver pairs are located in the unit square (The latter require a transmitter s sending a message to a receiver r to use power proportional to d(s, r) α , the distance to power α.) This gives an interesting complement to the capacity bounds of =-=[12, 17]-=- discussed above. A measure called disturbance, capturing the intrinsic difficulty of finding a short schedule for a problem instance, is defined in [18], along with an algorithm that achieves provabl   </text>
<query_num> 11908 </query_num>
<text>   pects of scheduling in the SINR model, and in particular presents an O(log(n)) approximation algorithm for scheduling wireless requests. On the negative side, some impossibility results are proved in =-=[11]-=-, and the NP-hardness of the scheduling problem in the SINR model is established in [10, 23]. In summary, the literature reviewed above presents several examples where the SINR model allows better per   </text>
<query_num> 11909 </query_num>
<text>   resource (e.g., the used energy or the general area in which the network resides) is bounded, then the ratio between the two models is proportional to log B, where B is the bound on the resource; see =-=[1, 3, 13, 22]-=-. Capacity and scheduling in the SINR model Some recent studies aim at achieving a better understanding of the SINR model. In particular, in their seminal work [12], Gupta and Kumar analyze the capaci   </text>
<query_num> 11910 </query_num>
<text>   resource (e.g., the used energy or the general area in which the network resides) is bounded, then the ratio between the two models is proportional to log B, where B is the bound on the resource; see =-=[1, 3, 13, 22]-=-. Capacity and scheduling in the SINR model Some recent studies aim at achieving a better understanding of the SINR model. In particular, in their seminal work [12], Gupta and Kumar analyze the capaci is means that the capacity increases when we allow transmission power to vary between the different sides of a communication channel. The problem of optimizing a single scheduling step was studied in =-=[1, 6]-=-. Given a collection of senderreceiver pairs (si, ti) in the plane, the goal is to assign each transmitter si a power (possibly 0) so as to maximize the number of successful pairs. Let ∆ = maxi d(si,  sful connections that is within O(∆2α ) of optimal. The main result ACM SIGACT News 80 June 2010 Vol. 41, No. 2of [6] is that if all transmitters use no-regret algorithms to play the game defined in =-=[1]-=-, then the average number of successful connections (over a sufficient number of rounds) will be within O(∆2α ) of optimal. This is the first distributed algorithm (in some sense) for this problem. In   </text>
<query_num> 11911 </query_num>
<text>   sociated with each station, the smaller representing its reception region and the larger representing its area of interference. Another interference model, also based on the UDG model, is proposed in =-=[24]-=-. One way to bridge the gap between the two models is using emulation. Lebhar et al. [16] consider the case of α &amp;gt; 2 and nodes that are deployed uniformly at random in a given area. For this setting,   </text>
<query_num> 11912 </query_num>
<text>   sting complement to the capacity bounds of [12, 17] discussed above. A measure called disturbance, capturing the intrinsic difficulty of finding a short schedule for a problem instance, is defined in =-=[18]-=-, along with an algorithm that achieves provably efficient performance (in terms of schedule length) in any network and request setting that exhibits low disturbance. For the special case of many-to-o   </text>
<query_num> 11913 </query_num>
<text>   the distance between the transmitter and the intended receiver. (In particular, E(r) = c for constant c is a function determined by the distances, so uniform networks employ an oblivious scheme.) In =-=[7]-=-, Fanghänel et al. prove that there is an exponential gap between the directed and bidirectional versions of the scheduling problem. Specifically, they show that in the case where each pair of points   </text>
<query_num> 11914 </query_num>
<text>   then the average number of successful connections (over a sufficient number of rounds) will be within O(∆2α ) of optimal. This is the first distributed algorithm (in some sense) for this problem. In =-=[8]-=-, Fanghänel et al. introduce an instance-based measure I of interference, and prove for general power assignments in the plane a lower bound of Ω( ) steps for α &amp;gt; 2. They also show that when restricte   </text>
</top>
<top>
<paper_num> 120 </paper_num>
<paper_title>   SafeQ: Secure and Efficient Query Processing in Sensor Networks.  </paper_title>
<paper_abstract>   Abstract—The architecture of two-tiered sensor networks, where storage nodes serve as an intermediate tier between sensors and a sink for storing data and processing queries, has been widely adopted because of the benefits of power and storage saving for sensors as well as the efficiency of query processing. However, the importance of storage nodes also makes them attractive to attackers. In this paper, we propose SafeQ, a protocol that prevents attackers from gaining information from both sensor collected data and sink issued queries. SafeQ also allows a sink to detect compromised storage nodes when they misbehave. To preserve privacy, SafeQ uses a novel technique to encode both data and queries such that a storage node can correctly process encoded queries over encoded data without knowing their values. To preserve integrity, we propose a new data structure called neighborhood chains that allows a sink to verify whether the result of a query contains exactly the data items that satisfy the query. In addition, we propose a solution to adapt SafeQ for event-driven sensor networks.  </paper_abstract>
<query_num> 12001 </query_num>
<text>   -th dimension is the end of the chain, this property would be violated. To reduce the communication cost between sensors and storage nodes, we propose an optimization technique based on Bloom filters =-=[34]-=-. Our basic idea is to use a Bloom filter to represent the HMAC hashes HMAC g(N (S([d0,d1]))), ···, HMAC g(N (S([dn,dn+1]))). Thus, a sensor only needs to send the Bloom filter instead of the hashes.   </text>
<query_num> 12002 </query_num>
<text>   AC g(N (F(b)))}, the storage node processes this query on the n data items (d1)ki , ···, (dn)ki received from each nearby sensor si at time slot t based on the following theorem, proof of which is in =-=[35]-=-. Theorem 4.1: Given n numbers sorted in the ascending order d1 &amp;lt; ··· &amp;lt;dn, where dj ∈ (d0,dn+1) (1≤j ≤ n), and a range [a, b] (d0 &amp;lt;a≤ b&amp;lt;dn+1), dj ∈ [a, b] if and only if there exist 1 ≤ n1 ≤ j&amp;lt;n2≤ n + d of the hashes. The number of bits needed to represent the Bloom filter is much smaller than that needed to represent the hashes. Due to space limitations, for details of the technique, please check =-=[35]-=-. VII. RANGE QUERIES IN EVENT-DRIVEN NETWORKS So far we have assumed that at each time slot, a sensor sends to a storage node the data that it collected at that time slot. However, this assumption doe   </text>
<query_num> 12003 </query_num>
<text>   Basic idea of SafeQ for preserving privacy A. Prefix Membership Verification The basic building block of our privacy preserving scheme is the prefix membership verification scheme first introduced in =-=[27]-=- and later formalized in [28]. The key idea of the prefix membership verification scheme is to convert the verification of whether a number is in a range to several verifications of whether two number   </text>
<query_num> 12004 </query_num>
<text>   D1), ···, MNC (Dn), MNC (Dn+1) and a subquery [a l ,b l ] along dimension l, theright bounding item of [a l ,b l ] is the item MNC (Dj) where L l (d l j ) ≤ bl &amp;lt; d l j . Figure 5 shows a query ([2,6],=-=[3,8]-=-) with a query result QR = {MNC (3, 5), MNC (6, 8)} and VO= {MNC (7, 1)}. The Sink: Upon receiving QR and VO, the sink verifies the integrity of QR as follows. First, it verifies that every item in QR   </text>
<query_num> 12005 </query_num>
<text>   EE INFOCOM 2010. However, they have the same two drawbacks as we discussed above. Boneh and Waters proposed a public-key system for supporting conjunctive, subset, and range queries on encrypted data =-=[17]-=-. Although theoretically this seems possible, Boneh&amp;Waters’ scheme cannot be used to solve the privacy problem in our context because it is too computationally expensive for sensor networks. It would   </text>
<query_num> 12006 </query_num>
<text>   Merkle hash trees have been used for the authentication of data elements [24] and they were used for verifying the integrity of database queries in [18], [19]. Pang et al. [20] and Narasimha &amp; Tsudik =-=[21]-=- proposed similar schemes for verifying the integrity of relational database query results using signature aggregation and chaining. For each tuple in a database, Pang et al. computed the signature of lf as well as the tuple’s left and right neighbors [20]. Narasimha &amp; Tsudik computed the signature by signing the concatenation of the digests of the tuple and its left neighbors along each dimension =-=[21]-=-. Although our neighborhood chaining technique seems similar to the above signature aggregation and chaining technique, it is much more efficient and suitable for sensor networks because of the follow   </text>
<query_num> 12007 </query_num>
<text>   atabase-as-service model (DAS) where sensitive data are outsourced to an untrusted server [12]. Agrawal et al. further used the bucket partitioning idea to investigate range queries on numerical data =-=[14]-=-. Hore et al. explored the optimal partitioning of buckets [13]. Authorized licensed use limited to: Michigan State University. Downloaded on July 09,2010 at 01:39:51 UTC from IEEE Xplore. Restriction   </text>
<query_num> 12008 </query_num>
<text>   e, it incurs too much communication cost between sensors and storage nodes. D. Secure File Systems on Untrusted Servers Secure file systems on untrusted servers have been studied in prior work (e.g., =-=[25]-=-, [26]), which aims to design a system where users can securely store their files on an untrusted server and the server cannot read the content of the files. These solutions cannot solve our secure ra   </text>
<query_num> 12009 </query_num>
<text>   hemes, a compromised sensor cannot jeopardize the querying and verification of data collected by other sensors. B. Privacy Preserving in Databases Database privacy has been studied in prior work [12]–=-=[16]-=-. Hacigumus et al. first proposed the bucket partitioning idea for querying encrypted data in the database-as-service model (DAS) where sensitive data are outsourced to an untrusted server [12]. Agraw   </text>
<query_num> 12010 </query_num>
<text>   imensional data by dividing the domain of each dimension into multiple buckets. S&amp;L scheme has two main drawbacks, which are inherited from the bucket partitioning technique. First, as pointed out in =-=[13]-=-, the bucket partitioning technique allows compromised storage nodes to obtain a reasonable estimation on the actual value of both data items and queries. In comparison, in SafeQ, such estimations are ced to an untrusted server [12]. Agrawal et al. further used the bucket partitioning idea to investigate range queries on numerical data [14]. Hore et al. explored the optimal partitioning of buckets =-=[13]-=-. Authorized licensed use limited to: Michigan State University. Downloaded on July 09,2010 at 01:39:51 UTC from IEEE Xplore. Restrictions apply.This full text paper was peer reviewed at the directio   </text>
<query_num> 12011 </query_num>
<text>   incurs too much communication cost between sensors and storage nodes. D. Secure File Systems on Untrusted Servers Secure file systems on untrusted servers have been studied in prior work (e.g., [25], =-=[26]-=-), which aims to design a system where users can securely store their files on an untrusted server and the server cannot read the content of the files. These solutions cannot solve our secure range qu   </text>
<query_num> 12012 </query_num>
<text>   ing becomes more efficient because the sink only communicates with storage nodes for queries. The inclusion of storage nodes in sensor networks was first introduced in [1] and has been widely adopted =-=[2]-=-–[6]. Several products of storage nodes, such as StarGate [7] and RISE [8], are commercially available. However, the inclusion of storage nodes also brings significant security challenges. As storage   of the item and (dj|dj+1)ki the right neighbor of the item. Figure 4 shows the neighborhood chain for the 5 data items 1, 3, 5, 7 and 9. (d0|1) ki (1|3) ki (3|5) ki (5|7) ki (7|9) ki (9|d6) ki Range =-=[2, 8]-=- Query result Verification object Fig. 4. An example neighborhood chain Preserving query result integrity using neighborhood chaining works as follows. After collecting n data items d1, ···,dn, sensor  MNC (D1), ···, MNC (Dn), MNC (Dn+1) and a subquery [a l ,b l ] along dimension l, theright bounding item of [a l ,b l ] is the item MNC (Dj) where L l (d l j ) ≤ bl &amp;lt; d l j . Figure 5 shows a query (=-=[2,6]-=-,[3,8]) with a query result QR = {MNC (3, 5), MNC (6, 8)} and VO= {MNC (7, 1)}. The Sink: Upon receiving QR and VO, the sink verifies the integrity of QR as follows. First, it verifies that every item   </text>
<query_num> 12013 </query_num>
<text>   mension. Here D could be large and each encryption is expensive due to the use of public key cryptography. C. Integrity Preserving in Databases Database integrity has also been explored in prior work =-=[18]-=-–[23], independent of the privacy issues. The focus of such work is on verifying the completeness of the result of relational database queries, such as select, join, set union, and set intersect. Merk   </text>
<query_num> 12014 </query_num>
<text>   on. Here D could be large and each encryption is expensive due to the use of public key cryptography. C. Integrity Preserving in Databases Database integrity has also been explored in prior work [18]–=-=[23]-=-, independent of the privacy issues. The focus of such work is on verifying the completeness of the result of relational database queries, such as select, join, set union, and set intersect. Merkle ha onical Range Trees (CRTs) to store the counting information for multi-dimensional data such that these counting information can be used for integrity verification without leaking boundary information =-=[23]-=-. However, protecting boundary information is unnecessary in our context because the sink has the right to all data collected by sensors. Therefore, the price for protecting boundary information is un   </text>
<query_num> 12015 </query_num>
<text>   rage nodes. Third, query processing becomes more efficient because the sink only communicates with storage nodes for queries. The inclusion of storage nodes in sensor networks was first introduced in =-=[1]-=- and has been widely adopted [2]–[6]. Several products of storage nodes, such as StarGate [7] and RISE [8], are commercially available. However, the inclusion of storage nodes also brings significant   </text>
<query_num> 12016 </query_num>
<text>   rving privacy A. Prefix Membership Verification The basic building block of our privacy preserving scheme is the prefix membership verification scheme first introduced in [27] and later formalized in =-=[28]-=-. The key idea of the prefix membership verification scheme is to convert the verification of whether a number is in a range to several verifications of whether two numbers are equal. A prefix {0, 1}   </text>
<query_num> 12017 </query_num>
<text>   s select, join, set union, and set intersect. Merkle hash trees have been used for the authentication of data elements [24] and they were used for verifying the integrity of database queries in [18], =-=[19]-=-. Pang et al. [20] and Narasimha &amp; Tsudik [21] proposed similar schemes for verifying the integrity of relational database query results using signature aggregation and chaining. For each tuple in a d   </text>
<query_num> 12018 </query_num>
<text>   t union, and set intersect. Merkle hash trees have been used for the authentication of data elements [24] and they were used for verifying the integrity of database queries in [18], [19]. Pang et al. =-=[20]-=- and Narasimha &amp; Tsudik [21] proposed similar schemes for verifying the integrity of relational database query results using signature aggregation and chaining. For each tuple in a database, Pang et a   </text>
<query_num> 12019 </query_num>
<text>   whether a number a is in a range [d1,d2], we first convert the range [d1,d2] to a minimum set of prefixes, denoted S([d1,d2]), such that the union of the prefixes is equal to [d1,d2]. For example, S(=-=[11, 15]-=-) ={1011,11**}. Givena range [d1,d2], where d1 and d2 are two numbers of w bits, the number of prefixes in S([d1,d2]) is at most 2w − 2 [29]. Second, we compute the prefix family F(a) for number a. Th of prefixes S, weuseN (S) to denote the resulting set of numericalized prefixes. Therefore, a ∈ [d1,d2] if and only if N (F(a)) ∩N(S([d1,d2])) ̸= ∅. Figure 3 illustrates the process of verifying 12 ∈ =-=[11, 15]-=-. [11, 15] ⇓ Prefix conversion 1011 11** ⇓ Prefix numericalization 10111 11100 (a) Fig. 3. 12 (=1100) ⇓ Prefix family construction 1100 11** **** 110* 1*** ⇓ Prefix numericalization 11001 11100 10000   </text>
</top>
<top>
<paper_num> 121 </paper_num>
<paper_title>   SODA: A Service-On-Demand Architecture for Application Service Hosting Utility Platforms.  </paper_title>
<paper_abstract>   as utility: computation jobs can be scheduled on-demand in Grid hosts based on available computation capacity. In this paper, we study another emerging usage of Grid utility: the hosting of application services. Different from a computation job, an application service such as e-Laboratory or on-line shopping has longer lifetime, and performs multiple jobs requested by its clients. A service Hosting Utility Platform (HUP) is formed by a set of servers in the Grid, and multiple application services will be hosted on the HUP. We present the design and implementation of SODA, a Service-On-Demand Architecture that enables on-demand creation of application services on a HUP. With SODA, an application service will be created in the form of a set of virtual service nodes; each node is a virtual machine which is physically a `slice&amp;apos; of a real host in the HUP. SODA involves both OS and middleware level techniques, and has the following salient capabilities: (1) on-demand service priming: the image of an application service as well as the OS on which it runs will be created on-demand and bootstrapped automatically; (2) better service isolation: services sharing the same HUP host are isolated with respect to administration, faults, attacks, and resources; (3) integrated service request management: for each service, a service switch will be created to direct client requests to appropriate virtual service nodes. Moreover, the application service provider can replace the default request switching policy with a service-specific policy.  </paper_abstract>
<query_num> 12101 </query_num>
<text>   As a result, services created by SODA enjoy better isolation. Resource isolation has been realized by existing techniques of resource reservation, allocation, and scheduling, such as those in QLinux =-=[32]-=-, Resource Kernel [30], GARA [21], and Virtual Services [31]. Resources such as CPU, bandwidth, and memory can be reserved and allocated to different processes, users, or service classes; and the allo   </text>
<query_num> 12102 </query_num>
<text>   Harness [27] and Cactus [4]. Meanwhile, the Grid resources can also be used as a data storage and management utility, such as in the experiments.sStorage Resource Broker [9], NeST [10], the Data Grid =-=[14]-=-, and OceanStore [26]. Industry efforts parallel those of the academia, such as Oceano [6] at IBM and Planetary Computing [2] at HP. In this paper, we focus on application service hosting as another e   </text>
<query_num> 12103 </query_num>
<text>   ], Legion [24], NetSolve [7], Harness [27] and Cactus [4]. Meanwhile, the Grid resources can also be used as a data storage and management utility, such as in the experiments.sStorage Resource Broker =-=[9]-=-, NeST [10], the Data Grid [14], and OceanStore [26]. Industry efforts parallel those of the academia, such as Oceano [6] at IBM and Planetary Computing [2] at HP. In this paper, we focus on applicati   </text>
<query_num> 12104 </query_num>
<text>   d and allocated to different processes, users, or service classes; and the allocation is enforced by various resource scheduling algorithms. Resource partitioning is also proposed for server clusters =-=[8]-=- and overlay networks [11], where slices of resources in multiple hosts are reserved for different applications or services. However, these works do not address other aspects of isolation such as admi   </text>
<query_num> 12105 </query_num>
<text>   e service software runs. It is challenging to enable automatic bootstrapping of both the service and the guest OS on top of the host OS in a HUP host. Unfortunately, current active service techniques =-=[5, 34]-=- are not adequate to handle such an ‘active virtual machine’ scenario. In this paper, we present the design and implementation of SODA, a Service-On-Demand Architecture that enables the hosting of app of a service can be dynamically downloaded and started where it is needed (for example, to perform QoS adaptation or to handle increasing load). Examples include the Berkeley Active Service Framework =-=[5]-=-, WebOS [34], Darwin [12] and the Adaptive Service Grid [36]. Different from active services, SODA supports on-demand creation of both services and guest OSes on top of which the services will run. As   </text>
<query_num> 12106 </query_num>
<text>   ederate multiple local HUPs, each having its own SODA Agent and Master. However, we will need to address challenges including autonomous management (under different ownership), distributed monitoring =-=[33]-=-, and platform heterogeneity of multiple HUPs. The purpose of this paper is to demonstrate the architecture of SODA and its prototype implementation, rather than the final design and implementation. T   </text>
<query_num> 12107 </query_num>
<text>   ge and management utility, such as in the experiments.sStorage Resource Broker [9], NeST [10], the Data Grid [14], and OceanStore [26]. Industry efforts parallel those of the academia, such as Oceano =-=[6]-=- at IBM and Planetary Computing [2] at HP. In this paper, we focus on application service hosting as another emerging usage of Grid utility. Recently, the vision of integrating Grid and Web service co   </text>
<query_num> 12108 </query_num>
<text>   n. Resource isolation has been realized by existing techniques of resource reservation, allocation, and scheduling, such as those in QLinux [32], Resource Kernel [30], GARA [21], and Virtual Services =-=[31]-=-. Resources such as CPU, bandwidth, and memory can be reserved and allocated to different processes, users, or service classes; and the allocation is enforced by various resource scheduling algorithms   </text>
<query_num> 12109 </query_num>
<text>   ndling; nor do they support on-demand service creation. The concept of virtual machines has been proposed and realized. Examples include Berkeley NOW [17] and High Performance Virtual Machines (HPVM) =-=[15, 16]-=-. Their goal is to aggregate the computing power of commodity machines (interconnected using high-performance communication techniques) and create a virtual high-performance supercomputer. On the othe   </text>
<query_num> 12110 </query_num>
<text>   rimed with a host OS based on bare hardware; while each virtual service node created by SODA is a virtual machine primed with guest OS and application service, based on the host OS. Finally, Virtuoso =-=[20]-=- realizes Grid computing in virtual machines, paralleling our efforts in service hosting in virtual service nodes. [20] also provides a view on resource management and virtual networking under virtual   </text>
<query_num> 12111 </query_num>
<text>   sophy of utility computing. A number of projects have proposed the usage of Grid resources as computation utility, such as Globus [22], Condor [19], Legion [24], NetSolve [7], Harness [27] and Cactus =-=[4]-=-. Meanwhile, the Grid resources can also be used as a data storage and management utility, such as in the experiments.sStorage Resource Broker [9], NeST [10], the Data Grid [14], and OceanStore [26].   </text>
<query_num> 12112 </query_num>
<text>   us [4]. Meanwhile, the Grid resources can also be used as a data storage and management utility, such as in the experiments.sStorage Resource Broker [9], NeST [10], the Data Grid [14], and OceanStore =-=[26]-=-. Industry efforts parallel those of the academia, such as Oceano [6] at IBM and Planetary Computing [2] at HP. In this paper, we focus on application service hosting as another emerging usage of Grid   </text>
</top>
<top>
<paper_num> 122 </paper_num>
<paper_title>   Toward Picture-perfect Streaming on the Internet.  </paper_title>
<paper_abstract>   Quality of service (QoS) in streaming of continuous media over the Internet is poor, which is partly due to variations in delays, bandwidth limitations, and packet losses. Although continuous media applications can tolerate some missing data, non-recoverable information loss degrades these applications ’ QoS. Consequently, a number of application areas have backed away from streaming of their content over the Internet. Inability to control the resulting visual and auditory quality of the streamed presentation is an important reason for such a trend. We believe that this trend can be reversed. To this end, this paper gives an overview of our efforts in exploring high quality streaming through the exploitation of multiple paths existing in the network. By high quality, we mean with significant bandwidth requirements, of relatively long duration, and without information loss or hiccups in data delivery. We believe this to be a promising approach. 1  </paper_abstract>
<query_num> 12201 </query_num>
<text>   an adaptation scheme which reacts to time-varying path characteristics is our ongoing effort. The work in [27] proposes the use of multi-servers for video streaming on the Internet. Their later work =-=[26]-=- extends the scheme by considering the use of FEC. In their work, they focus on designing a receiver-driven transport protocol which includes (a) a rate allocation algorithm (i.e., how receiver split   </text>
<query_num> 12202 </query_num>
<text>   cation’s sending rate and the resulting effects on the loss characteristics of the CM streaming application. Other works try to achieve multi-path streaming with assistance of the lower layers, e.g., =-=[19]-=- focuses a protocol which utilizes bandwidth available on multiple paths and redistributes workload according to congestion detected by a receiver. However, it requires network layer knowledge as well   </text>
<query_num> 12203 </query_num>
<text>   entralized routing server to compute the multipath routing which minimizes the number of shared links while satisfying the bandwidth requirement. Given information about the underlying network graph, =-=[17]-=- proposes multi-path routing heuristics for unicast and multicast scenarios to achieve high bandwidth and low delay for video traffic. A data scheduling algorithm at the server is also proposed to min   </text>
<query_num> 12204 </query_num>
<text>   hs. The goal here would be to optimize the viewing quality of the resulting CM data and to perform the load distribution based on the quality of the various paths. This problem is explored in [1] and =-=[18]-=- under different path modeling methods. Example results are also presented in Section 2. • Adaptation schemes under changes in network conditions. When network conditions change, one can improve the q f lost packets as measured at the receiver), and it supports the hypothesis that a conventional Gilbert model may not be sufficient for characterizing the loss process of a path. Given this evidence, =-=[18]-=- proposes to use a functional Gilbert model (FGM) as a more general approach to characterizing the bursty loss nature of a path as well as its dependency on an application’s bandwidth requirements. Sp  lost if Xk(t) =1. The transition rate from state 0 to state 1 takes a functional form of F(λ). The transition rate from state 1 to state 0 also takes a functional form of B(λ). It is also assumed in =-=[18]-=- that F(λ) and B(λ) are continuous and furthermore that F(λ) is a non-decreasing function of λ and B(λ) is a non-increasing function of λ. Intuitively these assumptions make sense, and hence, in pract performed using NS2 [22], and the results were qualitatively similar.s2.5 Load Distribution Revisited We now revisit the load distribution problem under the FGM. Consider the following example result =-=[18]-=- with a 120 pkts/sec bandwidth requirement. We first consider a system with two heterogeneous paths wherein F1(b) = 0.4 × b, B1(b) = 21000/b, F2(b) = 0.0667 × b, and B2(b) = 28500/b, as an illustratio is not the best of the three, yet it allows us to reduce the loss rate further. The above results were obtained analytically and without the use of erasure codes. We now consider simulation results 9 =-=[18]-=- where an error erasure code is used to reconstruct lost packets. Here the bandwidth requirements of the streaming application are increased by 12.5% due to the added overhead of the erasure code. Not   </text>
<query_num> 12205 </query_num>
<text>   in the benefits of multi-path streaming described above, one must first determine the paths to be used in delivery of the data. Since it is reasonable to characterize a path using its bottleneck link =-=[12]-=-, what we need to be able to do is determine whether a number of paths share points of congestion, i.e., have joint or disjoint bottlenecks [21, 29]. In addition, it may be useful (for some approaches   </text>
<query_num> 12206 </query_num>
<text>   king network bandwidth, channel characteristics and FEC parameters into account) and (b) a packet partitioning algorithm (which ensures non-overlapping packet sending and minimizes startup delay). In =-=[25]-=-, they examine the case where the last mile connection is the bottleneck and employ multi-path streaming otherwise. In these proposals, conventional GM is adopted while a more expressive model is used   </text>
<query_num> 12207 </query_num>
<text>   n. Their later work [28] extends the idea of using FEC with path diversity on an overlay framework. The motivation is to emulate multiple sources by the use of an overlay network, which is similar to =-=[3]-=- but using multiple redundant paths simultaneously. They propose a heuristic scheme to select redundant paths. It is suggested that 10% of the network nodes participating in the overlay is enough for   </text>
<query_num> 12208 </query_num>
<text>   on packet loss characteristics (e.g., loss correlations) rather than bandwidth and delay. Sophisticated media coding techniques (e.g., MD coding) have been developed for use with multi-path streaming =-=[4, 8, 6, 9, 5, 24, 30, 13, 14, 16, 15]-=-. The basic idea is to partition the media into multiple roughly equally important independently decodeable bitstreams (descriptions). Each of them contains complementary information and is sent throu   </text>
<query_num> 12209 </query_num>
<text>   on packet loss characteristics (e.g., loss correlations) rather than bandwidth and delay. Sophisticated media coding techniques (e.g., MD coding) have been developed for use with multi-path streaming =-=[4, 8, 6, 9, 5, 24, 30, 13, 14, 16, 15]-=-. The basic idea is to partition the media into multiple roughly equally important independently decodeable bitstreams (descriptions). Each of them contains complementary information and is sent throu media quality can be improved as the number of received descriptions increases. In this context, issues such as dealing with heterogeneous path bandwidth constraints [8], rate-distortion optimization =-=[24, 13, 14]-=-, coding efficiency, adaptation schemes [30, 16], etc. are explored. Specifically, in [6, 9], a model based on the conventional GM is also used in the analysis and path selection for MD coded multi-pa   </text>
<query_num> 12210 </query_num>
<text>   on packet loss characteristics (e.g., loss correlations) rather than bandwidth and delay. Sophisticated media coding techniques (e.g., MD coding) have been developed for use with multi-path streaming =-=[4, 8, 6, 9, 5, 24, 30, 13, 14, 16, 15]-=-. The basic idea is to partition the media into multiple roughly equally important independently decodeable bitstreams (descriptions). Each of them contains complementary information and is sent throu ssues such as dealing with heterogeneous path bandwidth constraints [8], rate-distortion optimization [24, 13, 14], coding efficiency, adaptation schemes [30, 16], etc. are explored. Specifically, in =-=[6, 9]-=-, a model based on the conventional GM is also used in the analysis and path selection for MD coded multi-path streaming. In our work, we focus on a more general study of multi-path streaming without   </text>
<query_num> 12211 </query_num>
<text>   on packet loss characteristics (e.g., loss correlations) rather than bandwidth and delay. Sophisticated media coding techniques (e.g., MD coding) have been developed for use with multi-path streaming =-=[4, 8, 6, 9, 5, 24, 30, 13, 14, 16, 15]-=-. The basic idea is to partition the media into multiple roughly equally important independently decodeable bitstreams (descriptions). Each of them contains complementary information and is sent throu ved descriptions increases. In this context, issues such as dealing with heterogeneous path bandwidth constraints [8], rate-distortion optimization [24, 13, 14], coding efficiency, adaptation schemes =-=[30, 16]-=-, etc. are explored. Specifically, in [6, 9], a model based on the conventional GM is also used in the analysis and path selection for MD coded multi-path streaming. In our work, we focus on a more ge   </text>
<query_num> 12212 </query_num>
<text>   onable to characterize a path using its bottleneck link [12], what we need to be able to do is determine whether a number of paths share points of congestion, i.e., have joint or disjoint bottlenecks =-=[21, 29]-=-. In addition, it may be useful (for some approaches to multi-path streaming) to be able to estimate current capacities and loss characteristics of these bottlenecks. Although this has not been critic re fairly accurate estimation of various network characteristics (refer to Section 3). These are non-trivial problems, but they are outside the scope of this discussion. We note that currently we use =-=[29]-=- in our system for detecting shared points of congestion. • Effects of redundancy and error erasure schemes. Some amount of lost data can be reconstructed in CM applications through the use of redunda eads and complexity due to measurements needed to achieve better performance with MP streaming should also be considered. For instance, in our case, we employ detection of shared points of congestion =-=[29]-=-. Other approaches to MP streaming might require even more detailed information about the network (refer to Section 3) which is likely to result in a need for more “intrusive” and complex measurements  network-level routing algorithm; furthermore, our system does not require specific knowledge of the paths, only the ability to determine whether two paths share a point of congestion, e.g., by using =-=[29]-=-.sering server-related problems such as the load balancing issues mentioned above) 2 . 2 Overview of Results We now give a brief overview of our results thus far, to illustrate the potential of multi- the network and hence allows easy deployment on the Internet. Our only requirement is that chosen paths do not share points of congestion, which can be detected at the end-hosts using schemes such as =-=[29]-=-. Also, we focus on packet loss characteristics (e.g., loss correlations) rather than bandwidth and delay. Sophisticated media coding techniques (e.g., MD coding) have been developed for use with mult   </text>
<query_num> 12213 </query_num>
<text>   ved descriptions increases. In this context, issues such as dealing with heterogeneous path bandwidth constraints [8], rate-distortion optimization [24, 13, 14], coding efficiency, adaptation schemes =-=[30, 16]-=-, etc. are explored. Specifically, in [6, 9], a model based on the conventional GM is also used in the analysis and path selection for MD coded multi-path streaming. In our work, we focus on a more ge   </text>
<query_num> 12214 </query_num>
<text>   well as be made to achieving multiple paths by using relay nodes. In fact, we have a P2P prototype implementation of our approach [2]. Best-path type approaches have also been studied. For instance, =-=[31, 32]-=- perform path switching to select the best path by estimating the “goodness” of a path from the perspective of a video and a VoIP stream. However, they do not exploit the benefits of path diversity (e   </text>
</top>
<top>
<paper_num> 123 </paper_num>
<paper_title>   Partitioning-Based Approach to Fast On-Chip Decoupling Capacitor Budgeting and Minimization.  </paper_title>
<paper_abstract>   Abstract — This paper proposes a fast decoupling capacitance (decap) allocation and budgeting algorithm for both early stage decap estimation and later stage decap minimization in today’s VLSI physical design. The new method is based on a sensitivitybased conjugate gradient (CG) approach. But we adopt several new techniques, which significantly improve the efficiency of the optimization process. First, we propose an efficient search step scheme to replace the time-consuming line search phase in conventional conjugate gradient method for decap budget optimization. Second, instead of optimizing an entire large circuit, we partition the circuit into a number of smaller subcircuits and optimize them separately by exploiting the locality of adding decaps. Third, we apply time-domain merged adjoint method to compute the sensitivity information and show that the partitioning-based merged adjoint method leads to better results than the flat merged adjoint method with the improved search scheme. Experimental results show that the proposed algorithm achieves at least 10X speed-up over the similar decap allocation methods reported so far with similar budget quality, and a power grid circuit with about one million nodes can be optimized using the new method in half an hour on the latest Linux workstations.  </paper_abstract>
<query_num> 12301 </query_num>
<text>   cy of the algorithm is therefore more obvious. Since the efficiency of our proposed method depends on the linear solver used, any speedup techniques like hierarchical approach [22], multigrid methods =-=[10]-=-, iterative methods [3], model reduction methods [5], [20] and random walk based algorithm [15] can be used to speed up and increase the capacity of the proposed method. Fig. 6. Comparison of decap bu   </text>
<query_num> 12302 </query_num>
<text>   d decap candidate locations through the entire chip for each violation node. A. General Partition Algorithm In our paper, we use the graph-based multilevel minimum cut algorithm for partitioning task =-=[9]-=-, which provides an extremely fast speed on large graph sizes, where a graph with one million vertices can be processed roughly within one minute. This ensures that the partition phase will not bottle   </text>
<query_num> 12303 </query_num>
<text>   fficiency of our proposed method depends on the linear solver used, any speedup techniques like hierarchical approach [22], multigrid methods [10], iterative methods [3], model reduction methods [5], =-=[20]-=- and random walk based algorithm [15] can be used to speed up and increase the capacity of the proposed method. Fig. 6. Comparison of decap budget and CPU time between different partition sizes. . We  des can be optimized in about half an hour on the latest Linux workstation. In the future, more efficient circuit simulation techniques, like hierarchical approach [22] and model reduction approaches =-=[20]-=- will be used to improve the transient simulation of the power/ground networks. Also, parallel simulation will be explored to further improve the efficiency of the decap budgeting algorithm. VII. ACKN   </text>
<query_num> 12304 </query_num>
<text>   herefore more obvious. Since the efficiency of our proposed method depends on the linear solver used, any speedup techniques like hierarchical approach [22], multigrid methods [10], iterative methods =-=[3]-=-, model reduction methods [5], [20] and random walk based algorithm [15] can be used to speed up and increase the capacity of the proposed method. Fig. 6. Comparison of decap budget and CPU time betwe   </text>
<query_num> 12305 </query_num>
<text>   illions of nodes and extracted onchip and off-chip RLC components in modern VLSI design. Mathematically, optimal decap allocation is a nonlinear optimization problem and many existing approaches [7], =-=[18]-=- use sensitivity-based optimization methods to solve the problem. To compute the sensitivity, transient simulations of the whole P/G networks have to be carried out at every optimization step. Given t er of sub-circuits and optimize them individually. A noise-aware partition scheme is proposed to perform the required partitioning. (3) We applies the timedomain merged adjoint network method in [7], =-=[18]-=- for sensitivity calculation in the partitioning-based optimization framework. We show that combining the proposed partitioning scheme with the merged adjoint method leads to much faster optimization  can be optimized more efficiently by wire sizing. We notice that the time-domain merged adjoint method has been used for transistor sizing [6] and then later applied to the decap optimization in [7], =-=[18]-=-. But we show in this paper that with the proposed partitioning scheme, the merged adjoint method works better than the simple application of the merged adjoint method in the CG optimization framework  is sensitivity-based approaches based on actual circuit simulation, which is a more accurate and wellaccepted method. Fig. 2 gives an illustration of VDD fluctuation of a node within one clock cycle =-=[18]-=-. The violation area at node j is defined as: gj(c1, ..., cn) = � T 0 max(Vmin − vj(t), 0)dt (1) which equals to the shaded area below a certain VDD threshold in the graph. The sensitivity of decap ad adjoint network under unit step current excitation at violation node j.sIEEE TRANSACTIONS ON COMPUTER-AIDED DESIGN OF INTEGRATED CIRCUITS AND SYSTEMS, VOL XX, NO. XX, DECEMBER 200X 3 In previous work =-=[18]-=-, a sensitivity-based algorithm is developed to optimize the allocation of decaps in a standard cell design format. In its problem formulation, it minimizes the following objective function m� gj(c1,   </text>
<query_num> 12306 </query_num>
<text>   ith millions of nodes and extracted onchip and off-chip RLC components in modern VLSI design. Mathematically, optimal decap allocation is a nonlinear optimization problem and many existing approaches =-=[7]-=-, [18] use sensitivity-based optimization methods to solve the problem. To compute the sensitivity, transient simulations of the whole P/G networks have to be carried out at every optimization step. G llions of nodes is already an extremely time-consuming task, the CPU and memory cost of the optimization method that uses transient simulations as the internal loops will be prohibitive. Recent study =-=[7]-=- shows that allocating decaps for a P/G grid with about one million nodes will take about 10 hours in modern workstations with improved simulation techniques. Given the increasing sizes of P/G network  number of sub-circuits and optimize them individually. A noise-aware partition scheme is proposed to perform the required partitioning. (3) We applies the timedomain merged adjoint network method in =-=[7]-=-, [18] for sensitivity calculation in the partitioning-based optimization framework. We show that combining the proposed partitioning scheme with the merged adjoint method leads to much faster optimiz  and can be optimized more efficiently by wire sizing. We notice that the time-domain merged adjoint method has been used for transistor sizing [6] and then later applied to the decap optimization in =-=[7]-=-, [18]. But we show in this paper that with the proposed partitioning scheme, the merged adjoint method works better than the simple application of the merged adjoint method in the CG optimization fra ty for solving sparse n × n matrices [19]. It is clear that the speed of this algorithm depends on the number of violation nodes, and will become intolerable for extremely large circuits. Recent work =-=[7]-=- improved the sensitivity calculation by applying the time domain merged adjoint network method, which will be covered later in the next section. The decap area is incorporated into the objective func   </text>
<query_num> 12307 </query_num>
<text>   ivery. For removing dynamic IR drop which Some preliminary results of this paper appeared in Proc. Design Automation Conference (DAC’05), Proc. Int. Symposium. on Quality Electronic Design (ISQED’05) =-=[11]-=-, [14] Hang Li, Jeffrey Fan, and Sheldon X.-D. Tan are with Department of Electrical Engineering, University of California, Riverside, Riverside, CA 92521 USA (e-mail: {hli,jfan,stan}@ee.ucr.edu) Life   </text>
<query_num> 12308 </query_num>
<text>   nds on the linear solver used, any speedup techniques like hierarchical approach [22], multigrid methods [10], iterative methods [3], model reduction methods [5], [20] and random walk based algorithm =-=[15]-=- can be used to speed up and increase the capacity of the proposed method. Fig. 6. Comparison of decap budget and CPU time between different partition sizes. . We also notice that the difference among   </text>
<query_num> 12309 </query_num>
<text>   o verify our method, with conclusions in section VI. II. REVIEW OF PREVIOUS DECAP OPTIMIZATION ALGORITHMS Existing on-chip decap budgeting algorithms basically fall into two categories. In [2], [12], =-=[16]-=-, [17], the current pattern around hot spots (where violation of IR drop occurs) is derived and the amount of electric charge needed to supply that current demand is estimated. To obtain an optimal de   </text>
<query_num> 12310 </query_num>
<text>   on V to verify our method, with conclusions in section VI. II. REVIEW OF PREVIOUS DECAP OPTIMIZATION ALGORITHMS Existing on-chip decap budgeting algorithms basically fall into two categories. In [2], =-=[12]-=-, [16], [17], the current pattern around hot spots (where violation of IR drop occurs) is derived and the amount of electric charge needed to supply that current demand is estimated. To obtain an opti   </text>
<query_num> 12311 </query_num>
<text>   section V to verify our method, with conclusions in section VI. II. REVIEW OF PREVIOUS DECAP OPTIMIZATION ALGORITHMS Existing on-chip decap budgeting algorithms basically fall into two categories. In =-=[2]-=-, [12], [16], [17], the current pattern around hot spots (where violation of IR drop occurs) is derived and the amount of electric charge needed to supply that current demand is estimated. To obtain a   </text>
<query_num> 12312 </query_num>
<text>   size. The time efficiency of the algorithm is therefore more obvious. Since the efficiency of our proposed method depends on the linear solver used, any speedup techniques like hierarchical approach =-=[22]-=-, multigrid methods [10], iterative methods [3], model reduction methods [5], [20] and random walk based algorithm [15] can be used to speed up and increase the capacity of the proposed method. Fig. 6 id circuit with about one million nodes can be optimized in about half an hour on the latest Linux workstation. In the future, more efficient circuit simulation techniques, like hierarchical approach =-=[22]-=- and model reduction approaches [20] will be used to improve the transient simulation of the power/ground networks. Also, parallel simulation will be explored to further improve the efficiency of the   </text>
<query_num> 12313 </query_num>
<text>   sizing step, as leakage currents are mainly DC currents and can be optimized more efficiently by wire sizing. We notice that the time-domain merged adjoint method has been used for transistor sizing =-=[6]-=- and then later applied to the decap optimization in [7], [18]. But we show in this paper that with the proposed partitioning scheme, the merged adjoint method works better than the simple application   </text>
</top>
<top>
<paper_num> 124 </paper_num>
<paper_title>   Featherweight generic confinement.  </paper_title>
<paper_abstract>   Existing approaches to object encapsulation and confinement either rely on restrictions to programs or require the use of specialised ownership type systems. Syntactic restrictions are difficult to scale and to prove correct, while specialised type systems require extensive changes to programming languages. We demonstrate that confinement can be enforced cheaply in Featherweight Generic Java, with no essential change to the underlying language or type system. This result delineates the differences between parametric polymorphism and ownership type systems, demonstrates that polymorphic type parameters can simultaneously act as ownership parameters, and should facilitate the adoption of ownership and confinement type systems in general-purpose programming languages. 1.  </paper_abstract>
<query_num> 12401 </query_num>
<text>   ODUCTION Two main approaches to object instance encapsulation are under investigation in the literature. On one hand, programming conventions, such as Islands [13] and various kinds of Confined Types =-=[4, 10]-=- use tailored restrictions on programs to provide containment guarantees for programs in existing programming languages, but until recently had not been proven sound [22]. On the other hand, ownership berg, and Vitek [12] show how type inference can be used to avoid the need for annotation, making a system that can provide per-package encapsulation in practical programs. Clarke, Richmond and Noble =-=[10]-=- apply these ideas in the context of Enterprise Java Beans, and by exploiting special architecture specific constraints, provide per-object encapsulation without annotations or inference. Recent work   </text>
<query_num> 12402 </query_num>
<text>   ODUCTION Two main approaches to object instance encapsulation are under investigation in the literature. On one hand, programming conventions, such as Islands [13] and various kinds of Confined Types =-=[4, 10]-=- use tailored restrictions on programs to provide containment guarantees for programs in existing programming languages, but until recently had not been proven sound [22]. On the other hand, ownership ms by general-purpose programming languages. The next section of this paper briefly introduces the notion of encapsulation or confinement, in particular the kind of confinement used in Confined Types =-=[4]-=-, the primary topic of this paper. We then present FGJ+c, which leverages the generic type rules of FGJ to support a simple confinement invariant, ensuring that confined classes may not be accessed ou  method names. These schemes enforce a containment invariant which simply states that objects outside a particular boundary may not access objects inside that boundary. For example, in Confined Types =-=[4]-=-, the unit of confinement is a Java package: all the instances of public classes within that package form the encapsulation boundary; all the instances of private classes (known as confined classes) a ntary threads of research have evolved. On one hand are expressive but weighty type systems based on ownership types [9]. On the other hand are lightweight but limited systems based on confined types =-=[4]-=-. The systems based on ownership types differ essentially in only one characteristic, which Clarke and Wrigstad distinguish as shallow vs. deep ownership [11]. A deep ownership type permits only a sin defaults, this need not be a problem in practice [1, 5]. Confined type systems have achieved their more limited goals while keeping the amount of annotations low. Vitek and Bokowski’s original system =-=[4]-=-, which had security as its application, required certain classes to be annotated as confined to indicate classes confined within the present package, and certain methods to be annotated as anonymous,   </text>
<query_num> 12403 </query_num>
<text>   concepts. We do not need to distinguish anonymous methods, because “this” is parameterised to record its ownership. Banerjee and Naumann prove a per-object representation independence result for Java =-=[2, 3]-=-. They adopt a confinement discipline resembling ownership types, except that they apply the confinement only at the point they wish to reason about. They require that confined classes extend a specia   </text>
<query_num> 12404 </query_num>
<text>   lus, such as that underpinning Joe1 [8], although that calculus will first have to be extended with FGJ-like genericity. 6.2 Capabilities We also plan to extend FGJ+c to model capability-like systems =-=[6, 20]-=- where types control invocations of individual methods, rather than access to whole objects. Again, we have to alter the method invocation rule: ∆; Γ ⊢ visible(e, D) ∆; Γ ⊢ visible(ē, D) ∆; Γ ⊢ e : N   </text>
<query_num> 12405 </query_num>
<text>   ms in existing programming languages, but until recently had not been proven sound [22]. On the other hand, ownership type systems [9], originating from the formalisation of Flexible Alias Protection =-=[19]-=-, require quite significant modifications to programming languages. In particular, languages like Joe, Universes, AliasJava, and SafeConcurrentJava, depend upon ownership parameterisation within the t   </text>
<query_num> 12406 </query_num>
<text>   opyright is held by the author/owner. ACM 0-89791-88-6/97/05. 1 Institute of Information and Computing Sciences Utrecht University, Netherlands dave@cs.uu.nl confinement in Featherweight Generic Java =-=[14]-=- “almost for free” — with no change to the underlying language or type system — by additionally enforcing some simple visibility rules and constraints on program structure. We hope that this result wi Second, we hope to obtain a simpler formalism, with few new concepts. Third, we are developing an extension to Generic Java that will merge ownership and generic types: our Featherweight Generic Java =-=[14]-=- model will support soundness proofs for our language. Finally, we hope this approach will facilitate the adoption of ownership and confinement type systems by general-purpose programming languages. T  Figure 1: FGJ Classes separated into FGJ+c classes and owner classes. extra restrictions that leverage FGJ’s proven type soundness to provide confinement. Every FGJ+c program must meet the FGJ rules =-=[14]-=- along with additional rules presented in figure 2. For reference, figure 3 shows the FGJ syntax from Igarashi et al. [14]. 4.1 FGJ+c Programs Any FGJ+c program is an FGJ program that meets the follow   </text>
<query_num> 12407 </query_num>
<text>   s confined to indicate classes confined within the present package, and certain methods to be annotated as anonymous, to indicate that such methods do not reveal “this”. Grothoff, Palsberg, and Vitek =-=[12]-=- show how type inference can be used to avoid the need for annotation, making a system that can provide per-package encapsulation in practical programs. Clarke, Richmond and Noble [10] apply these ide   </text>
<query_num> 12408 </query_num>
<text>   uite significant modifications to programming languages. In particular, languages like Joe, Universes, AliasJava, and SafeConcurrentJava, depend upon ownership parameterisation within the type system =-=[8, 16, 1, 5]-=-. All these type systems are distinct, but they only support ownership parameterisation, not type parameters. This paper continues the efforts to provide effective object encapsulation within practica Ownership types require additional annotations to use them, raising issues about their role in programming. Some authors argue that, with appropriate defaults, this need not be a problem in practice =-=[1, 5]-=-. Confined type systems have achieved their more limited goals while keeping the amount of annotations low. Vitek and Bokowski’s original system [4], which had security as its application, required ce e current object (i.e., owned by “this”). Then, we need to ensure that we can only access objects owned by This from the instance to which they belong — that is, only when derefencing this. Following =-=[1]-=-, we sketch such an ownership rule: ∆; Γ ⊢ e : T This ∈ owners(mtype(m, bound∆(T))) ⇒ e ≡ this ∆; Γ ⊢ ownership(e.m(ē), D) Where owners finds the set of all ownership parameters or bounds in a type. T   </text>
<query_num> 12409 </query_num>
<text>   us kinds of Confined Types [4, 10] use tailored restrictions on programs to provide containment guarantees for programs in existing programming languages, but until recently had not been proven sound =-=[22]-=-. On the other hand, ownership type systems [9], originating from the formalisation of Flexible Alias Protection [19], require quite significant modifications to programming languages. In particular,  irect accesses: indirect access is permitted — indeed, is encouraged. Public classes (or instances of public classes) thus provide an interface to the private instances in their package. Zhao et. al. =-=[22]-=- have formulated a containment invariant in terms of the expressions within methods. Basically, if an expression (or any of its subexpressions) can possibly evaluate to some object o, that object must ntext of Enterprise Java Beans, and by exploiting special architecture specific constraints, provide per-object encapsulation without annotations or inference. Recent work by Zhao, Palsberg and Vitek =-=[22]-=- has formalised Vitek and Bokowski’s approach to per-package confinement, with an operational semantics and a static type system based on Featherweight Java, and augmented by a number of specific rule   </text>
<query_num> 12410 </query_num>
<text>   ut limited systems based on confined types [4]. The systems based on ownership types differ essentially in only one characteristic, which Clarke and Wrigstad distinguish as shallow vs. deep ownership =-=[11]-=-. A deep ownership type permits only a single object as entry point to the collection of objects it owns, whereas a shallow ownership type permits multiple entry points into the confined collection. C   </text>
</top>
<top>
<paper_num> 125 </paper_num>
<paper_title>   Convicting exploitable software vulnerabilities: An efficient input provenance based approach.  </paper_title>
<paper_abstract>   Software vulnerabilities are the root cause of a wide range of attacks. Existing vulnerability scanning tools are able to produce a set of suspects. However, they often suffer from a high false positive rate. Convicting a suspect and vindicating false positives are mostly a highly demanding manual process, requiring a certain level of understanding of the software. This limitation significantly thwarts the application of these tools by system administrators or regular users who are concerned about security but lack of understanding of, or even access to, the source code. It is often the case that even developers are reluctant to inspect/fix these numerous suspects unless they are convicted by evidence. In this paper, we propose a lightweight dynamic approach which generates evidence for various security vulnerabilities in software, with the goal of relieving the manual procedure. It is based on data lineage tracing, a technique that associates each execution point precisely with a set of relevant input values. These input values can be mutated by an offline analysis to generate exploits. We overcome the efficiency challenge by using Binary Decision Diagrams (BDD). Our tool successfully generates exploits for all the known vulnerabilities we studied. We also use it to uncover a number of new vulnerabilities, proved by evidence.  </paper_abstract>
<query_num> 12501 </query_num>
<text>   Xiangyu Zhang Dongyan Xu Department of Computer Sciences and CERIAS Purdue University {zlin, xyzhang, dxu}@cs.purdue.edu such as TaintCheck [15], Control Flow Integrity [16], and Data Flow Integrity =-=[17]-=-. However, most of these techniques are often active during program execution, thereby incurring non-trivial runtime overhead. Moreover, they aim to detect attacks, and thus vulnerabilities that are n detect attacks on the fly. Reducing runtime overhead is its major concern. This is also true for other dynamic techniques such as control flow integrity checking [16] and data flow integrity checking =-=[17]-=-. Our technique aims to generate evidence off-line. 7 Conclusions In this paper, we propose a data lineage tracing based dynamic approach to generate evidence for remote exploitable vulnerabilities in   </text>
<query_num> 12502 </query_num>
<text>   ce. Therefore, it becomes a pressing need to develop new techniques to automatically or semiautomatically generate evidence to convict real vulnerabilities. Random test generation (e.g., fuzz testing =-=[7, 8]-=-) that randomly mutates benign inputs has been used to construct exploits. However, it is known that random test generation is not effective in many cases, e.g. it might take 2 32 tries to satisfy a s /DL(si) (use0) (use1) 2311 fread(...) pc1 ∗ 1 READ (buf,size,...) ∀0 ≤ i &amp;lt; size buf[i] ∀0 ≤ i &amp;lt; size DL(buf[i]@pc11) = get new id() ∗∗∗ pc2 ∗∗ 7 pc28 *p = buf[i] *p = buf[i] wide lo wide hi buf[6] buf=-=[7]-=- pc11 pc11 DL(∗p@pc27) = DL(buf[6]@pc11) = {6} DL(∗p@pc28) = DL(buf[7]@pc11) = {7} pc29 *p = buf[i] high lo buf[8] pc11 DL(∗p@pc29) = DL(buf[8]@pc11) = {8} pc210 *p = buf[i] high hi buf[9] pc11 DL(∗p@   </text>
<query_num> 12503 </query_num>
<text>   ce. Therefore, it becomes a pressing need to develop new techniques to automatically or semiautomatically generate evidence to convict real vulnerabilities. Random test generation (e.g., fuzz testing =-=[7, 8]-=-) that randomly mutates benign inputs has been used to construct exploits. However, it is known that random test generation is not effective in many cases, e.g. it might take 2 32 tries to satisfy a s 1) = get new id() ∗∗∗ pc2 ∗∗ 7 pc28 *p = buf[i] *p = buf[i] wide lo wide hi buf[6] buf[7] pc11 pc11 DL(∗p@pc27) = DL(buf[6]@pc11) = {6} DL(∗p@pc28) = DL(buf[7]@pc11) = {7} pc29 *p = buf[i] high lo buf=-=[8]-=- pc11 DL(∗p@pc29) = DL(buf[8]@pc11) = {8} pc210 *p = buf[i] high hi buf[9] pc11 DL(∗p@pc210) = DL(buf[9]@pc11) = {9} 2451 width=... width wide hi pc28 wide lo pc27 DL(width@2451) = DL(wide hi@pc28) ∪D   </text>
<query_num> 12504 </query_num>
<text>   def@si)/DL(si) (use0) (use1) 2311 fread(...) pc1 ∗ 1 READ (buf,size,...) ∀0 ≤ i &amp;lt; size buf[i] ∀0 ≤ i &amp;lt; size DL(buf[i]@pc11) = get new id() ∗∗∗ pc2 ∗∗ 7 pc28 *p = buf[i] *p = buf[i] wide lo wide hi buf=-=[6]-=- buf[7] pc11 pc11 DL(∗p@pc27) = DL(buf[6]@pc11) = {6} DL(∗p@pc28) = DL(buf[7]@pc11) = {7} pc29 *p = buf[i] high lo buf[8] pc11 DL(∗p@pc29) = DL(buf[8]@pc11) = {8} pc210 *p = buf[i] high hi buf[9] pc11  width 2451 height 2461 DL(4941) = DL(width@2451) ∪ DL(height@2461) = {6, 7, 8, 9} * pc1, pc2 are statements in libc functions. ** the input byte with offset 7, with the value “0x00”, is loaded to buf=-=[6]-=- by the 6th instance of pc2 *** since the input sequence starts with buf[0] and the id assignment starts at 0, DL(buf[i]@pc11) ≡ i. As roBDD is capable of efficiently representing the power set domain bilities with evidence pays off the loss of completeness. 5 Implementation and Evaluation We have implemented the whole system in Linux. The data lineage tracer module is built on top of Valgrind-2.2 =-=[6]-=- with roBDD [4] support. We instrument data movement (e.g.,LOAD,STORE,MOV), arithmetic operation, and logic operation instructions (e.g., ADD,SUB,AND) to keep track of data dependence and generate lin   </text>
<query_num> 12505 </query_num>
<text>   hus, recently, there has been significant advance in combining static software verification principles with symbolic execution in test generation to identify software errors including vulnerabilitiess=-=[9, 11, 10, 12, 13, 14]-=-. These techniques aim to explore all feasible program paths to expose potential defects. Such an ambitious goal with symbolic execution incurs scalability issues. For instance, using symbolic executi   </text>
<query_num> 12506 </query_num>
<text>   hus, recently, there has been significant advance in combining static software verification principles with symbolic execution in test generation to identify software errors including vulnerabilitiess=-=[9, 11, 10, 12, 13, 14]-=-. These techniques aim to explore all feasible program paths to expose potential defects. Such an ambitious goal with symbolic execution incurs scalability issues. For instance, using symbolic executi [13] is capable of handling hundreds of millions of instructions, which only accounts for a few seconds of execution. Furthermore, it often requires the user to annotate symbolic variables (e.g., EXE =-=[9]-=-), which implies understanding of program semantics. In this paper, we propose a practical dynamic approach that is intended to use in combination with other static tools. We observe that although the ghlighted as follows. • We propose a novel dynamic technique which generates evidence to convict a wide range of real vulnerabilities. Compared with the state of the art of test generation techniques =-=[9, 10, 11]-=-, it is less expensive. The output of our tool is a runnable program input to the whole software system instead of a module, and such an input can be easily turned into an exploit. • The technique is  usex@DEF (usex)). Otherwise, it is treated as having an empty lineage set, corresponding to statically initialized variables. Identifying Input Values. It is non-trivial to label input values. In EXE =-=[9]-=-, users are required to annotate input variables. We had considered such a strategy. However, since we are working at binary level and we handle whole system inputs such as those read from files or ne i buf[6] buf[7] pc11 pc11 DL(∗p@pc27) = DL(buf[6]@pc11) = {6} DL(∗p@pc28) = DL(buf[7]@pc11) = {7} pc29 *p = buf[i] high lo buf[8] pc11 DL(∗p@pc29) = DL(buf[8]@pc11) = {8} pc210 *p = buf[i] high hi buf=-=[9]-=- pc11 DL(∗p@pc210) = DL(buf[9]@pc11) = {9} 2451 width=... width wide hi pc28 wide lo pc27 DL(width@2451) = DL(wide hi@pc28) ∪DL(wide lo@pc27) = {6, 7} 2461 height=... height high hi pc210 high lo pc29   </text>
<query_num> 12507 </query_num>
<text>   hus, recently, there has been significant advance in combining static software verification principles with symbolic execution in test generation to identify software errors including vulnerabilitiess=-=[9, 11, 10, 12, 13, 14]-=-. These techniques aim to explore all feasible program paths to expose potential defects. Such an ambitious goal with symbolic execution incurs scalability issues. For instance, using symbolic executi ghlighted as follows. • We propose a novel dynamic technique which generates evidence to convict a wide range of real vulnerabilities. Compared with the state of the art of test generation techniques =-=[9, 10, 11]-=-, it is less expensive. The output of our tool is a runnable program input to the whole software system instead of a module, and such an input can be easily turned into an exploit. • The technique is  t much more space than our roBDD based approach, especially for dataintensive applications. 6 Related Work In recent years, there have been significant advance in automated code based test generation =-=[9, 10, 11, 13]-=-. Theoretically, these techniques can be applied to our problem of automated evidence generation. However in practice, they have inherent limitations that constrain their application. First, most thes   </text>
<query_num> 12508 </query_num>
<text>   hus, recently, there has been significant advance in combining static software verification principles with symbolic execution in test generation to identify software errors including vulnerabilitiess=-=[9, 11, 10, 12, 13, 14]-=-. These techniques aim to explore all feasible program paths to expose potential defects. Such an ambitious goal with symbolic execution incurs scalability issues. For instance, using symbolic executi ints and resolve them by a solver, e.g., solving the negated constraint ¬C1 provides a new input value satisfying x!=c, which drives the execution to take the false branch of P1. The state of the art =-=[13]-=- is capable of handling hundreds of millions of instructions, which only accounts for a few seconds of execution. Furthermore, it often requires the user to annotate symbolic variables (e.g., EXE [9]) t much more space than our roBDD based approach, especially for dataintensive applications. 6 Related Work In recent years, there have been significant advance in automated code based test generation =-=[9, 10, 11, 13]-=-. Theoretically, these techniques can be applied to our problem of automated evidence generation. However in practice, they have inherent limitations that constrain their application. First, most thes   </text>
<query_num> 12509 </query_num>
<text>   overhead. Moreover, they aim to detect attacks, and thus vulnerabilities that are not under attack are invisible. The second type of approaches are static analysis, and notable examples include BOON =-=[18]-=-, Splint [19], and Archer [21]. Static analysis is not bound to execution and thus often capable of identifying potential vulnerabilities in a program, and also it imposes no overhead at runtime. Thus   </text>
<query_num> 12510 </query_num>
<text>   pproaches have been proposed in this category, Zhiqiang Lin Xiangyu Zhang Dongyan Xu Department of Computer Sciences and CERIAS Purdue University {zlin, xyzhang, dxu}@cs.purdue.edu such as TaintCheck =-=[15]-=-, Control Flow Integrity [16], and Data Flow Integrity [17]. However, most of these techniques are often active during program execution, thereby incurring non-trivial runtime overhead. Moreover, they spect is convicted. Otherwise, the suspect is considered innocent. Since the runtime detector is not our focus, we simply use a segmentation fault detector. More advanced detectors such as TaintCheck =-=[15]-=- can be adopted for higher accuracy. Our technique does not rely on a specific static analysis tool, which provides flexibility to the system. More specifically, it can be easily shaped into a system  o. Eventually, at 4941, we acquire the exact lineage as demonstrated earlier in Figure 2. Efficient Lineage Representation. Compared with existing techniques with similar functions such as TaintCheck =-=[15]-=-, in which one bit is required for one byte, we are facing a much harder space problem because we are computing a set for each byte, which potentially has the same cardinality of the entire input set. ue is a light-weight whole program technique that explore a subset of program paths. It does not require source code access and it does not require understanding the program in most cases. TaintCheck =-=[15]-=- represents another type of dynamic techniques that are relevant. Our technique can be considered as a generalization of TaintCheck. More specifically, TaintCheck uses one bit to color program executi   </text>
<query_num> 12511 </query_num>
<text>   reover, they aim to detect attacks, and thus vulnerabilities that are not under attack are invisible. The second type of approaches are static analysis, and notable examples include BOON [18], Splint =-=[19]-=-, and Archer [21]. Static analysis is not bound to execution and thus often capable of identifying potential vulnerabilities in a program, and also it imposes no overhead at runtime. Thus, these techn   </text>
<query_num> 12512 </query_num>
<text>   romise. Unfortunately, most static techniques suffer from a high falsepositive rate and generate a large volume of warnings. For example, the static analysis tool Splint has nearly 50% false positive =-=[20]-=-, and tools like Flawfinder [1] and RATS [2] often produce hundreds of warnings, in which only a few of them are the real defects. The procedure of convicting real defects and vindicating false positi   </text>
<query_num> 12513 </query_num>
<text>   sed implementation may be devastating. For example, sets with thousands of elements may have to be traversed for the execution of a single instruction. Fortunately, recent research on dynamic slicing =-=[24]-=- reveals that reduced ordered Binary Decision Diagram (roBDD) [4] can be used to achieve both space and time efficiency in representing sets, especially when these sets have the characteristics of ove   </text>
<query_num> 12514 </query_num>
<text>   stance si. For example, DL(width@4941) = {6, 7}, with the numbers denoting the values’ indices in the input sequence 2 . DL(4941) = {6, 7, 8, 9}. One may raise the question whether control dependence =-=[22]-=- needs to be considered. Our experience shows that simply including control dependence in lineage computation often leads to undesirably oversized lineage sets. Therefore we consider control dependenc T covers the suspect s. SCD contains the static control dependence information, which is precomputed from the binary. Readers who are interested in computing static control dependence are referred to =-=[22]-=-. The implementation of our SCD component is discussed in Section 5. If s is not covered by the benign execution, the driver calls the methodDirectedTGen (line 2-27), which is a directed input generat   </text>
<query_num> 12515 </query_num>
<text>   t much more space than our roBDD based approach, especially for dataintensive applications. 6 Related Work In recent years, there have been significant advance in automated code based test generation =-=[9, 10, 11, 13]-=-. Theoretically, these techniques can be applied to our problem of automated evidence generation. However in practice, they have inherent limitations that constrain their application. First, most thes   </text>
<query_num> 12516 </query_num>
<text>   to detect attacks, and thus vulnerabilities that are not under attack are invisible. The second type of approaches are static analysis, and notable examples include BOON [18], Splint [19], and Archer =-=[21]-=-. Static analysis is not bound to execution and thus often capable of identifying potential vulnerabilities in a program, and also it imposes no overhead at runtime. Thus, these techniques are more de   </text>
</top>
<top>
<paper_num> 126 </paper_num>
<paper_title>   Incremental Syntactic Parsing of Natural Language Corpora with Simple Synchrony Networks.  </paper_title>
<paper_abstract>   This article explores the use of Simple Synchrony Networks (SSNs) for learning to parse  English sentences drawn from a corpus of naturally occurring text. Parsing natural  language sentences requires taking a sequence of words and outputting a hierarchical  structure representing how those words fit together to form constituents. Feed-forward  and Simple Recurrent Networks have had great difficulty with this task, in part because  the number of relationships required to specify a structure is too large for the number  of unit outputs they have available. SSNs have the representational power to output  the necessary O(n  2  ) possible structural relationships, because SSNs extend the O(n)  incremental outputs of Simple Recurrent Networks with the O(n) entity outputs provided  by Temporal Synchrony Variable Binding. This article presents an incremental  representation of constituent structures which allows SSNs to make effective use of both  these dimensions. Experiments on learning to ...  </paper_abstract>
<query_num> 12601 </query_num>
<text>   . The generation aspect of this task largely distinguishes our approach from other extensions to SRNs for handling structured data. For example, the Backpropagation Through Structure (BPTS) algorithm =-=[35, 11]-=- assumes that the network is being trained to process structured input data, either for classification [10] or for transformation [11]. The transformation task is closer to that of training a parser,   </text>
<query_num> 12602 </query_num>
<text>   ach for learning language is based around the SRN, in which the network is trained to predict the next word in a sentence [7, 8, 30] or else trained to assess whether a sentence is grammatical or not =-=[24, 25]-=-. However, the simple SRN has not produced results comparable with the statistical parsers, because its basic output representation is flat and unstructured. The reason the simple SRN does not produce riments with natural language using SRNs have typically used a restricted form of input representation, either predicting the next word in a sentence [6, 8, 30] or assessing whether it is grammatical =-=[24, 25]-=-. Our extension to the SRN, the SSN, corrects this limitation by enhancing the range of output representations to include structured parse trees. Our approach is designed to generate a structured repr   </text>
<query_num> 12603 </query_num>
<text>   and describe a training algorithm. Finally we give three example SSN architectures; SSNs are defined by a restriction on the space of possible TSVB networks. 1 Methods such as Long Short-Term Memory =-=[19]-=- can help learn regularities between distant items in a sequence, but they cannot totally overcome this unhelpful bias. 2 Note that this argument applies to any domain where structured representations   </text>
<query_num> 12604 </query_num>
<text>   for learning to parse English sentences drawn from a corpus of naturally occurring text. The SSN has been defined in previous work [17, 23], and is an extension of the Simple Recurrent Network (SRN) =-=[6, 7]-=-. The SSN extends SRNs with Temporal Synchrony Variable Binding (TSVB) [33], which enables the SSN to represent structures and generalise across structural constituents. We apply SSNs to syntactic par As noted in the Introduction to this article, experiments with natural language using SRNs have typically used a restricted form of input representation, either predicting the next word in a sentence =-=[6, 8, 30]-=- or assessing whether it is grammatical [24, 25]. Our extension to the SRN, the SSN, corrects this limitation by enhancing the range of output representations to include structured parse trees. Our ap   </text>
<query_num> 12605 </query_num>
<text>   for learning to parse English sentences drawn from a corpus of naturally occurring text. The SSN has been defined in previous work [17, 23], and is an extension of the Simple Recurrent Network (SRN) =-=[6, 7]-=-. The SSN extends SRNs with Temporal Synchrony Variable Binding (TSVB) [33], which enables the SSN to represent structures and generalise across structural constituents. We apply SSNs to syntactic par rom statistical language learning [2, 3, 5, 20]. The basic connectionist approach for learning language is based around the SRN, in which the network is trained to predict the next word in a sentence =-=[7, 8, 30]-=- or else trained to assess whether a sentence is grammatical or not [24, 25]. However, the simple SRN has not produced results comparable with the statistical parsers, because its basic output represe e layers (solid lines) indicate that every unit in the source layer is connected to every unit in the target layer. As discussed above, recurrence is implemented with context units, just as with SRNs =-=[7]-=-, and the dotted lines indicate that activation from each unit in the source layer is copied to a corresponding context unit in the target layer. All three architectures possess a layer of pulsing inp   </text>
<query_num> 12606 </query_num>
<text>   ing structured data. For example, the Backpropagation Through Structure (BPTS) algorithm [35, 11] assumes that the network is being trained to process structured input data, either for classification =-=[10]-=- or for transformation [11]. The transformation task is closer to that of training a parser, but, as the conclusion of [11] makes clear, the use of BPTS relies on the input and output 18 having the sa   </text>
<query_num> 12607 </query_num>
<text>   on This article explores the use of Simple Synchrony Networks (SSNs) for learning to parse English sentences drawn from a corpus of naturally occurring text. The SSN has been defined in previous work =-=[17, 23]-=-, and is an extension of the Simple Recurrent Network (SRN) [6, 7]. The SSN extends SRNs with Temporal Synchrony Variable Binding (TSVB) [33], which enables the SSN to represent structures and general egularities that motivated the use of a structured representation in the first place. 2 For the SSN, the use of TSVB means that learned regularities inherently generalise over structural constituents =-=[15, 17]-=-, thereby capturing important linguistic properties such as systematicity [15]. It is this generalisation ability which allows the SSN parser presented in this article to scale up to naturally occurri ate way. Indeed, the performance of the SSN parser actually improves with the addition of the STM. 2 Simple Synchrony Networks In this section we provide a summary of Simple Synchrony Networks (SSNs) =-=[17, 23]-=-. We begin by describing the basic principles of Temporal Synchrony Variable Binding (TSVB) [33] which extend standard connectionist networks with pulsing units; pulsing units enable a network to prov   </text>
<query_num> 12608 </query_num>
<text>   resenting how those words fit together to form constituents, such as noun phrases and verb phrases. The state-of-the-art techniques for tackling this task are those from statistical language learning =-=[2, 3, 5, 20]-=-. The basic connectionist approach for learning language is based around the SRN, in which the network is trained to predict the next word in a sentence [7, 8, 30] or else trained to assess whether a   </text>
<query_num> 12609 </query_num>
<text>   resenting how those words fit together to form constituents, such as noun phrases and verb phrases. The state-of-the-art techniques for tackling this task are those from statistical language learning =-=[2, 3, 5, 20]-=-. The basic connectionist approach for learning language is based around the SRN, in which the network is trained to predict the next word in a sentence [7, 8, 30] or else trained to assess whether a   for enforcing syntactic grammars, which define what is and isn&amp;apos;t a possible constituent structure for a sentence. More recent work has focused on how to incorporate probabilities into these grammars =-=[2, 20]-=- and how to estimate these probabilities from a corpus of naturally occurring text. The output structure is taken to be the structure with the highest probability according to the estimates. For examp nd recall) and the percentage of correct responses on each output unit. The measure of precision and recall used for constituent evaluation is a standard measure used in statistical language learning =-=[20]-=-. The precision is the number of correct constituents output by the parser divided by the total number of constituents output by the parser. The recall is the number of correct constituents divided by that the network has learnt a robust mapping from input sentences to output parse trees. This level of generalisation (around 80% average precision/recall) is similar to that achieved by PCFG parsers =-=[20]-=-, although for a fair comparison identical experiments must be performed with each algorithm: again, we return to this point in Section 5. 4.3 Analysis of results The basic experimental results above   such as the simple Probabilistic Context Free Grammar (PCFG), direct comparisons can be made between the two approaches. For instance, the simple PCFG can achieve around 72% average precision/recall =-=[20]-=- on parsing from sequences of word-tags. In comparison, the SSN in the above experiments achieves 80% average precision/recall when trained and tested on sentences with fewer than 15 words. However, t   </text>
<query_num> 12610 </query_num>
<text>   resenting how those words fit together to form constituents, such as noun phrases and verb phrases. The state-of-the-art techniques for tackling this task are those from statistical language learning =-=[2, 3, 5, 20]-=-. The basic connectionist approach for learning language is based around the SRN, in which the network is trained to predict the next word in a sentence [7, 8, 30] or else trained to assess whether a  he network towards learning word-specific generalisations. The word-specific nature of linguistic generalisations is manifested in the current popularity of lexicalised grammar representations, as in =-=[5]-=-. Also, the short-term memory mechanism discussed later in this section is motivated by psycholinguistic phenomena. Other particular motivations will be discussed as they arise. 3.2 Structured output   </text>
</top>
<top>
<paper_num> 127 </paper_num>
<paper_title>   Limiting Result Cardinalities for Multidatabase Queries Using Histograms.  </paper_title>
<paper_abstract>   Abstract. Integrating, cleaning and analyzing data from heterogeneous sources is often complicated by the large amounts of data and its physical distribution which can result in poor query response time. One approach to speed up the processing is to reduce the cardinality of results – either by querying only the first tuples or by obtaining a sample for further processing. In this paper we address the processing of such queries in a multidatabase environment. We discuss implementations of the query operators, strategies for their placement in a query plan and particularly the usage of histograms for estimating attribute value distributions and result cardinalities in order to parameterize the operators.  </paper_abstract>
<query_num> 12701 </query_num>
<text>   a there is often also a need for cleaning and analyzing the data in order to obtain qualitatively appropriate results. By means of multidatabase languages (e.g MSQL [GLRS93], SchemaSQL [LSS96], FRAQL =-=[SCS00]-=-) we have the tools at hand which are needed for querying across diverse data sources. Querying using multiple data sources usually produces complete result sets. This requires processing large amount p a global query optimization taking this meta-data into account – focusing on the processing of ‘first Ò’and ‘sample’ queries. Our work is based on the object-relational multidatabase language FRAQL =-=[SCS00]-=- which in particular allows the dynamic addition of user-defined conflict reconciliation functions. For this language a query engine has been implemented which is able to access heterogeneous database  of a query in order to improve the response time of query evaluation. In the following we will focus on two approaches which are implemented as part of the multidatabase system FRAQL query processor =-=[SCS00]-=-: LIMIT FIRST and LIMIT SAMPLE. Both are extensions to the standard SQLSELECT statement: SELECT &amp;lt;projection list&amp;gt; FROM &amp;lt;table expression&amp;gt; [ WHERE &amp;lt;condition&amp;gt; ] [ ORDER BY &amp;lt;order spec&amp;gt; ] LIMIT FIRST|SA ue then can be inserted as parameter Ò in equation (1). A special treatment is required for histograms of relations containing attributes which are transformed by applying so-called mapping functions =-=[SCS00]-=- as part of the view definition. Because the mapping is implemented as a special query operator in the query plan the involved histogram also has to be mapped. A straightforward solution is to apply t   </text>
<query_num> 12702 </query_num>
<text>   ecution plan and ‘cuts’ the tuple stream after the desired cardinality. Following the operator introduced in [CK97], we added an operator stop to our query engine, which implements the iterator model =-=[Gra93]-=- and passes a given number of tuples from the input stream. At physical level the stop operator has several implementations: a simple pipelined scan-stop operation for unordered limitations and a bloc   </text>
<query_num> 12703 </query_num>
<text>   egration of data there is often also a need for cleaning and analyzing the data in order to obtain qualitatively appropriate results. By means of multidatabase languages (e.g MSQL [GLRS93], SchemaSQL =-=[LSS96]-=-, FRAQL [SCS00]) we have the tools at hand which are needed for querying across diverse data sources. Querying using multiple data sources usually produces complete result sets. This requires processi   </text>
<query_num> 12704 </query_num>
<text>   es are rewritten using the histograms instead of the base relations. An interactive and iterative way to provide approximate answers for aggregated queries, called online-aggregation, is described in =-=[HHW97]-=-. Here the user starts with a relative imprecise answer provided by a first small random sample of the data. This initial value will be improved during the processing. The user can observe online the   </text>
<query_num> 12705 </query_num>
<text>   issing random access to the data, sequential sampling algorithms have to be utilized. There are two types of scenarios: known and unknown relation sizes. In the first case, algorithms as described in =-=[Vit87]-=- can be used, which have the advantage of not blocking. In the second case sampling with reservoir [Li94] is necessary. These algorithms do not need the relation size, but provide the first tuples onl   </text>
<query_num> 12706 </query_num>
<text>   mputed histograms for determining approximate answers is yet another possibility to reduce the query result size and to achieve short query response times. This technique is among others described in =-=[PGI99]-=-. The authors claim that it is possible to execute non-aggregate and aggregate queries using this method. Thereby the queries are rewritten using the histograms instead of the base relations. An inter   </text>
<query_num> 12707 </query_num>
<text>   roximate answers for join aggregation queries. These synopses are well suited for star or snowflake schemas which are usual in the data warehouse area. This approach is implemented in the AQUA system =-=[AGP99]-=-, which works on top of any commercial DBMS and stores its precomputed statistical data in relations within the DBMS. For providing fast approximate answers for user queries, the system rewrites the q   </text>
<query_num> 12708 </query_num>
<text>   s. A plan for a query containing aLIMIT FIRST clause is constructed as follows: after substituting global view definitions, performing the usual transformations (e.g., standardization, simplification =-=[JK84]-=-) and decomposing the query into sub-queries processable by the sources, the optimizer seeks to insert a stop operator according to thesrules given above at the root of the sub-query. If the global re   </text>
<query_num> 12709 </query_num>
<text>   se system can increase the performance of the sample computation. They discuss several techniques for uniform random sampling from base relations or the output of relational operators. In [CMN99] and =-=[AGPR99]-=- the join sampling problem is pointed out as an example of the problem of commuting the sample operator with relational operators. [AGPR99] uses precomputed join samples, so-called join synopses, to p   </text>
<query_num> 12710 </query_num>
<text>   sired result size. Several commercial database management systems provide a similar technique to compute the top-Ò results. Sampling is another technique for data reduction. The authors of [OR86] and =-=[Olk93]-=- describe different kinds of uniform random sampling techniques in a DBMS, because the integration of sampling in a database system can increase the performance of the sample computation. They discuss  using the notationLIMIT SAMPLE &amp;lt;value expr&amp;gt; [PERCENT] the system generates a simple uniform sample of size n of the query result. An efficient computation requires a sample operator, as described in =-=[Olk93]-=-, being applied as low as possible in the query plan. As we are in a multi database environment, there are several constraints, which have to be considered. Our system uses virtual integrated relation CMN99]. Possible strategies are: – Naive sampling includes a first complete computation of the join of Ê and Ì followed by the application of the sample operator. – The second strategy is proposed in =-=[Olk93]-=- and includes the following steps. Consider the computation of a join of Ê and Ì . First sample uniform randomly one (1)stuple from Ê and join it with Ì and getting the result Î . Select randomly one   . The result is Ë . 3. The last step consists of picking out one tuple from each group of Ë using a unweighted random sample algorithm. With the above constraints, the sampling approach according to =-=[Olk93]-=- cannot be applied in our environment because it requires an index as well as full statistics. However, the naive sampling algorithm or group-sample is possible. To support the latter strategy the FRA   </text>
<query_num> 12711 </query_num>
<text>   tes c.cust id and i.cust id (denoted as hist (cust id) and hist (cust id)) of the base relations the buckets of the histogram hist�(cust id) for the join �can be calculated using the following formula=-=[SS94]-=-: �� � ���� ��Ù ��Ø � � � �Ù ��Ø � £ �Ù ��Ø� � Ñ�Ü � ��� Here, � and �� are the numbers of distinct values present in the join column from or � respectively. If the histograms do not coincide a preced   </text>
</top>
<top>
<paper_num> 128 </paper_num>
<paper_title>   Analyzing an Infinite Parallel Job Allocation Process.  </paper_title>
<paper_abstract>   . In recent years the task of allocating jobs to servers has been studied with the &amp;quot;balls and bins&amp;quot; abstraction. Results in this area exploit the large decrease in maximum load that can be achieved by allowing each job (ball) a very small amount of choice in choosing its destination server (bin). The scenarios considered can be divided into two categories: sequential, where each job can be placed at a server before the next job arrives, and parallel, where the jobs arrive in large batches that must be dealt with simultaneously. Another, orthogonal, classification of load balancing scenarios is into fixed time and infinite. Fixed time processes are only analyzed for an interval of time that is known in advance, and for all such results thus far either the number of rounds or the total expected number of arrivals at each server is a constant. In the infinite case, there is an arrival process and a deletion process that are both defined over an infinite time line. In this paper, we presen...  </paper_abstract>
<query_num> 12801 </query_num>
<text>   alls and achieves the optimal load of O( m n ) using log log n log(m=n) rounds of communication, w.h.p., or load max \Phi r p log n; O \Gamma m n \Delta\Psi using r rounds of communication, w.h.p. In =-=[BMS97]-=- the authors extend the lower bound of [ACMR95] to arbitrary rslog log n, implying that the result of Stemanns protocol is optimal for all r. Their main result is a generalization of Stemanns upper bo   </text>
<query_num> 12802 </query_num>
<text>   and each bin accepts up to T balls during each round. They show that with T = r q (2r+o(1))\Deltalog n log log n this algorithm terminates after r rounds with maximum load r \Delta T , w.h.p. Stemann =-=[St96]-=- extends the results for the case where the number of balls m is larger than the number n of bins. For m = n, he analyzes a very simple class of algorithms achieving maximum load O i r q log n log log   </text>
<query_num> 12803 </query_num>
<text>   are rather limited in their applicability to parallel and distributed settings, and thus much of the work on allocation processes has concentrated on the parallel scenario ([ACMR95,Mit96,St96,BMS97], =-=[Mit97]-=-), where the jobs arrive in large batches that must be processed simultaneously. The seminal paper on the sequential scenario [ABKU94] considered both a fixed time (which they call finite) process, an ting time to be O(1) for N ! 1, and the maximum queue length to be O(log log n+o(1)), w.h.p. His analysis makes use of deep results of Kurtz ([Kur81]) on so called density dependent Markov Chains. In =-=[Mit97]-=- the author extends his results to several different load generation and consumption schemes. For example, he analyzes the same process with constant service times, the customers having a different nu   </text>
<query_num> 12804 </query_num>
<text>   at least 1 \Gamma 1 n ff for any constant ff ? 0. It is easy to show that the same result holds if the jobs are generated by n generators which can be arbitrarily distributed over the processors (see =-=[SV96]-=-). Each generator is allowed to produce a job with a probability smaller than 1 2de per round. It is also possible to use generators with different generation probabilities if the expected overall gen   </text>
<query_num> 12805 </query_num>
<text>   balls, w.h.p., where n is both the number of balls and bins in the system. The simple sequential game has many applications and is also used as an online algorithm for competitive Load Balancing (see =-=[ABK94]-=-, [AKP + 93], and [PW93]). Recently, Czumaj et al. ([CS97]) extended the results in several directions. They present an adaptive process where the number of choices made in order to place a ball depen   </text>
<query_num> 12806 </query_num>
<text>   both the number of balls and bins in the system. The simple sequential game has many applications and is also used as an online algorithm for competitive Load Balancing (see [ABK94], [AKP + 93], and =-=[PW93]-=-). Recently, Czumaj et al. ([CS97]) extended the results in several directions. They present an adaptive process where the number of choices made in order to place a ball depends on the load of the pr   </text>
<query_num> 12807 </query_num>
<text>   f the jobs are generated asynchronously, for instance by a Poisson Process. The analysis of the process is performed using a type of delay sequence argument, using a structure known as a witness tree =-=[MSS95]-=-. To use this type of argument, we first show that every time the process fails a witness tree must exist. We then show that that it is very unlikely for a witness tree to occur. Usually the latter in   </text>
<query_num> 12808 </query_num>
<text>   ion process that are both defined over an infinite time line. Despite a large subsequent body of work on the parallel scenario as well as further work on infinite processes in the sequential scenario =-=[CS97]-=-, up to now there has been no analysis of an infinite process for the parallel scenario. In fact, for all previous parallel results, either the number of rounds or the total expected number of arrival  in the system. The simple sequential game has many applications and is also used as an online algorithm for competitive Load Balancing (see [ABK94], [AKP + 93], and [PW93]). Recently, Czumaj et al. (=-=[CS97]-=-) extended the results in several directions. They present an adaptive process where the number of choices made in order to place a ball depends on the load of the previously chosen bins, and an off-l   </text>
<query_num> 12809 </query_num>
<text>   of bins. For m = n, he analyzes a very simple class of algorithms achieving maximum load O i r q log n log log n j if r rounds of communication are allowed. This matches the lower bound presented in =-=[ACMR95]-=-. He generalizes the algorithm for m ? n balls and achieves the optimal load of O( m n ) using log log n log(m=n) rounds of communication, w.h.p., or load max \Phi r p log n; O \Gamma m n \Delta\Psi u   </text>
<query_num> 12810 </query_num>
<text>   rdination between jobs and servers. Often, this problem is stated in terms of balls and bins; such occupancy results have several applications in load balancing, hashing, and PRAM simulation [KLM92], =-=[ABKU94]-=-, [ACMR95]. A simple distributed algorithm for allocating jobs to servers is to place each job at a random server. This requires no coordination, but it is well known that if there are n jobs and n se  concentrated on the parallel scenario ([ACMR95,Mit96,St96,BMS97], [Mit97]), where the jobs arrive in large batches that must be processed simultaneously. The seminal paper on the sequential scenario =-=[ABKU94]-=- considered both a fixed time (which they call finite) process, and an infinite process. Fixed time processes are only analyzed for a fixed interval of time that is known in advance, where infinite pr st cannot be bounded. We develop a method for using a witness tree argument in the analysis of an infinite process; this is the main technical contribution of our paper. 1.2 Previous Work Azar et al. =-=[ABKU94]-=- examine a sequential protocol called greedy process to place n balls into n bins. For each ball they choose d bins i.u.r. and put the ball into the bin with minimum load at the time of placement. The h the new one) among the chosen bins. They show that a process with reassignments yields a maximum load that is never smaller by more than an constant factor than the maximum load of the process from =-=[ABKU94]-=-. Adler et al. [ACMR95] explore the problem in parallel and distributed settings for the case of placing n balls into n bins. They provide a lower bound for non-adaptive (possible destinations are cho onstant number r 2 IN of communication rounds, the maximum load is shown to be at least\Omega i r q log n log log n j . Additionally, they present parallelizations of the sequential strategy found in =-=[ABKU94]-=-. They give a two-round parallelization of the greedy process, matching the lower bound. Furthermore, they introduce a multiple-round strategy requiring log log n +O(1) rounds of communication and ach   </text>
</top>
<top>
<paper_num> 129 </paper_num>
<paper_title>   Detecting Communities in Social Networks Using Max-Min Modularity.  </paper_title>
<paper_abstract>   Many datasets can be described in the form of graphs or networks where nodes in the graph represent entities and edges represent relationships between pairs of entities. A common property of these networks is their community structure, considered as clusters of densely connected groups of vertices, with only sparser connections between groups. The identification of such communities relies on some notion of clustering or density measure. which defines the communities that can be found. However, previous community detection methods usually apply the same structural measure on all kinds of networks, despite their distinct dissimilar features. In this paper, we present a new community mining measure, Max-Min Modularity, which considers both connected pairs and criteria defined by domain experts in finding communities, and then specify a hierarchical clustering algorithm to detect communities in networks. When applied to real world networks for which the community structures are already known, our method shows improvement over previous algorithms. In addition, when applied to randomly generated networks for which we only have approximate information about communities, it gives promising results which shows the algorithm’s robustness against noise.  </paper_abstract>
<query_num> 12901 </query_num>
<text>   , making it prohibitively difficult to solve for large graphs. However, a wide variety of heuristic algorithms have been developed and give good solutions in many cases, e.g., multilevel partitioning =-=[22]-=-, kpartite graph partitioning [24], relational clustering [25], flow-based methods [12], information-theoretic methods [9] and spectral clustering [33]. The main problem for these methods is that inpu   </text>
<query_num> 12902 </query_num>
<text>   However, a wide variety of heuristic algorithms have been developed and give good solutions in many cases, e.g., multilevel partitioning [22], kpartite graph partitioning [24], relational clustering =-=[25]-=-, flow-based methods [12], information-theoretic methods [9] and spectral clustering [33]. The main problem for these methods is that input parameters such as the number of the partitions and their si   </text>
<query_num> 12903 </query_num>
<text>   algorithm does not require parameters. The idea of considering domain knowledge as related/unrelated pairs in this paper is analogous to the notions of must/cannot links in semi-supervised clustering =-=[3, 39, 40]-=-. However, in semi-supervised clustering, the labeled data is used for cluster initialization [3] and the link constraints must be satisfied [39, 40]. Moreover, the number of clusters k is usually req   </text>
<query_num> 12904 </query_num>
<text>   an overlapping community structure, which is hard to grasp with classical clustering methods where every vertex of the graph belongs to only one community. Based on these observations, fuzzy methods =-=[18, 26, 32, 46]-=- and dynamic approaches [4, 38] have been proposed for overlapping structure and dynamic community detection. Recent work by Xu et al. [42] proposed a fast SCAN algorithm to detect not only clusters,   </text>
<query_num> 12905 </query_num>
<text>   and many of the edges are unobserved, i.e. labeled as 0. Moreover, while existing methods [15, 16] for correlation clustering require the user to specify parameters that are usually hard to determine =-=[1]-=-, e.g., the number of clusters, our algorithm does not require parameters. The idea of considering domain knowledge as related/unrelated pairs in this paper is analogous to the notions of must/cannot   </text>
<query_num> 12906 </query_num>
<text>   er positive or negative, such assumption is not true for community detection, where graphs are usually sparse and many of the edges are unobserved, i.e. labeled as 0. Moreover, while existing methods =-=[15, 16]-=- for correlation clustering require the user to specify parameters that are usually hard to determine [1], e.g., the number of clusters, our algorithm does not require parameters. The idea of consider   </text>
<query_num> 12907 </query_num>
<text>   nd may fail to identify communities smaller than a certain scale [13]. Possible solutions include recursive algorithms based on modularity optimization [34]. At last, as pointed out by Scripps et al. =-=[35]-=-, the modularity only measures existing links on the network, but does not explicitly consider the absent links between two nodes in the same community. In other words, the modularity only measures ho nected node pairs within community 1, which is not considered by the Q measure. Therefore, modularity fails to compare the community structure between different graphs. To solve this problem, Scripps =-=[35]-=- proposed two ratios p and q, measuring the fraction of links within communities and absent links between communities, respectively. The drawback of their method is that the interpretation is not clea   </text>
<query_num> 12908 </query_num>
<text>   ons in many cases, e.g., multilevel partitioning [22], kpartite graph partitioning [24], relational clustering [25], flow-based methods [12], information-theoretic methods [9] and spectral clustering =-=[33]-=-. The main problem for these methods is that input parameters such as the number of the partitions and their sizes are usually required, but we do not typically know how many communities there are, an   </text>
<query_num> 12909 </query_num>
<text>   sets of web pages on related topics, which can enable search engines and portals to increase the precision and recall of search results by focusing on narrow but topically-related subsets of the web =-=[12]-=-; groups within social networks might correspond to social communities, which can be used to understand the data, such as organization structures, academic collaborations and the communities in tele-c of heuristic algorithms have been developed and give good solutions in many cases, e.g., multilevel partitioning [22], kpartite graph partitioning [24], relational clustering [25], flow-based methods =-=[12]-=-, information-theoretic methods [9] and spectral clustering [33]. The main problem for these methods is that input parameters such as the number of the partitions and their sizes are usually required,   </text>
<query_num> 12910 </query_num>
<text>   the connected pairs might be inaccurate for community structure detection. While other algorithms cannot handle such case, our MM modularity-based methods can achieve information from link prediction =-=[23]-=-, and extract appropriate criteria for community detection. Additionally, our algorithm still runs in O(mD log n) time, which is the same as previous modularity-based algorithms. 7 Conclusions We have   </text>
<query_num> 12911 </query_num>
<text>   to equalsized parts and thus still suffer from the same drawbacks that make graph partitioning inappropriate for community mining. In the field of theoretical computer science, correlation clustering =-=[2, 5, 37]-=- considers a complete graph on n vertices, where each edge (u, v) is labeled either + or − depending on whether u and v have been deemed to be similar or different. Similar to our problem, the goal is   </text>
<query_num> 12912 </query_num>
<text>   veloped and give good solutions in many cases, e.g., multilevel partitioning [22], kpartite graph partitioning [24], relational clustering [25], flow-based methods [12], information-theoretic methods =-=[9]-=- and spectral clustering [33]. The main problem for these methods is that input parameters such as the number of the partitions and their sizes are usually required, but we do not typically know how m   </text>
<query_num> 12913 </query_num>
<text>   without considering their linking structure would definitely lead to unsatisfactory results for queries. Various relation-based methods have been developed, such as the spectral clustering approaches =-=[10, 36, 41]-=- and modularity-based algorithms [7, 30]. However, none of them distinguish the intrinsic features of the domain of the network in question. In other words, the same 978 Copyright © by SIAM. Unauthori ies there are, and there is no reason that they should be roughly the same size. Various benefit functions have been proposed to avoid the problem, such as the normalized cut [36] and the min-max cut =-=[10]-=-. However, these approaches are biased in favor of divisions into equalsized parts and thus still suffer from the same drawbacks that make graph partitioning inappropriate for community mining. In the   </text>
<query_num> 12914 </query_num>
<text>   without considering their linking structure would definitely lead to unsatisfactory results for queries. Various relation-based methods have been developed, such as the spectral clustering approaches =-=[10, 36, 41]-=- and modularity-based algorithms [7, 30]. However, none of them distinguish the intrinsic features of the domain of the network in question. In other words, the same 978 Copyright © by SIAM. Unauthori ly know how many communities there are, and there is no reason that they should be roughly the same size. Various benefit functions have been proposed to avoid the problem, such as the normalized cut =-=[36]-=- and the min-max cut [10]. However, these approaches are biased in favor of divisions into equalsized parts and thus still suffer from the same drawbacks that make graph partitioning inappropriate for  is found for these “structureless” networks is an important question to answer. To solve this problem, computer scientists proposed several benefit functions based on cut sizes, e.g., normalized cut =-=[36]-=-. However, cut sizes do not accurately reflect the intuitive concept of social network communities and thus are the wrong measure to optimize. A good division of a network into communities is not mere   </text>
<query_num> 12915 </query_num>
<text>   without considering their linking structure would definitely lead to unsatisfactory results for queries. Various relation-based methods have been developed, such as the spectral clustering approaches =-=[10, 36, 41]-=- and modularity-based algorithms [7, 30]. However, none of them distinguish the intrinsic features of the domain of the network in question. In other words, the same 978 Copyright © by SIAM. Unauthori strong community structure. Q typically falls in the range from 0.3 to 0.7 [17] and high values are rare. 3.2 Drawbacks of Modularity Q The modularity approach has been pursued by a number of authors =-=[11, 19, 28, 41]-=-, and has been proved highly effective in practice for community evaluation [8]. However, there are three major problems for the Q measure. At first, the modularity requires information of the entire   </text>
</top>
<top>
<paper_num> 130 </paper_num>
<paper_title>   An Image Morphing Technique Based on Optimal Mass Preserving Mapping.  </paper_title>
<paper_abstract>   Abstract—Image morphing, or image interpolation in the time domain, deals with the metamorphosis of one image into another. In this paper, a new class of image morphing algorithms is proposed based on the theory of optimal mass transport. The P mass moving energy functional is modified by adding an intensity penalizing term, in order to reduce the undesired double exposure effect. It is an intensity-based approach and, thus, is parameter free. The optimal warping function is computed using an iterative gradient descent approach. This proposed morphing method is also extended to doubly connected domains using a harmonic parameterization technique, along with finite-element methods. Index Terms—Image interpolation, image morphing, image warping, mass preserving mapping, Monge–Kantorovich flow, optimal transport. I.  </paper_abstract>
<query_num> 13001 </query_num>
<text>   asonable as it reflects physical reality. Registration and interpolation of smoke, clouds, flames and fluids is an active area of research in the computer graphics and image processing community. See =-=[10]-=-, [24], and [33], for example, for related methods. This can serve as an image compression method by storing the selected frames and generating the omitted frames on-the-fly when the video is played.   </text>
<query_num> 13002 </query_num>
<text>   eflects physical reality. Registration and interpolation of smoke, clouds, flames and fluids is an active area of research in the computer graphics and image processing community. See [10], [24], and =-=[33]-=-, for example, for related methods. This can serve as an image compression method by storing the selected frames and generating the omitted frames on-the-fly when the video is played. Our extension of   </text>
<query_num> 13003 </query_num>
<text>   g the pseudo mass density to unity on the entire domain of the image. Other approaches to using various classes of diffeomorphisms for registration and warping may be found for example in [25], [31], =-=[32]-=-, and the references therein. In this paper, we present an improved approach for image morphing, with special efforts taken to reduce the double exposure effect (also referred to as the “fade-in and f   </text>
<query_num> 13004 </query_num>
<text>   ieee.org. Digital Object Identifier 10.1109/TIP.2007.896637 Lei Zhu, Yan Yang, Steven Haker, and Allen Tannenbaum 1057-7149/$25.00 © 2007 IEEE field warping, and energy-based warping. In mesh warping =-=[35]-=-, features are specified by a nonuniform control mesh, and the warping function is usually generated by a spline interpolation. This class of mesh warping algorithm usually shows good distortion behav   </text>
<query_num> 13005 </query_num>
<text>   ing, Monge–Kantorovich flow, optimal transport. I. INTRODUCTION IMAGE morphing, sometimes referred to as “image interpolation in the time domain,” deals with the metamorphosis of one image to another =-=[21]-=-. It is a technique widely used in television commercials, music videos and motion pictures. Image morphing has also been used for facial recognition [37]. Given a pair of images, the goal of image mo to as “ghosts”). Energy minimization-based warpings usually guarantee the one-to-one mapping property, which prevents the warped image from folding back upon itself. For example, in Lee et al.’s work =-=[21]-=-, points, polylines, and curves are sampled and reduced to a collection of points. These points are then used to generate the warping function by minimizing an energy functional. A similar method has  med that when time varies from 0 to 1, the starting image continuously changes to the ending image . We further require that the same transition rate is applied to all points on the in-between images =-=[21]-=-. Hence, the image warping map at any time is simply given by (21) and the corresponding cross-dissolved image at time is given by (22) and can also be color images and (22) can be applied to three co   </text>
<query_num> 13006 </query_num>
<text>   ity between the transformed source image and the target image, e.g., the sum of squared difference (SSD), likelihood measurement, correlation ratio, normalized correlation, or mutual information (MI) =-=[34]-=-. In this paper, SSD and MI are adopted as the similarity measures, based on the characteristics of our testing images. If SSD is used as the similarity measure [14], we are minimizing the following e   </text>
<query_num> 13007 </query_num>
<text>   le as it reflects physical reality. Registration and interpolation of smoke, clouds, flames and fluids is an active area of research in the computer graphics and image processing community. See [10], =-=[24]-=-, and [33], for example, for related methods. This can serve as an image compression method by storing the selected frames and generating the omitted frames on-the-fly when the video is played. Our ex   </text>
<query_num> 13008 </query_num>
<text>   or image retrieval. Our interest in MKP arose from our visualization work in medical applications. For example, a flattened representation of colon surface is helpful for the detection of colon polys =-=[13]-=- and a flattened representation of vessel surface is useful for the study of the correlation between wall shear stress and the development of atherosclerosis [39]. Among various flattening techniques,  connected domain (here denotes a square root of 1). Similar techniques have been applied for colon surface visualization Fig. 1. Dubly onnected domain 6 with inner boundary &amp;apos; and outter boundary &amp;apos; . =-=[13]-=-, tissue thickness measurement [36], and defining orthogonal curves for template matching [30]. Assume we have a doubly connected domain , which has two boundaries: the inner boundary and the outer bo is solved to give the value of inside domain . Numerically, we are working on a triangulated domain and the Laplace equations are solved by a standard FEM technique as we used for colon visualization =-=[13]-=-. Once the analytic function is obtained, a curvilinear harmonic polar coordinate system can be defined by taking as one coordinate and as another. can be thought of as a curvilinear “radius” and as t   </text>
<query_num> 13009 </query_num>
<text>   ply assigning the pseudo mass density to unity on the entire domain of the image. Other approaches to using various classes of diffeomorphisms for registration and warping may be found for example in =-=[25]-=-, [31], [32], and the references therein. In this paper, we present an improved approach for image morphing, with special efforts taken to reduce the double exposure effect (also referred to as the “f   </text>
<query_num> 13010 </query_num>
<text>   quare root of 1). Similar techniques have been applied for colon surface visualization Fig. 1. Dubly onnected domain 6 with inner boundary &amp;apos; and outter boundary &amp;apos; . [13], tissue thickness measurement =-=[36]-=-, and defining orthogonal curves for template matching [30]. Assume we have a doubly connected domain , which has two boundaries: the inner boundary and the outer boundary , as shown in Fig. 1. The re   </text>
<query_num> 13011 </query_num>
<text>   separate the myocardium and the ventricles, and then perform image morphing only on the myocardial region of interest. In this section, we extend the MP mapping algorithm to doubly connected domains =-=[41]-=-. The main difficulty comes from the construction of the initial MP mapping. We propose an algorithm that finds by using harmonic parameterization. In this approach, the two domains are first harmonic   </text>
<query_num> 13012 </query_num>
<text>   special shapes or features of the objects in the image. See [31] for a review of the literature and an extensive list of references on this subject. We should also mention the nice work of Iwanowski =-=[17]-=- in which an image morphing approach is proposed that combines morphological interpolation and linear filtering and does not require control points or landmarks. In this paper, we present an automatic   </text>
<query_num> 13013 </query_num>
<text>   transportation, statistical physics, shape optimization, expert systems, and meteorology [26]. Recently, this problem has been studied within the context of content-based image retrieval [22], [27], =-=[28]-=-. Pixels in an image are divided into several bins according to their color and spatial locations. The Earth Mover’s distance (EMD) is then calculated between the bins of two images and used for image   </text>
</top>
<top>
<paper_num> 131 </paper_num>
<paper_title>   Surface Mesh Smoothing, Regularization, and Feature Detection.  </paper_title>
<paper_abstract>   We describe a hybrid algorithm that is designed to reconstruct a piecewise smooth surface mesh from noisy input. While denoising, our method simultaneously regularizes triangle meshes on flat regions for further mesh processing and preserves crease sharpness for faithful reconstruction. A clustering technique, which combines K-means and geometric a priori information, is first developed and refined. It is then used to implement vertex classification so that we can not only apply different smoothing operators on different vertex groups for different purposes, but also succeed in crease detection, where the tangent plane of the surface is discontinuous, without any significant cost increase. Consequently we are capable of efficiently obtaining different mesh segmentations, depending on user input and  </paper_abstract>
<query_num> 13101 </query_num>
<text>   avy noise, K-means may not give us a very satisfactory vertex partition. One more advanced clustering approach, which has been shown to handle more complicated structured data, is spectral clustering =-=[1, 22, 32, 40, 35, 30]-=-. It does not require estimating an explicit model of data distribution, but rather employs a spectral analysis of the matrix of point-to-point similarities. However, the real power of spectral cluste   </text>
<query_num> 13102 </query_num>
<text>   clustering was applied on triangular mesh faces for hierarchical mesh decomposition in [24]. Liu and Zhang [27] proposed a 3D mesh segmentation algorithm through spectral clustering of mesh faces. In =-=[7]-=-, Chen et al used Bayesian discriminant analysis to determine the decision boundary for separating potential feature and non-feature vertices in curvature space. However, we believe that the present p   </text>
<query_num> 13103 </query_num>
<text>   erators based upon geometric isotropic diffusion were proposed in the 1990s [38, 16]. Later, algorithms based on anisotropic diffusion were introduced to avoid smearing out of real important features =-=[11, 8, 3, 37, 4, 17, 9]-=-. These methods are typically expensive, both in terms of cost per iteration and in the number of iterations required to achieve satisfactory results. Moreover, they often require the user to provide   </text>
<query_num> 13104 </query_num>
<text>   ex neighborhood choices and the manner in which tangent planes are approximated. Further, sampling irregularities in the given mesh occasionally distort results and significantly slow algorithms down =-=[14, 25, 10, 18]-=-. Several discrete operators were designed to overcome this numerical difficulty and maintain sampling irregularity in the smoothed meshes. To satisfy different application requirements, a smoothing m h sampling rates, a better choice is the scale-dependent version [14]. Further, to solve problems arising from unequal face angles, a better approximation to the mean curvature normal was proposed in =-=[10, 28]-=- which doesn’t produce vertex tangential drifting when surfaces are relatively flat and compensates both for unequal edge lengths and for unequal face angles. This approach does not improve the mesh i e given vertex set, thus increasing the influence of the given data during the denoising process at all iterates [39]. This also helps to reduce the effect of volume shrinkage over several iterations =-=[38, 10, 18]-=-. Since the surface scale is local and the mesh is generally nonuniform, we choose λi at each such iteration depending on the spatial location i. Also, since noise is carried in the position of vertic   </text>
<query_num> 13105 </query_num>
<text>   fusion process. The approach of bilateral filtering has given rise to methods that more rapidly yield results of a quality similar to anisotropic diffusion, albeit with less theoretical justification =-=[13, 21, 44]-=-. The latter methods usually require very few, cheap iterations, and cost but a tiny fraction of the computational effort required to carry out an elaborate anisotropic diffusion process, especially o   </text>
<query_num> 13106 </query_num>
<text>   ge processing literature uses tools from robust statistics to sharpen image edges and corners by automatically estimating σ as the median absolute deviation of the given image intensity gradient, see =-=[5, 33]-=-. Extending this selection to polygonal meshes, our numerical experiments show that for preserving edge sharpness (4b) works much better than (4a). Similar conclusions are reached in [42], where mesh   </text>
<query_num> 13107 </query_num>
<text>   i + τ∆xi + λi(vi − xi), i = 1, . . . , N, λ &amp;gt; 0, (6) where {vi; i = 1, . . . , N} is the given vertex set, thus increasing the influence of the given data during the denoising process at all iterates =-=[39]-=-. This also helps to reduce the effect of volume shrinkage over several iterations [38, 10, 18]. Since the surface scale is local and the mesh is generally nonuniform, we choose λi at each such iterat   </text>
<query_num> 13108 </query_num>
<text>   ient, see [5, 33]. Extending this selection to polygonal meshes, our numerical experiments show that for preserving edge sharpness (4b) works much better than (4a). Similar conclusions are reached in =-=[42]-=-, where mesh smoothing algorithms via mean and median filters applied to face normals are compared. Unfortunately, median absolute deviation scaling cannot generate enough smoothing on large flat regi   </text>
<query_num> 13109 </query_num>
<text>   in time involving such surfaces often require subsequent smoothing and regularization. Fast and simple 3D mesh smoothing operators based upon geometric isotropic diffusion were proposed in the 1990s =-=[38, 16]-=-. Later, algorithms based on anisotropic diffusion were introduced to avoid smearing out of real important features [11, 8, 3, 37, 4, 17, 9]. These methods are typically expensive, both in terms of co   </text>
<query_num> 13110 </query_num>
<text>   in time involving such surfaces often require subsequent smoothing and regularization. Fast and simple 3D mesh smoothing operators based upon geometric isotropic diffusion were proposed in the 1990s =-=[38, 16]-=-. Later, algorithms based on anisotropic diffusion were introduced to avoid smearing out of real important features [11, 8, 3, 37, 4, 17, 9]. These methods are typically expensive, both in terms of co a noisy mesh of this sort,4 and we denote its vertices x(0) = v = {vi; i = 1, . . . , N}. 2.1 Discrete isotropic Laplacian The simplest discrete isotropic Laplacian operator is the Umbrella operator =-=[38]-=-. It averages the neighboring edges ∆xi = 1 ∑ ei,k, (1) mi k∈N (i) where mi = |N (i)|, the number of neighbors of vertex xi. This is a linear form implying the assumption that all neighboring edge len e given vertex set, thus increasing the influence of the given data during the denoising process at all iterates [39]. This also helps to reduce the effect of volume shrinkage over several iterations =-=[38, 10, 18]-=-. Since the surface scale is local and the mesh is generally nonuniform, we choose λi at each such iteration depending on the spatial location i. Also, since noise is carried in the position of vertic   </text>
</top>
<top>
<paper_num> 132 </paper_num>
<paper_title>   Modeling and Integrating Background Knowledge in Data Anonymization.  </paper_title>
<paper_abstract>   Recent work has shown the importance of considering the adversary’s background knowledge when reasoning about privacy in data publishing. However, it is very difficult for the data publisher to know exactly the adversary’s background knowledge. Existing work cannot satisfactorily model background knowledge and reason about privacy in the presence of such knowledge. This paper presents a general framework for modeling the adversary’s background knowledge using kernel estimation methods. This framework subsumes different types of knowledge (e.g., negative association rules) that can be mined from the data. Under this framework, we reason about privacy using Bayesian inference techniques and propose the skyline (B, t)privacy model, which allows the data publisher to enforce privacy requirements to protect the data against adversaries with different levels of background knowledge. Through an extensive set of experiments, we show the effects of probabilistic background knowledge in data anonymization and the effectiveness of our approach in both privacy protection and utility preservation.  </paper_abstract>
<query_num> 13201 </query_num>
<text>   [15], [16], [4], [38], [39]. Other anonymization techniques include clustering [40], [41], [42], [43], space mapping [44], spatial indexing [45], marginals releasing [46], and data perturbation [47], =-=[48]-=-. On the theoretical side, optimal k-anonymity has been proved to be NP-hard for k ≥ 3 in [49], and approximation algorithms for finding the anonymization that suppresses the fewest cells have been pr   </text>
<query_num> 13202 </query_num>
<text>   [22], [33], [34], [35], [36], [21], [37] and bucketization [15], [16], [4], [38], [39]. Other anonymization techniques include clustering [40], [41], [42], [43], space mapping [44], spatial indexing =-=[45]-=-, marginals releasing [46], and data perturbation [47], [48]. On the theoretical side, optimal k-anonymity has been proved to be NP-hard for k ≥ 3 in [49], and approximation algorithms for finding the   </text>
<query_num> 13203 </query_num>
<text>   ach to analyze how an adversary can gain sensitive information from the published data. Anonymization Techniques. Most anonymization solutions adopt generalization [13], [14], [32], [22], [33], [34], =-=[35]-=-, [36], [21], [37] and bucketization [15], [16], [4], [38], [39]. Other anonymization techniques include clustering [40], [41], [42], [43], space mapping [44], spatial indexing [45], marginals releasi   </text>
<query_num> 13204 </query_num>
<text>   analyze how an adversary can gain sensitive information from the published data. Anonymization Techniques. Most anonymization solutions adopt generalization [13], [14], [32], [22], [33], [34], [35], =-=[36]-=-, [21], [37] and bucketization [15], [16], [4], [38], [39]. Other anonymization techniques include clustering [40], [41], [42], [43], space mapping [44], spatial indexing [45], marginals releasing [46   </text>
<query_num> 13205 </query_num>
<text>   and Optimization Objectives. Knowledge about the algorithms and optimization objectives for anonymizing data can be used to help adversaries infer the original data, as shown recently by Wong et al. =-=[12]-=-. This kind of knowledge cannot be modeled using prior belief function about individuals. It is an interesting research direction to study this and other kinds of knowledge that may enable an adversar e in the overall table should also be public information. Recently, the σ-presence measure [28] observes that knowing an individual is in the database poses privacy risks. The m-confidentiality model =-=[12]-=- recognizes that knowledge of the mechanism or algorithm of anonymization can leak extra sensitive information. Dynamic dataset re-publication [29], [30], [25], [31] considers the scenario where an ad und Knowledge: This paper considers background knowledge that can be mined• • from the data to be released. In practice, the adversary may have access to additional background knowledge. Wong et al. =-=[12]-=- study how to protect the data against an adversary who has knowledge of the mechanism or algorithm of anonymization. It is interesting to examine other kinds of adversarial knowledge that can lead to   </text>
<query_num> 13206 </query_num>
<text>   approach in both privacy protection and utility preservation. I. INTRODUCTION A number of privacy models have been proposed for data anonymization, e.g., k-anonymity [1], ℓ-diversity [2], tcloseness =-=[3]-=-, and so on. A key limitation of these models is that they cannot guarantee that the sensitive attribute values of individuals are protected when the adversary has additional knowledge (called backgro 1 qi However, none of the above distance measures satisfy the semantic awareness property. One distance measure that takes value semantics into consideration is the Earth Mover’s Distance (EMD) [20], =-=[3]-=-. The EMD is based on the minimal amount of work needed to transform one distribution to another by moving distribution mass between each other. Unfortunately, EMD does not have the probability scalin and who is not in the table. A few subsequent works [2], [26], [27] recognize that the adversary also has knowledge of the distribution of the sensitive attribute in each group. The t-closeness model =-=[3]-=- proposes that the distribution of the sensitive attribute in the overall table should also be public information. Recently, the σ-presence measure [28] observes that knowing an individual is in the d   </text>
<query_num> 13207 </query_num>
<text>   ase poses privacy risks. The m-confidentiality model [12] recognizes that knowledge of the mechanism or algorithm of anonymization can leak extra sensitive information. Dynamic dataset re-publication =-=[29]-=-, [30], [25], [31] considers the scenario where an adversary has knowledge of previous releases of the dataset. None of these models consider correlational knowledge. Background Knowledge Integration.   </text>
<query_num> 13208 </query_num>
<text>   ation [13], [14], [32], [22], [33], [34], [35], [36], [21], [37] and bucketization [15], [16], [4], [38], [39]. Other anonymization techniques include clustering [40], [41], [42], [43], space mapping =-=[44]-=-, spatial indexing [45], marginals releasing [46], and data perturbation [47], [48]. On the theoretical side, optimal k-anonymity has been proved to be NP-hard for k ≥ 3 in [49], and approximation alg   </text>
<query_num> 13209 </query_num>
<text>   bility ditribution for each tj, we obtain Ω(si|tj) = ni × ∑ m r=1 nr × P(si|tj) ∑ k j ′ =1 P(si|t j ′) P(sr|tj) ∑ k j ′ =1 P(sr|t j ′) The above estimation technique makes the random world assumption =-=[18]-=-, where every reasonable mapping between individuals and sensitive attribute values is equally probable. Specifically, Equation (5) can be directly derived from the formula shown in Equation (4) by as   </text>
<query_num> 13210 </query_num>
<text>   ctiveness of our approach in both privacy protection and utility preservation. I. INTRODUCTION A number of privacy models have been proposed for data anonymization, e.g., k-anonymity [1], ℓ-diversity =-=[2]-=-, tcloseness [3], and so on. A key limitation of these models is that they cannot guarantee that the sensitive attribute values of individuals are protected when the adversary has additional knowledge  values is equally probable. Specifically, Equation (5) can be directly derived from the formula shown in Equation (4) by assuming P(S − {si}|E − {tj}) = P(S − {si}|E − {tj ′}) for all 1 ≤ j′ ≤ k. In =-=[2]-=-, Machanavajjhala et al. studied the problem of calculating the posterior belief under the framework of generalization by employing the random world theory. Not surprisingly, (5)t1 t2 t3 P(HIV |t1) = [1], [14] assumes that the adversary has access to some publiclyavailable databases (e.g., a vote registration list) and the adversary knows who is and who is not in the table. A few subsequent works =-=[2]-=-, [26], [27] recognize that the adversary also has knowledge of the distribution of the sensitive attribute in each group. The t-closeness model [3] proposes that the distribution of the sensitive att   </text>
<query_num> 13211 </query_num>
<text>   does not have ovarian cancer. Correlational knowledge is one kind of adversarial background knowledge. Integrating background knowledge into privacy quantification has been recently studied [4], [5], =-=[6]-=-, [7]. They propose different approaches (a formal language [4], [5] or ME constraints [7]) for expressing background knowledge and analyze the privacy risk when the adversary has a certain amount of  sumes different types of background knowledge, including correlational knowledge. A. Motivation Background knowledge poses significant challenges in defining privacy for the anonymized data [4], [5], =-=[6]-=-, [7]. For example, when background knowledge is present, we cannot simply say that no adversary knows any individual’s sensitive attribute value after seeing the released data, because there may exis tributions &amp; Organization In this paper, we formalize the above intuitive definition. First, we model all background knowledge that is consistent with the original data. We build on our previous work =-=[6]-=-, namely, mining background knowledge from the data to be released. Our rationale is that if certain facts or knowledge exist in the data (e.g., males cannot have ovarian cancer), they should manifest ledge is to calculate estimations of the adversary’s prior belief function Ppri, which is defined over all possible QI values in D[QI]. B. Estimating the Prior Belief Function We build on the work of =-=[6]-=- and generate background knowledge by mining the data to be released. The general rationale is that the adversary’s background knowledge about the data should be consistent with the data in T and shou tribute values. These methods provide a framework for defining and analyzing background knowledge, but they are unaware of the exact background knowledge the adversary may have. The injector approach =-=[6]-=- considers negative association rules but it does not model other types of background knowledge and does not provide an approach to analyze how an adversary can gain sensitive information from the pub   </text>
<query_num> 13212 </query_num>
<text>   e attribute in each group. The t-closeness model [3] proposes that the distribution of the sensitive attribute in the overall table should also be public information. Recently, the σ-presence measure =-=[28]-=- observes that knowing an individual is in the database poses privacy risks. The m-confidentiality model [12] recognizes that knowledge of the mechanism or algorithm of anonymization can leak extra se   </text>
<query_num> 13213 </query_num>
<text>   ide an approach to analyze how an adversary can gain sensitive information from the published data. Anonymization Techniques. Most anonymization solutions adopt generalization [13], [14], [32], [22], =-=[33]-=-, [34], [35], [36], [21], [37] and bucketization [15], [16], [4], [38], [39]. Other anonymization techniques include clustering [40], [41], [42], [43], space mapping [44], spatial indexing [45], margi   </text>
<query_num> 13214 </query_num>
<text>   l algorithms are implemented in Java and the experiments are performed on a 3.4GHZ Pentium 4 machine with 2.0GB of RAM. Given the dataset, we use the variations of Mondrian multidimensional algorithm =-=[21]-=- to compute the anonymized tables using different privacy models: (1) distinct ℓ-diversity; (2) probabilistic ℓ-diversity; (3) t-closeness; and (4) (B, t)privacy. The variations of Mondrian use the or ze how an adversary can gain sensitive information from the published data. Anonymization Techniques. Most anonymization solutions adopt generalization [13], [14], [32], [22], [33], [34], [35], [36], =-=[21]-=-, [37] and bucketization [15], [16], [4], [38], [39]. Other anonymization techniques include clustering [40], [41], [42], [43], space mapping [44], spatial indexing [45], marginals releasing [46], and   </text>
<query_num> 13215 </query_num>
<text>   lem and even known estimation algorithms have too high a complexity to be practical. To overcome the complexity of exact inference, we generalize the approximation technique used by Lakshmanan et al. =-=[9]-=- and propose an approximate inference method called Ω-estimate. We show that Ω-estimate is practical and accurate through experimental evaluation. Thirdly, we propose a novel privacy model called (B,  itive values in the group. Each edge from tuple tj to sensitive value si is associated with the probability P(si|tj). Our approach is a generalized version of the O-estimate used by Lakshmanan et al. =-=[9]-=-, where they estimate the number of correct mappings between original items and anonymized items. In that context, a item either can be linked to an anonymized item or cannot be linked to the anonymiz ground knowledge in other contexts. Yang and Li [51] studied the problem of information disclosure in XML publishing when the adversary has knowledge of functional dependencies about the XML data. In =-=[9]-=-, Lakshmanan et al. studied the problem of protecting the true identities of data objects in the context of frequent set mining when an adversary has partial information of the items in the domain. VI   </text>
<query_num> 13216 </query_num>
<text>   lete problem. A number of approximation algorithms have been proposed to compute the permanent of a matrix. The state of the art is the polynomial-time randomized approximation algorithm presented in =-=[17]-=-. However, the time complexity is of order of O(k 22 ). It is thus not feasible for the general formula to work for a large k. In the following, we turn to approximation algorithms for computing the p   </text>
<query_num> 13217 </query_num>
<text>   nference is hard to compute, we propose an approximation inference method called Ω-estimate. A. Anonymization Techniques Two widely-studied data anonymization techniques are generalization [13], [1], =-=[14]-=- and bucketization [15], [4], [16]. In generalization, quasi-identifier values are replaced with values that are less-specific but semantically consistent. Bucketization, on the other hand, first part edge integration, and (3) anonymization techniques. We then examine several research works that have studied background knowledge in other contexts. General Privacy Models. The k-anonymity model [1], =-=[14]-=- assumes that the adversary has access to some publiclyavailable databases (e.g., a vote registration list) and the adversary knows who is and who is not in the table. A few subsequent works [2], [26]  and does not provide an approach to analyze how an adversary can gain sensitive information from the published data. Anonymization Techniques. Most anonymization solutions adopt generalization [13], =-=[14]-=-, [32], [22], [33], [34], [35], [36], [21], [37] and bucketization [15], [16], [4], [38], [39]. Other anonymization techniques include clustering [40], [41], [42], [43], space mapping [44], spatial in   </text>
<query_num> 13218 </query_num>
<text>   nonymization solutions adopt generalization [13], [14], [32], [22], [33], [34], [35], [36], [21], [37] and bucketization [15], [16], [4], [38], [39]. Other anonymization techniques include clustering =-=[40]-=-, [41], [42], [43], space mapping [44], spatial indexing [45], marginals releasing [46], and data perturbation [47], [48]. On the theoretical side, optimal k-anonymity has been proved to be NP-hard fo   </text>
<query_num> 13219 </query_num>
<text>   oes not provide an approach to analyze how an adversary can gain sensitive information from the published data. Anonymization Techniques. Most anonymization solutions adopt generalization [13], [14], =-=[32]-=-, [22], [33], [34], [35], [36], [21], [37] and bucketization [15], [16], [4], [38], [39]. Other anonymization techniques include clustering [40], [41], [42], [43], space mapping [44], spatial indexing   </text>
<query_num> 13220 </query_num>
<text>   racy in aggregate query answering. 1) General Utility Measures: We first compare data utility based on two general utility measures: Discernibility Metric (DM) [22] and Global Certainty Penalty (GCP) =-=[23]-=-. Figure 5(a) shows the DM cost while Figure 5(b) shows the GCP cost for the four anonymized tables. In both experiments, we evaluate the utility measure as a function of the privacy parameters shown   </text>
<query_num> 13221 </query_num>
<text>   re. i=1 qi However, none of the above distance measures satisfy the semantic awareness property. One distance measure that takes value semantics into consideration is the Earth Mover’s Distance (EMD) =-=[20]-=-, [3]. The EMD is based on the minimal amount of work needed to transform one distribution to another by moving distribution mass between each other. Unfortunately, EMD does not have the probability s   </text>
<query_num> 13222 </query_num>
<text>   rms of general utility measures and accuracy in aggregate query answering. 1) General Utility Measures: We first compare data utility based on two general utility measures: Discernibility Metric (DM) =-=[22]-=- and Global Certainty Penalty (GCP) [23]. Figure 5(a) shows the DM cost while Figure 5(b) shows the GCP cost for the four anonymized tables. In both experiments, we evaluate the utility measure as a f t provide an approach to analyze how an adversary can gain sensitive information from the published data. Anonymization Techniques. Most anonymization solutions adopt generalization [13], [14], [32], =-=[22]-=-, [33], [34], [35], [36], [21], [37] and bucketization [15], [16], [4], [38], [39]. Other anonymization techniques include clustering [40], [41], [42], [43], space mapping [44], spatial indexing [45],   </text>
<query_num> 13223 </query_num>
<text>   ssumes that the adversary has access to some publiclyavailable databases (e.g., a vote registration list) and the adversary knows who is and who is not in the table. A few subsequent works [2], [26], =-=[27]-=- recognize that the adversary also has knowledge of the distribution of the sensitive attribute in each group. The t-closeness model [3] proposes that the distribution of the sensitive attribute in th   </text>
<query_num> 13224 </query_num>
<text>   tion from the published data. Anonymization Techniques. Most anonymization solutions adopt generalization [13], [14], [32], [22], [33], [34], [35], [36], [21], [37] and bucketization [15], [16], [4], =-=[38]-=-, [39]. Other anonymization techniques include clustering [40], [41], [42], [43], space mapping [44], spatial indexing [45], marginals releasing [46], and data perturbation [47], [48]. On the theoreti   </text>
</top>
<top>
<paper_num> 133 </paper_num>
<paper_title>   Managing commitments in multiple concurrent negotiations.  </paper_title>
<paper_abstract>   Automated negotiation by software agents is a key enabling technology for agent mediated e-commerce. To this end, this paper considers an important class of such negotiations – namely those in which an agent engages in multiple concurrent bilateral negotiations for a good or service. In particular, we consider the situation in which a buyer agent is looking for a single service provider from a number of available ones in its environment. By bargaining simultaneously with these providers and interleaving partial agreements that it makes with them, a buyer can reach good deals in an efficient manner. However, a key problem in such encounters is managing commitments since an agent may want to make intermediate deals (so that it has a definite agreement) with other agents before it gets to finalize a deal at the end of the encounter. To do this effectively, however, the agents need to have a flexible model of commitments that they can reason about in order to determine when to commit and to decommit. This paper provides and evaluates such a commitment model and integrates it into a concurrent negotiation model.  </paper_abstract>
<query_num> 13301 </query_num>
<text>   . It has been empirically demonstrated that the latter type allows the agents to be more flexible in deliberating about their behaviors and enables them to gain a higher utility value than the former =-=[9]-=-. Consequently, we use the percentage of contract value in our model. 3 This factor is incorporated to discourage the agent from dropping its commitment towards the end of the negotiation (where it is   </text>
<query_num> 13302 </query_num>
<text>   cific future events. Thus, if these specified contingencies aries, the agents are allowed to drop their commitments [21]. However, there are a number of problems associated with this type of contract =-=[5]-=-. First, not all possible future events are known to the agents beforehand, thus, they cannot always make optimal use of contingency contracts. Second, this type of contract is useful when the number   on the value of the contract [13]. Asa result, this approach is not adopted in our work. The most advanced work in the area, and also the basis to our work, is the leveled commitment contracts (LVC) =-=[5]-=-. Our commitment manager is built upon the same basic intuition that any agent can freely decommit from a contract, for whatever reason they deem appropriate, by simply paying a decommitment fee to th   </text>
<query_num> 13303 </query_num>
<text>   e: a seller always breaks a committed deal if it is presented with a better option. Partial: if a seller finds a better option, it will break a committed deal with a percentage of probability (as per =-=[12]-=-). In this experiment, we set this percentage to be 50%, meaning that half of the time a seller finds a better deal, it will renege and half of the time it will stay with its current deal. After each  e future [13–15]. This is also the case for existing concurrent negotiation models [4,16,17]. However, this view is very limiting for the agents and it may lead to irrational and inefficient behavior =-=[12]-=-.Asa result, a number of methods have been developed to overcome this limitation. One of the first pieces of work in this area was the contract net protocol [18], where there is a possibility for a de   </text>
<query_num> 13304 </query_num>
<text>   each round, the threads report back their status to the coordi1 Given the time-constrained nature of our encounters, the types of strategy that we consider are the time-dependent family introduced in =-=[6]-=-. These can be broadly divided into three classes: the conceder strategy quickly lowers its value until it reaches its reservation (minimum acceptable) value. The linear strategy drops to its reservat max Š that satisfies x b j min 6 x a j min 6 x b j max 6 x a j max . The negotiation strategy: Each seller is assigned a random strategy selected from a predefined set of alternations (as outlined in =-=[6]-=-). This set is composed of time-dependant functions (like conceder, boulware and linear) and behaviordependant tactics (such as tit-for-tat in its various forms). The negotiation deadline: The deadlin   </text>
<query_num> 13305 </query_num>
<text>   greedier in making commitments. Our extended model is also currently being used in a number of real world applications to form and maintain coalitions in business and e-science virtual organizations =-=[22]-=- and in an internal project of BT concerned with logistics planning [23]. For the future, there are a number of ways in which our model can be improved. First, we would like to experiment with differe   </text>
<query_num> 13306 </query_num>
<text>   is made in a negotiation, it is binding on all participants. Neither party can back out no matter what happens in the future [13–15]. This is also the case for existing concurrent negotiation models =-=[4,16,17]-=-. However, this view is very limiting for the agents and it may lead to irrational and inefficient behavior [12].Asa result, a number of methods have been developed to overcome this limitation. One of   </text>
<query_num> 13307 </query_num>
<text>   n, non}. This information is represented as a probability distribution over the agent types, which may be based on past experiences, obtained from a trusted third party, or from a system of referrals =-=[8]-=-. If no such information is available, all agents are assumed to have a uniform distribution. There are two further sources of information that aid the coordinatorÕs decision making: the percentage of Table 3, other control variables are selected as per [11]. Specifically, the number of seller agents (n) is set in the range of [1, 30] and the number of negotiation issues (m) is set in the range of =-=[1, 8]-=-. An agent aÕs preference for issue j is represented by the tuple fx a j min ; x a j max ; w a j g. The tuple xa j min ; x a j max Š Table 3 The independent variables Variables Descriptions Values q0   </text>
<query_num> 13308 </query_num>
<text>   t mind losing their effort (even without any form of compensation). In a similar fashion, the role of commitment for cooperative agents was examined in the context of automated scheduling of meetings =-=[19]-=-. In e-commerce settings, however, these models are inappropriate because the agents are not always cooperative and they seek to maximize their individual gains. For self-interested agents, contingenc   </text>
<query_num> 13309 </query_num>
<text>   troduction Electronic Commerce Research and Applications 4 (2005) 362–376 Automated negotiation is a key form of interaction in agent-based systems and such negotiations exist in many different forms =-=[1]-=-. In this paper, we focus on one such form, namely one-to-many negotiations in service-oriented contexts. Here, a * Corresponding author. E-mail addresses: duong.nguyen@bt.com (T.D. Nguyen), nrj@ecs.s es are listed in Table 4. Apart from the control variables described in Table 3, other control variables are selected as per [11]. Specifically, the number of seller agents (n) is set in the range of =-=[1, 30]-=- and the number of negotiation issues (m) is set in the range of [1, 8]. An agent aÕs preference for issue j is represented by the tuple fx a j min ; x a j max ; w a j g. The tuple xa j min ; x a j m  independent variables Variables Descriptions Values q0 The initial penalty fee [5,100] qmax The final penalty fee (qmax P q0) [5,100] s The l threshold [0,1.5] x The number of concurrent commitments =-=[1,4]-=- Table 4 The dependent variables Variables Descriptions U The utility value of the final agreement N The number of successful negotiations D The number of decommitments made by buyersT.D. Nguyen, N.R. t to a maximum of two contracts. Fig. 9 shows the number of agreements obtained by the buyer at the end of the encounter when it varies the number of commitments it can hold at any one time (here x 2 =-=[1, 4]-=-). As can be seen, when holding more than one commitment, the buyer increases its chance of reaching an agreement when dealing with non-loyal sellers. In particular, when x is increased from 1 to 2, t   </text>
<query_num> 13310 </query_num>
<text>   uation, the independent variables are given in Table 3 and the dependent ones are listed in Table 4. Apart from the control variables described in Table 3, other control variables are selected as per =-=[11]-=-. Specifically, the number of seller agents (n) is set in the range of [1, 30] and the number of negotiation issues (m) is set in the range of [1, 8]. An agent aÕs preference for issue j is represente   </text>
</top>
<top>
<paper_num> 134 </paper_num>
<paper_title>   Resilient multicast using overlays.  </paper_title>
<paper_abstract>   We introduce PRM (Probabilistic Resilient Multicast): a multicast data recovery scheme that improves data de-livery ratios while maintaining low end-to-end latencies. PRM has both a proactive and a reactive component; in this paper we describe how PRM can be used to improve the performance of application-layer multicast protocols, espe-cially when there are high packet losses and host failures. Further, using analytic techniques, we show that PRM can guarantee arbitrarily high data delivery ratios and low latency bounds. As a detailed case study, we show how PRM can be applied to the NICE application-layer multicast protocol. We present detailed simulations of the PRM-enhanced NICE protocol for 10,000 node Internet-like topologies. Simulations show that PRM achieves a high delivery ratio ( ¢ 97%) with a low latency bound (600 ms) for environments with high end-to-end network losses (1-5%) and high topology change rates (5 changes per second) while incurring very low overheads ( £ 5%). I.  </paper_abstract>
<query_num> 13401 </query_num>
<text>   K A large number of research proposals have addressed reliable delivery for multicast data, most notably in the context of network-layer multicast. A comparative survey of these protocols is given in =-=[12]-=- and [22]. In SRM [7] receivers send NAKs to the source to indicate missing data packets. Each such NAK is multicast to the entire group and is used to suppress NAKs from other receivers that did not   </text>
<query_num> 13402 </query_num>
<text>   ach to provide improved reliability performance for multicast data. There exists some well-known forward error correcting code based approaches that are also proactive in nature. For example, Huitema =-=[10]-=- had proposed the use of packet level FECs for reliable multicast. Nonnenmacher et. al. [16] studied and demonstrated that additional benefits can be achieved when an FEC-based technique is combined w  ACKs with local scope Parity-based [16] Network multicast Reactive NAKs and Moderate Moderate (APES [21]) (and directed subcast) FEC-based repairs FEC-based Network multicast Proactive FECs High Low =-=[10]-=-, [16], [3], [15] or App-layer multicast 2 PRM App-layer multicast Proactive randomized forwarding Low Low and reactive NAKs TABLE III COMPARISON OF DIFFERENT RELIABILITY/RESILIENCE MECHANISMS FOR MUL   </text>
<query_num> 13403 </query_num>
<text>   alable because the acknowledgments are aggregated along the tree in a bottom-up fashion and also allows local recovery and repair of data losses. Protocols like RMTP [18], TMTP [23], STORM [20], LVMR =-=[14]-=- and Lorax [13] construct this structure using TTL-scoped network-layer multicast as a primitive. In contrast, LMS [17] uses an additional mechanism, called directed subcast, to construct its data rec multicast Reactive NAKs Low Moderate on ack tree LMS [17] Network multicast Reactive NAKs Low Moderate and directed subcast on ack-tree RMTP [18] Network multicast Reactive/periodic Low Moderate LVMR =-=[14]-=- ACKs with local scope TMTP [23] Network multicast Reactive NAKs and Low Moderate periodic ACKs with local scope Parity-based [16] Network multicast Reactive NAKs and Moderate Moderate (APES [21]) (an   </text>
<query_num> 13404 </query_num>
<text>   and repair of data losses. Protocols like RMTP [18], TMTP [23], STORM [20], LVMR [14] and Lorax [13] construct this structure using TTL-scoped network-layer multicast as a primitive. In contrast, LMS =-=[17]-=- uses an additional mechanism, called directed subcast, to construct its data recovery structure. Our work differs from of all these above approaches in two key aspects. First, unlike all these protoc eads Recovery latency SRM [7] Network multicast Reactive NAKs High (for high High with global scope network losses) STORM [20], Lorax [13] Network multicast Reactive NAKs Low Moderate on ack tree LMS =-=[17]-=- Network multicast Reactive NAKs Low Moderate and directed subcast on ack-tree RMTP [18] Network multicast Reactive/periodic Low Moderate LVMR [14] ACKs with local scope TMTP [23] Network multicast Re   </text>
<query_num> 13405 </query_num>
<text>   delivery resilient multicast. In this paper we describe PRM in the context of overlay-based multicast. The basic idea of multicast using overlays (also known as application-layer multicast) [6], [8], =-=[2]-=-, [24], [5], [19], [11] is shown in Figure 1. Unlike native multicast where data packets are replicated at routers inside the network, in application-layer multicast data packets are replicated at end high resilience guarantees, both in theory and practice. PRM can be used to significantly augment the data delivery ratios of any application-layer multicast protocol (e.g. Narada [6], Yoid [8], NICE =-=[2]-=-, HMTP [24], Scribe [5], CAN-multicast [19], Delaunay Triangulation-based [11]) while maintaining low latency bounds. B. Contributions The contributions of this paper are three-fold: We propose a simp pproach is sufficient to achieve high data delivery ¤ rates within low latency bounds. We demonstrate how our proposed scheme can be used with an existing application-layer multicast protocol ¤ (NICE =-=[2]-=-) to provide a low overhead, low latency and high delivery ratio multicast technique for realistic applications and scenarios. 2s0 1 B A C D G H J K L M N P E F Q A C E F T D T G H J K L M N P Fig. 2. ol like Narada [6], each node maintains state information about all other nodes. Therefore, no additional discovery of nodes is necessary in this case. For some other protocols like Yoid [8] and NICE =-=[2]-=- overlay nodes maintain information of only a small subset of other nodes in the topology. Therefore we implement a node discovery mechanism, using a random-walk on the overlay tree. A similar techniq ions only model random failures of overlay nodes and random packet losses on overlay links for an overlay tree. We have also implemented PRM over a specific application-layer multicast protocol, NICE =-=[2]-=-. We present detailed performance studies which includes the consequent interactions between PRM and NICE in Section IV. A. Analysis of Simplified PRM For the tractability of the analysis we consider   </text>
<query_num> 13406 </query_num>
<text>   ditional benefits can be achieved when an FEC-based technique is combined with automatic retransmission requests. APES uses a related approach for data recovery [21]. Digital Fountain [3] and RPB/RBS =-=[15]-=- are two other efficient FEC-based approaches that provide significantly improved performance. All these FEC based approaches can recover from network losses. However, they alone are not sufficient fo scope Parity-based [16] Network multicast Reactive NAKs and Moderate Moderate (APES [21]) (and directed subcast) FEC-based repairs FEC-based Network multicast Proactive FECs High Low [10], [16], [3], =-=[15]-=- or App-layer multicast 2 PRM App-layer multicast Proactive randomized forwarding Low Low and reactive NAKs TABLE III COMPARISON OF DIFFERENT RELIABILITY/RESILIENCE MECHANISMS FOR MULTICAST DATA. rout   </text>
<query_num> 13407 </query_num>
<text>   ges and is responsible for fast data recovery in PRM under high failure rates of overlay nodes. Existing approaches for resilient and reliable multicast use either reactive retransmissions (e.g. RMTP =-=[18]-=-, STORM [20] Lorax [13]) or proactive error correction codes (e.g. Digital Fountain [3]) and can only recover from packet losses on the overlay links. Therefore the proactive randomized forwarding is  ers. An overlay node can detect missing data using gaps in the sequence numbers. This information is used to trigger NAK-based retransmissions. This technique has been applied for loss repair in RMTP =-=[18]-=-. In our implementation each overlay node, � , piggybacks a bit-mask with each forwarded data packet indicating which of the prior sequence numbers it has correctly received. The recipient of the data  as the root. This structure is scalable because the acknowledgments are aggregated along the tree in a bottom-up fashion and also allows local recovery and repair of data losses. Protocols like RMTP =-=[18]-=-, TMTP [23], STORM [20], LVMR [14] and Lorax [13] construct this structure using TTL-scoped network-layer multicast as a primitive. In contrast, LMS [17] uses an additional mechanism, called directed  lobal scope network losses) STORM [20], Lorax [13] Network multicast Reactive NAKs Low Moderate on ack tree LMS [17] Network multicast Reactive NAKs Low Moderate and directed subcast on ack-tree RMTP =-=[18]-=- Network multicast Reactive/periodic Low Moderate LVMR [14] ACKs with local scope TMTP [23] Network multicast Reactive NAKs and Low Moderate periodic ACKs with local scope Parity-based [16] Network mu   </text>
<query_num> 13408 </query_num>
<text>   l of data delivery resilient multicast. In this paper we describe PRM in the context of overlay-based multicast. The basic idea of multicast using overlays (also known as application-layer multicast) =-=[6]-=-, [8], [2], [24], [5], [19], [11] is shown in Figure 1. Unlike native multicast where data packets are replicated at routers inside the network, in application-layer multicast data packets are replica  path for the two cases. are more significant than regular packet losses in the network and may cause data outage in the order of tens of seconds (e.g. the Narada application-layer multicast protocol =-=[6]-=- sets default timeouts between 30-60 seconds). A. Our Approach PRM uses two simple techniques: A proactive component called Randomized forwarding in which each overlay node chooses a constant number o  techniques provide high resilience guarantees, both in theory and practice. PRM can be used to significantly augment the data delivery ratios of any application-layer multicast protocol (e.g. Narada =-=[6]-=-, Yoid [8], NICE [2], HMTP [24], Scribe [5], CAN-multicast [19], Delaunay Triangulation-based [11]) while maintaining low latency bounds. B. Contributions The contributions of this paper are three-fol erlay node periodically discovers a set of random other nodes on the overlay and evaluates the number of losses that it shares with these random nodes. In an overlay construction protocol like Narada =-=[6]-=-, each node maintains state information about all other nodes. Therefore, no additional discovery of nodes is necessary in this case. For some other protocols like Yoid [8] and NICE [2] overlay nodes  50 60 Maximum Gap (in second) Fig. 11. Cumulative distribution of the maximum time gap over which an overlay node lost all data packets. best-effort application-layer multicast protocols, e.g. Narada =-=[6]-=-. Therefore, we believe that PRM-enhancements can significantly augment the data delivery ratios of all such protocols. An FEC-based scheme is typically able to recover from all network losses. Howeve   </text>
<query_num> 13409 </query_num>
<text>   nown forward error correcting code based approaches that are also proactive in nature. For example, Huitema [10] had proposed the use of packet level FECs for reliable multicast. Nonnenmacher et. al. =-=[16]-=- studied and demonstrated that additional benefits can be achieved when an FEC-based technique is combined with automatic retransmission requests. APES uses a related approach for data recovery [21].  -tree RMTP [18] Network multicast Reactive/periodic Low Moderate LVMR [14] ACKs with local scope TMTP [23] Network multicast Reactive NAKs and Low Moderate periodic ACKs with local scope Parity-based =-=[16]-=- Network multicast Reactive NAKs and Moderate Moderate (APES [21]) (and directed subcast) FEC-based repairs FEC-based Network multicast Proactive FECs High Low [10], [16], [3], [15] or App-layer multi   </text>
<query_num> 13410 </query_num>
<text>   nt multicast. In this paper we describe PRM in the context of overlay-based multicast. The basic idea of multicast using overlays (also known as application-layer multicast) [6], [8], [2], [24], [5], =-=[19]-=-, [11] is shown in Figure 1. Unlike native multicast where data packets are replicated at routers inside the network, in application-layer multicast data packets are replicated at end hosts. Logically  and practice. PRM can be used to significantly augment the data delivery ratios of any application-layer multicast protocol (e.g. Narada [6], Yoid [8], NICE [2], HMTP [24], Scribe [5], CAN-multicast =-=[19]-=-, Delaunay Triangulation-based [11]) while maintaining low latency bounds. B. Contributions The contributions of this paper are three-fold: We propose a simple, low-overhead scheme for resilient multi   </text>
<query_num> 13411 </query_num>
<text>   search proposals have addressed reliable delivery for multicast data, most notably in the context of network-layer multicast. A comparative survey of these protocols is given in [12] and [22]. In SRM =-=[7]-=- receivers send NAKs to the source to indicate missing data packets. Each such NAK is multicast to the entire group and is used to suppress NAKs from other receivers that did not get the same packet.  delivery when overlays are used. Overlay nodes are processes on regular end-hosts and are more prone to failures than network 15sScheme Data delivery Recovery mechanism Overheads Recovery latency SRM =-=[7]-=- Network multicast Reactive NAKs High (for high High with global scope network losses) STORM [20], Lorax [13] Network multicast Reactive NAKs Low Moderate on ack tree LMS [17] Network multicast Reacti   </text>
</top>
<top>
<paper_num> 135 </paper_num>
<paper_title>   Spam Filtering Using Inexact String Matching in Explicit Feature Space with On-Line Linear Classifiers.  </paper_title>
<paper_abstract>   Contemporary spammers commonly seek to defeat statistical spam filters through the use of word obfuscation. Such methods include character level substitutions, repetitions, and insertions to reduce the effectiveness of word-based features. We present an efficient method for combating obfuscation through the use of inexact string matching kernels, which were first developed to measure similarity among mutating genes in computational biology. Our system avoids the high classification costs associated with these kernel methods by working in an explicit feature space, and employs the Perceptron Algorithm using Margins for fast on-line training. No prior domain knowledge was incorporated into this system. We report strong experimental results on the TREC 2006 spam data sets and on other publicly available spam data, including near-perfect performance on the TREC 2006 Chinese spam data set. These results invite further exploration of the use of inexact string matching for spam filtering. 1.  </paper_abstract>
<query_num> 13501 </query_num>
<text>   The Perceptron Algorithm with Margins (PAM) (=-=Krauth and Mézard, 1987; Li et al., 2002-=-) attempts to establish such a margin, τ, during the training process. Following work on support vector machines (=-=Boser et al., 1992-=-) one may expect that providing the perceptron with higher margin will add to the stability and accuracy of the hypothesis produced (=-=Cristianini and Shawe-Taylor, 2000-=-). To establish the margin, inste   </text>
<query_num> 13502 </query_num>
<text>   ation. Finally, we use an on-line linear classifier, the Perceptron Algorithm with Margins (PAM) (=-=Krauth and Mézard, 1987; Li et al., 2002-=-), which has fast training time and good resistance to noise (=-=Khardon and Wachman, 2005-=-). The remainder of this paper provides details on the methods of efficient inexact string matching (Section 2) and the PAM classifier (Section 3). Experimental results in Section 4 show that this app ort vector machines. In this competition, we choose to classify the data with the PAM classifier (=-=Krauth and Mézard, 1987; Li et al., 2002-=-), which learns a linear classifier with tolerance for noise (=-=Khardon and Wachman, 2005-=-).sGiven: set of examples and their labels Z = ((x1,y1),... ,(xm,ym)) ∈ (R n × {−1,1}) m , τ Initialize w := 0 n for every (xj,yj) ∈ Z do: if yj(〈w,xj〉) &amp;lt; τ w := w + ηyjxj done Figure 3: The On-Line P   </text>
<query_num> 13503 </query_num>
<text>   ing matching features into an explicit feature space, enabling fast classification. Finally, we use an on-line linear classifier, the Perceptron Algorithm with Margins (PAM) (=-=Krauth and Mézard, 1987; Li et al., 2002-=-), which has fast training time and good resistance to noise (=-=Khardon and Wachman, 2005-=-). The remainder of this paper provides details on the methods of efficient inexact string matching (Section 2) a ification methods at our disposal, including Naive Bayes classifiers and support vector machines. In this competition, we choose to classify the data with the PAM classifier (=-=Krauth and Mézard, 1987; Li et al., 2002-=-), which learns a linear classifier with tolerance for noise (=-=Khardon and Wachman, 2005-=-).sGiven: set of examples and their labels Z = ((x1,y1),... ,(xm,ym)) ∈ (R n × {−1,1}) m , τ Initialize w := 0 n  ith Margins (PAM) The classical perceptron attempts to separate the data but has no guarantees on the separation margin obtained. The Perceptron Algorithm with Margins (PAM) (=-=Krauth and Mézard, 1987; Li et al., 2002-=-) attempts to establish such a margin, τ, during the training process. Following work on support vector machines (=-=Boser et al., 1992-=-) one may expect that providing the perceptron with higher margin wi classifier makes a mistake, PAM also updates on xj if yj(〈xj,w〉) &amp;lt; τ. When the data are linearly separable, the margin of the classifier produced by PAM can be lower-bounded (=-=Krauth and Mézard, 1987; Li et al., 2002-=-). The algorithm is summarized in Figure 3. It is important to select a reasonable value for τ. If τ is too large, the algorithm will not be able to find a stable hypothesis until the norm of w grows   </text>
<query_num> 13504 </query_num>
<text>   of times that the k-mer a appears in string s, with overlaps allowed. In spam filtering, there have been several results showing that binary features may be more effective than count-based features (=-=Drucker et al., 1999; Metsis et al., 2006-=-), perhaps because binary features offer greater resistance to the good word attack (=-=Wittel and Wu, 2004-=-). A stringsmay be represented as a binary k-mer vector x ∈ X, by mapping ∀ d classification. Second, strong performance has been achieved by other linear classifiers on spam such as a variant of logistic regression (=-=Goodman and Yin, 2006-=-) and linear support vector machines (=-=Drucker et al., 1999; Rios and Zha, 2004-=-); since Khardon and Wachman (2005) demonstrate experimentally that PAM is competitive with SVM on many data sets, it is reasonable to expect PAM to perform well on spam. Third, it   </text>
<query_num> 13505 </query_num>
<text>   ord-level features (=-=Graham, 2002; Metsis et al., 2006-=-), a combination that has proven efficient and reasonably effective. Recently, character-level features, such as those used by the PPM classifier (=-=Bratko and Filipic, 2005-=-), have improved performance. Character level features offer resistance to standard spam attacks such as obfuscation, the practice of intentionally misspelling words to defeat word-based spam filters  . An (i,j) k-mer mapping creates a space in which there is a dimension for all possible k-mers, where i ≤ k ≤ j. Indeed, the best spam filter from the 2005 TREC Spam Filtering Track was a PPM-filter (=-=Bratko and Filipic, 2005-=-), which implicitly uses a feature space of (i,j) k-mers (=-=Sculley and Brodley, 2006-=-). 2.2 Wildcards and Gaps In computational biology, it has been found that k-mers alone are not expressive enough to   </text>
<query_num> 13506 </query_num>
<text>   pam. Third, it has been shown that while generative models such as Naive Bayes may have steeper (faster) learning curves, discriminative models such as linear classifiers have lower asymptotic error (=-=Ng and Jordan, 2002-=-). In the remainder of this section, we will review the basic Perceptron Algorithm, and explore relevant features of the PAM variant that makes it an attractive choice for on-line spam filtering. 3.1   </text>
<query_num> 13507 </query_num>
<text>   several reasons. First, it offers fast on-line training and classification. Second, strong performance has been achieved by other linear classifiers on spam such as a variant of logistic regression (=-=Goodman and Yin, 2006-=-) and linear support vector machines (=-=Drucker et al., 1999; Rios and Zha, 2004-=-); since Khardon and Wachman (2005) demonstrate experimentally that PAM is competitive with SVM on many data sets, it is r   </text>
<query_num> 13508 </query_num>
<text>   tor. 1 An upper-bound on the number of mistakes committed by the perceptron algorithm can be shown both when the data are linearly separable (=-=Novikoff, 1962-=-) and when they are not linearly separable (=-=Freund and Schapire, 1999-=-). 3.2 Perceptron Algorithm with Margins (PAM) The classical perceptron attempts to separate the data but has no guarantees on the separation margin obtained. The Perceptron Algorithm with Margins (PA   </text>
</top>
<top>
<paper_num> 136 </paper_num>
<paper_title>   Fast Mesh interpolation and Mesh Decomposition with Applications.  </paper_title>
<paper_abstract>   Abstract A fast iterative method for constructing a smooth subdivision surface that interpolates the vertices of an arbitrary mesh is presented. The construction is done by iteratively adjusting vertices of the given mesh locally until control mesh of the required interpolating surface is reached. The new interpolation method has not only the simplicity of a local method, but also the capability of a global method in faithfully resembling the shape of a given mesh. The new method does not require solving a linear system, hence it can handle meshes with large number of vertices. Furthermore, the new method is fast and does not require a fairing step in the construction process because the iterative process converges to a unique solution at an exponential rate. Another important result of this work is, with the new iterative process, each mesh (surface) can be expanded as an infinite series of meshes (surfaces) which carry high and low frequency information of the given model. This mesh expansion scheme provides us with new approaches to some classic applications in computer graphics such as texture mapping, de-noising/smoothing/sharpening, and morphing. These new approaches are demonstrated in this paper and test results are included. 1  </paper_abstract>
<query_num> 13601 </query_num>
<text>   . 6 De-noising With the proliferation of 3D scanning devices, fairing, smoothing and denoising of noisy meshes have become more and more important. Several important works have been done in this area =-=[19, 17, 16, 18]-=-. In this section, we present a new denoising technique which is a straightforward application of our interpolation formula. Our purpose here is simply to show the versatility of our interpolation met   </text>
<query_num> 13602 </query_num>
<text>   and Q, the task here is to find a smooth transition from Q to M. We assume M and Q have the same topology. If this is not the case, simply resample them using one of the resampling techniques such as =-=[20]-=-. According to eq. (6), we have Q = M + ∞� (E − A) i A(Q − M). i=0 (Q − M) can be regarded as the difference of the two models, hence our goal is to transform this difference smoothly so that it appro   </text>
<query_num> 13603 </query_num>
<text>   cult to handle meshes with large number of vertices. There are also techniques that produce subdivision surfaces to interpolate given curves or surfaces that near- (or quasi-)interpolate given meshes =-=[7]-=-. But those techniques are either of different natures or of different concerns, hence will not be discussed here. As far as we know there does not have a subdivision surface based interpolation techn   </text>
<query_num> 13604 </query_num>
<text>   ications. Powerful interpolation techniques using subdivision surfaces as a representation scheme certainly are needed for subdivision surface based applications. Traditional interpolation techniques =-=[3, 4, 5, 6, 8, 9, 10, 15]-=- are mainly aimed at one purpose: shape reconstruction. The iterative interpolation technique proposed in this paper also provides us with a means to look at several classic applications from a differ ctions. Test results and concluding remarks are presented in the last two sections. 2 Previous Work There are two major methods for interpolating a given mesh with a subdivision surface: local method =-=[3, 5, 6, 9, 10, 15]-=- or global method [4, 8]. In a local method, new vertices are defined as affine combinations of nearby vertices. Interpolating subdivision is the most well-known local interpolation method. In this ca   </text>
<query_num> 13605 </query_num>
<text>   ications. Powerful interpolation techniques using subdivision surfaces as a representation scheme certainly are needed for subdivision surface based applications. Traditional interpolation techniques =-=[3, 4, 5, 6, 8, 9, 10, 15]-=- are mainly aimed at one purpose: shape reconstruction. The iterative interpolation technique proposed in this paper also provides us with a means to look at several classic applications from a differ ctions. Test results and concluding remarks are presented in the last two sections. 2 Previous Work There are two major methods for interpolating a given mesh with a subdivision surface: local method =-=[3, 5, 6, 9, 10, 15]-=- or global method [4, 8]. In a local method, new vertices are defined as affine combinations of nearby vertices. Interpolating subdivision is the most well-known local interpolation method. In this ca  are not so dense, the effect of undesired artifacts becomes obvious on the resulting interpolating surfaces (see Figure 1 for an example where the Butterfly scheme [3], the improved Butterfly scheme =-=[15]-=- and the technique proposed in this paper are compared on a given mesh with five vertices). The global method, usually needs to build a global linear system with some constraints [4]. The solution to   </text>
<query_num> 13606 </query_num>
<text>   ications. Powerful interpolation techniques using subdivision surfaces as a representation scheme certainly are needed for subdivision surface based applications. Traditional interpolation techniques =-=[3, 4, 5, 6, 8, 9, 10, 15]-=- are mainly aimed at one purpose: shape reconstruction. The iterative interpolation technique proposed in this paper also provides us with a means to look at several classic applications from a differ ctions. Test results and concluding remarks are presented in the last two sections. 2 Previous Work There are two major methods for interpolating a given mesh with a subdivision surface: local method =-=[3, 5, 6, 9, 10, 15]-=- or global method [4, 8]. In a local method, new vertices are defined as affine combinations of nearby vertices. Interpolating subdivision is the most well-known local interpolation method. In this ca  to see. But when the mesh vertices are not so dense, the effect of undesired artifacts becomes obvious on the resulting interpolating surfaces (see Figure 1 for an example where the Butterfly scheme =-=[3]-=-, the improved Butterfly scheme [15] and the technique proposed in this paper are compared on a given mesh with five vertices). The global method, usually needs to build a global linear system with so   </text>
<query_num> 13607 </query_num>
<text>   ications. Powerful interpolation techniques using subdivision surfaces as a representation scheme certainly are needed for subdivision surface based applications. Traditional interpolation techniques =-=[3, 4, 5, 6, 8, 9, 10, 15]-=- are mainly aimed at one purpose: shape reconstruction. The iterative interpolation technique proposed in this paper also provides us with a means to look at several classic applications from a differ marks are presented in the last two sections. 2 Previous Work There are two major methods for interpolating a given mesh with a subdivision surface: local method [3, 5, 6, 9, 10, 15] or global method =-=[4, 8]-=-. In a local method, new vertices are defined as affine combinations of nearby vertices. Interpolating subdivision is the most well-known local interpolation method. In this case, a subdivision scheme Butterfly scheme [15] and the technique proposed in this paper are compared on a given mesh with five vertices). The global method, usually needs to build a global linear system with some constraints =-=[4]-=-. The solution to the global linear system is a control mesh whose limit surface interpolates the vertices of the given mesh. This approach usually requires some fairness constraints [4] in the interp e construction of the interpolation surface because there exists only one such surface that interpolates the given mesh. Traditionally, people tried to directly find A −1 M by solving a linear system =-=[4, 8]-=-. It is difficult to deal with meshes with large number of vertices that way. With the new technique, P can be constructed not by solving a linear system, but by iteratively applying eq. (7) until som ntioned above, P is the only mesh that has the same topology as M and whose limit surface interpolates M. Therefore there is no need for a fairing process either. Traditional interpolation techniques =-=[4, 8]-=- need a fairing process because extra vertices are added in the interpolation process. These extra vertices, with possibly improperly assigned positions, lead to undulations in the interpolating surfa   </text>
<query_num> 13608 </query_num>
<text>   nd very easy to implement. Subdivision surfaces are the representation scheme used in this interpolation technique. Paper ID: 0350 Category: Research Paper Paper ID: 0350 Page: 1 Subdivision surfaces =-=[1, 11, 14]-=- have become popular recently because of their capability in modeling/representing any complex shape with only one surface [2]. Subdivision surfaces cover both parametric forms [12, 13] and discrete f   </text>
<query_num> 13609 </query_num>
<text>   on surfaces [1, 11, 14] have become popular recently because of their capability in modeling/representing any complex shape with only one surface [2]. Subdivision surfaces cover both parametric forms =-=[12, 13]-=- and discrete forms. Parametric forms are good for design and representation and discrete forms are good for machining and tessellation (including FE mesh generation). Therefore we have a representati ined. By applying our fast iterative interpolation technique to Ni, we get I(Ni). Because all I(Ni) are smooth subdivision surfaces, they can be added together (through subdivision or parametrization =-=[12]-=-) to form a new surface. Let N = � n i=0 I(Ni). After normalization, Nx, which is the x component of N, contains noises for the subdivision surface S(G0) everywhere. For a given list of colors, three   </text>
</top>
<top>
<paper_num> 137 </paper_num>
<paper_title>   A Comprehensive Survey of Multiagent Reinforcement Learning.  </paper_title>
<paper_abstract>   Abstract—Multiagent systems are rapidly finding applications in a variety of domains, including robotics, distributed control, telecommunications, and economics. The complexity of many tasks arising in these domains makes them difficult to solve with preprogrammed agent behaviors. The agents must, instead, discover a solution on their own, using learning. A significant part of the research on multiagent learning concerns reinforcement learning techniques. This paper provides a comprehensive survey of multiagent reinforcement learning (MARL). A central issue in the field is the formal statement of the multiagent learning goal. Different viewpoints on this issue have led to the proposal of many different goals, among which two focal points can be distinguished: stability of the agents ’ learning dynamics, and adaptation to the changing behavior of the other agents. The MARL algorithms described in the literature aim—either explicitly or implicitly—at one of these two goals or at a combination of both, in a fully cooperative, fully competitive, or more general setting. A representative selection of these algorithms is discussed in detail in this paper, together with the specific issues that arise in each category. Additionally, the benefits and challenges of MARL are described along with some of the problem domains where the MARL techniques have been applied. Finally, an outlook for the field is provided. Index Terms—Distributed control, game theory, multiagent systems, reinforcement learning. I.  </paper_abstract>
<query_num> 13701 </query_num>
<text>   ) Managers of resources, as in [5]. Each agent manages one resource, and the agents learn how to best service requests in order to optimize a given performance measure. 2) Clients of resources, as in =-=[107]-=-. The agents learn how to best select resources such that a given performance measure is optimized. A popular resource management domain is network routing [108]–[110]. Other examples include elevator   </text>
<query_num> 13702 </query_num>
<text>   , and are therefore, applicable to mixed SGs, although without any guarantees for success. 1) Single-Agent RL: Single-agent RL algorithms like Qlearning can be directly applied to the multiagent case =-=[69]-=-. However, the nonstationarity of the MARL problem invalidates most of the single-agent RL theoretical guarantees. Despite its limitations, this approach has found a significant number of applications   </text>
<query_num> 13703 </query_num>
<text>   , leading to a high degree of scalability. Several existing MARL algorithms often require some additional preconditions to theoretically guarantee and to fully exploit the potential of these benefits =-=[41]-=-, [53]. Relaxing these conditions and further improving the performance of the various MARL algorithms in this context is an active field of study.160 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETI Then, if all the agents learn the common Q-function in parallel with (7), they can safely use (8) to select these optimal joint actions and maximize their return. The Distributed Q-learning algorithm =-=[41]-=- solves the cooperative task without assuming coordination and with limited computation (its complexity is similar to that of single-agent Qlearning). However, the algorithm only works in deterministi   </text>
<query_num> 13704 </query_num>
<text>   13] given in Section IV. Some methods in the area of direct policy search use gradient update rules that guarantee convergence in specific classes of static games: Infinitesimal Gradient Ascent (IGA) =-=[66]-=-, Winor-Learn-Fast IGA (WoLF-IGA) [13], Generalized IGA (GIGA) [67], and GIGA-WoLF [57]. For instance, IGA and WoLF-IGA work in two-agent, two-action games, and use similar gradient update rules ⎧ ⎪⎨   </text>
<query_num> 13705 </query_num>
<text>   169 rewarding promising behaviors rather than only the achievement of the goal, could be provided to the agents [70], [86]. Humans or skilled agents could teach unskilled agents how to solve the task =-=[126]-=-. Shaping is a technique whereby the learning process starts by presenting the agents with simpler tasks, and progressively moves toward complex ones [127]. Preprogrammed reflex behaviors could be bui   </text>
<query_num> 13706 </query_num>
<text>   Benefits of MARL A speedup of MARL can be realized thanks to parallel computation when the agents exploit the decentralized structure of the task. This direction has been investigated in, e.g., [45]–=-=[50]-=-. Experience sharing can help agents with similar tasks to learn faster and better. For instance, agents can exchange information using communication [51], skilled agents may serve as teachers for the L is put to use [112], [113], [115]–[119]. A complementary avenue for improving scalability is the discovery and exploitation of the decentralized, modular structure of the multiagent task [45], [48]–=-=[50]-=-. Providing domain knowledge to the agents can greatly help them in learning solutions to realistic tasks. In contrast, the large size of the state-action space and the delays in receiving informative   </text>
<query_num> 13707 </query_num>
<text>   L problem invalidates most of the single-agent RL theoretical guarantees. Despite its limitations, this approach has found a significant number of applications, mainly because of its simplicity [70], =-=[71]-=-, [85], [86]. One important step forward in understanding how singleagent RL works in multiagent tasks was made recently in [87]. The authors applied results in evolutionary game theory to analyze the   </text>
<query_num> 13708 </query_num>
<text>   actions with high likelihood of getting good rewards given the models [62]. The Frequency Maximum Q-value (FMQ) heuristic is based on the frequency with which actions yielded good rewards in the past =-=[63]-=-. Agent i uses Boltzmann action selection (5), plugging in modified Q-values ˜ Qi computed with the formula ˜Qi(ui) =Qi(ui)+ν Ci max(ui) Ci (ui) rmax(ui) (12) where rmax(ui) is the maximum reward obse   </text>
<query_num> 13709 </query_num>
<text>   and better. For instance, agents can exchange information using communication [51], skilled agents may serve as teachers for the learner [52], or the learner may watch and imitate the skilled agents =-=[53]-=-. When one or more agents fail in a multiagent system, the remaining agents can take over some of their tasks. This implies that MARL is inherently robust. Furthermore, by design, most multiagent syst ing to a high degree of scalability. Several existing MARL algorithms often require some additional preconditions to theoretically guarantee and to fully exploit the potential of these benefits [41], =-=[53]-=-. Relaxing these conditions and further improving the performance of the various MARL algorithms in this context is an active field of study.160 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PAR   </text>
<query_num> 13710 </query_num>
<text>   being hardwired into the agents at inception. The agents learn social conventions in [80], role assignments in [82], and the structure of the coordination graph together with the local Q-functions in =-=[83]-=-. Example 2: Coordination using social conventions in a fully cooperative task. In the earlier Section VI-A (see Fig. 3), suppose the agents are ordered such that agent 1 &amp;lt; agent 2 (a &amp;lt;b means that a   within sensor range [94], [95]. Pursuit involves the capture of moving targets by the robotic team. In a popular variant, several “predator” robots have to capture a “prey” robot by converging on it =-=[83]-=-, [96]. Object transportation requires the relocation of a set of objects into given final positions and configurations. The mass or size of some of the objects may exceed the transportation capabilit   </text>
<query_num> 13711 </query_num>
<text>   compasses temporal-difference reinforcement learning, game theory, and direct policy search techniques. aware algorithms use some form of opponent modeling to keep track of the other agents’ policies =-=[40]-=-, [76], [77]. The field of origin of the algorithms is a taxonomy axis that shows the variety of research inspiration benefiting MARL. MARL can be regarded as a fusion of temporal-difference RL, game   equilibrium strategies for all agents. This means the equilibrium selection problem arises when the solution of solve is not unique. A particular instance of solve and eval for, e.g., Nash Qlearning =-=[40]-=-, [54] is { evali{Q.,k (x, ·)} = Vi(x, NE{Q.,k (x, ·)}) (19) solvei{Q.,k (x, ·)} = NEi{Q.,k (x, ·)} where NE computes a Nash equilibrium (a set of strategies), NEi is agent i’s strategy component of t ximal or 2) every stage game has a Nash equilibrium that is a saddle point, i.e., not only does the learner not benefit from deviating from this equilibrium, but the other agents do benefit from this =-=[40]-=-, [88]. This requirement is satisfied only in a small class of problems. In all other cases, some external mechanism for equilibrium selection is needed for convergence. Instantiations of correlated e   </text>
<query_num> 13712 </query_num>
<text>   easure. 2) Clients of resources, as in [107]. The agents learn how to best select resources such that a given performance measure is optimized. A popular resource management domain is network routing =-=[108]-=-–[110]. Other examples include elevator scheduling [5] and load balancing [107]. Performance measures include average job processing times, minimum waiting time for resources, resource usage, and fair   </text>
<query_num> 13713 </query_num>
<text>   ed reflex behaviors could be built into the agents [70], [86]. Knowledge about the task structure could be used to decompose it into subtasks, and learn a modular solution with, e.g., hierarchical RL =-=[128]-=-. Last, but not the least, if a (possibly incomplete) task model is available, this model could be used with model-based RL algorithms to initialize Q-functions to reasonable, rather than arbitrary, v   </text>
<query_num> 13714 </query_num>
<text>   er, in an environment that changes over time, a hardwired behavior may become inappropriate. A reinforcement learning (RL) agent learns by trial-and-error interaction with its dynamic environment [6]–=-=[8]-=-. At each time step, the agent perceives the complete state of the environment and takes an action, which causes the environment to transit into a new state. The agent receives a scalar reward signal  pectation is taken over the probabilistic state transitions. The quantity Rk compactly represents the reward accumulated by the agent in the long run. Other possibilities of defining the return exist =-=[8]-=-. The discount factor γ can be regarded as encoding increasing uncertainty about rewards that will be received in the future, or as a means to bound the sum that otherwise might grow infinitely. The t   </text>
<query_num> 13715 </query_num>
<text>   etely. A simple social convention relies on a unique164 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART C: APPLICATIONS AND REVIEWS, VOL. 38, NO. 2, MARCH 2008 ordering of agents and actions =-=[80]-=-. These two orderings must be known to all agents. Combining them leads to a unique ordering of joint actions, and coordination is ensured if in (8) the first joint action in this ordering is selected rdination approaches have also been investigated, where the coordination structures are learned online, instead of being hardwired into the agents at inception. The agents learn social conventions in =-=[80]-=-, role assignments in [82], and the structure of the coordination graph together with the local Q-functions in [83]. Example 2: Coordination using social conventions in a fully cooperative task. In th   </text>
<query_num> 13716 </query_num>
<text>   f such bounds include maximum time constraints for reaching a desired performance level, or a lower bound on instantaneous performance levels. Some steps in this direction have been taken in [55] and =-=[57]-=-. C. Joint Environment and Learning Dynamics The stagewise application of game-theoretic techniques to solve dynamic multiagent tasks is a popular approach. It may, however, not be the most suitable,   </text>
<query_num> 13717 </query_num>
<text>   form of reward functions. The Hyper-Q algorithm uses the other agents’ models as a state vector and learns a Q-function Qi(ˆσ1,...,ˆσi−1, ˆσi+1, ...,ˆσn,ui) with an update rule similar to Q-learning =-=[68]-=-. By learning values of strategies instead of only actions, HyperQ should be able to adapt better to nonstationary agents. One inherent difficulty is that the action selection probabilities in 4 There   </text>
<query_num> 13718 </query_num>
<text>   gleagent RL [111]–[122]. A fair number of approximate MARL algorithms have been proposed: for discrete, large state-action spaces, e.g., [123], for continuous states and discrete actions, e.g., [96], =-=[98]-=-, and [124], and finally for continuous states and actions, e.g., [95], and [125]. Unfortunately, most of these algorithms only work in a narrow set of problems and are heuristic in nature. Significan   </text>
<query_num> 13719 </query_num>
<text>   he MARL problem invalidates most of the single-agent RL theoretical guarantees. Despite its limitations, this approach has found a significant number of applications, mainly because of its simplicity =-=[70]-=-, [71], [85], [86]. One important step forward in understanding how singleagent RL works in multiagent tasks was made recently in [87]. The authors applied results in evolutionary game theory to analy ness to imperfect observations are necessary, and few MARL algorithms exhibit these properties. In real-life applications, more direct derivations of single-agent RL (see Section VI-D1) are preferred =-=[70]-=-, [85], [86], [90]. In this section, several representative application domains are reviewed: distributed control, multirobot teams, trading agents, and resource management. A. Distributed Control In  ses: retrieval of objects, coverage of as much of the environment surface as possible, and exploration, where the robots have to bring into sensor range as much of the environment surface as possible =-=[70]-=-, [85], [86]. Multitarget observation is an extension of the exploration task, where the robots have to maintain a group of moving targets within sensor range [94], [95]. Pursuit involves the capture  ions, alsoBUS¸ONIU et al.: A COMPREHENSIVE SURVEY OF MULTIAGENT REINFORCEMENT LEARNING 169 rewarding promising behaviors rather than only the achievement of the goal, could be provided to the agents =-=[70]-=-, [86]. Humans or skilled agents could teach unskilled agents how to solve the task [126]. Shaping is a technique whereby the learning process starts by presenting the agents with simpler tasks, and p   </text>
<query_num> 13720 </query_num>
<text>   iations and auctions. For instance, the Trading Agent Competition is a simulated contest where the agents need to arrange travel packages by bidding for goods such as plane tickets and hotel bookings =-=[101]-=-. MARL approaches to this problem typically involve temporal-difference [34] or Q-learning agents, using approximate representations of the Q-functions to handle the large state space [102]–[105]. In   </text>
<query_num> 13721 </query_num>
<text>   in a narrow set of problems and are heuristic in nature. Significant advances in approximate MARL can be made if the wealth of theoretical results on single-agent approximate RL is put to use [112], =-=[113]-=-, [115]–[119]. A complementary avenue for improving scalability is the discovery and exploitation of the decentralized, modular structure of the multiagent task [45], [48]–[50]. Providing domain knowl   </text>
<query_num> 13722 </query_num>
<text>   ing simpler, local maximizations in terms of the local value functions, and aggregating their solutions. Under certain conditions, coordinated selection of an optimal joint action is guaranteed [45], =-=[46]-=-, [48]. In general, all the coordination techniques described in Section VI-B next can be applied to the fully cooperative MARL task. For instance, a framework to explicitly reason about possibly cost   </text>
<query_num> 13723 </query_num>
<text>   ings [101]. MARL approaches to this problem typically involve temporal-difference [34] or Q-learning agents, using approximate representations of the Q-functions to handle the large state space [102]–=-=[105]-=-. In some cases, cooperative agents represent the interest of a single company or individual, and merely fulfil different functions in the trading process, such as buying and selling [103], [104]. In   </text>
<query_num> 13724 </query_num>
<text>   irection has been investigated in, e.g., [45]–[50]. Experience sharing can help agents with similar tasks to learn faster and better. For instance, agents can exchange information using communication =-=[51]-=-, skilled agents may serve as teachers for the learner [52], or the learner may watch and imitate the skilled agents [53]. When one or more agents fail in a multiagent system, the remaining agents can ese strict requirements, by providing a way for the agents to exchange interesting data (e.g., state measurements or portions of Q-tables) rather than rely on exact measurements to ensure consistency =-=[51]-=-. Most algorithms also suffer from the curse of dimensionality. Distributed Q-learning and FMQ are exceptions in the sense that their complexity is not exponential in the number of agents (but they on   </text>
<query_num> 13725 </query_num>
<text>   ithms for dynamic tasks, and review relevant game-theoretic algorithms for static games. Other authors have investigated more closely the relationship between game theory and MARL. Bowling and Veloso =-=[13]-=- discuss several MARL algorithms, showing that these algorithms combine temporal difference RL with gametheoretic solvers for the static games arising in each state of the dynamic environment. Shoham  es are typeset in italics throughout the paper, e.g., Q-learning. II. BACKGROUND: REINFORCEMENT LEARNING In this section, the necessary background on single-agent and multiagent RL is introduced [7], =-=[13]-=-. First, the single-agent task is defined and its solution is characterized. Then, the multiagent task is defined. Static multiagent tasks are introduced separately, together with necessary game-theor ], and model-learning methods that estimate a model, and then learn using model-based techniques [36], [37]. Most MARL algorithms are derived from a model-free algorithm called Q-learning [32], e.g., =-=[13]-=-, [38]–[42]. Q-learning [32] turns (2) into an iterative approximation procedure. The current estimate of Q ∗ is updated using estimated samples of the right-hand side of (2). These samples are comput Several types of MARL goals have been proposed in the literature, which consider stability of the agent’s learning dynamics [54], adaptation to the changing behavior of the other agents [55], or both =-=[13]-=-, [38], [56]–[58]. A detailed analysis of this open problem is given in Section IV. Nonstationarity of the multiagent learning problem arises because all the agents in the system are learning simultan ncerns have been voiced regarding their usefulness. For instance, in [14], it is argued that the link between stagewise convergence to Nash equilibria and performance in the dynamic SG is unclear. In =-=[13]-=- and [56], convergence is required for stability, and rationality is added as an adaptation criterion. For an algorithm to be convergent, the authors of [13] and [56] require that the learner converge   </text>
<query_num> 13726 </query_num>
<text>   l-learning methods that estimate a model, and then learn using model-based techniques [36], [37]. Most MARL algorithms are derived from a model-free algorithm called Q-learning [32], e.g., [13], [38]–=-=[42]-=-. Q-learning [32] turns (2) into an iterative approximation procedure. The current estimate of Q ∗ is updated using estimated samples of the right-hand side of (2). These samples are computed using ac ic game. In this case, the goals are formulated in terms of stage strategies instead of strategies, and expected returns instead of rewards. Convergence to equilibria is a basic stability requirement =-=[42]-=-, [54]. It means the agents’ strategies should eventually converge to a coordinated equilibrium. Nash equilibria are most frequently used. However, concerns have been voiced regarding their usefulness ependent of the other agents share a common structure based on Q-learning, where policies and state values are computed with game-theoretic solvers for the stage games arising in the states of the SG =-=[42]-=-, [61]. This is similar to (13) and (14); the only difference is that for mixed games, solvers can be different from minimax. Denoting by {Q.,k (x, ·)} the stage game arising in state x and given by a atisfied only in a small class of problems. In all other cases, some external mechanism for equilibrium selection is needed for convergence. Instantiations of correlated equilibrium Q-learning (CE-Q) =-=[42]-=- or asymmetric Q-learning [72] can be performed in a similar fashion, by using correlated or Stackelberg (leader–follower) equilibria, respectively. For asymmetric-Q, the follower does not need to mod oals and performance in the dynamic task is not clear, and is to the authors’ best knowledge not made explicit in the literature. This holds for stability requirements, like convergence to equilibria =-=[42]-=-, [54], as well as for adaptation requirements, like rationality [13], [56]. Stability of the learning process is needed, because the behavior of stable agents is more amenable to analysis and meaning   </text>
<query_num> 13727 </query_num>
<text>   lem invalidates most of the single-agent RL theoretical guarantees. Despite its limitations, this approach has found a significant number of applications, mainly because of its simplicity [70], [71], =-=[85]-=-, [86]. One important step forward in understanding how singleagent RL works in multiagent tasks was made recently in [87]. The authors applied results in evolutionary game theory to analyze the dynam o imperfect observations are necessary, and few MARL algorithms exhibit these properties. In real-life applications, more direct derivations of single-agent RL (see Section VI-D1) are preferred [70], =-=[85]-=-, [86], [90]. In this section, several representative application domains are reviewed: distributed control, multirobot teams, trading agents, and resource management. A. Distributed Control In distri etrieval of objects, coverage of as much of the environment surface as possible, and exploration, where the robots have to bring into sensor range as much of the environment surface as possible [70], =-=[85]-=-, [86]. Multitarget observation is an extension of the exploration task, where the robots have to maintain a group of moving targets within sensor range [94], [95]. Pursuit involves the capture of mov   </text>
<query_num> 13728 </query_num>
<text>   lose criterion in (25) is based either on a comparison of an average policy with the current one, in the original version of WoLF-PHC, or on the second-order difference of policy elements, in PD-WoLF =-=[74]-=-. The Extended Optimal Response (EXORL) heuristic [75] applies a complementary idea in two-agent tasks: the policy update is biased in a way that minimizes the other agent’s incentive to deviate from   </text>
<query_num> 13729 </query_num>
<text>   ly work in a narrow set of problems and are heuristic in nature. Significant advances in approximate MARL can be made if the wealth of theoretical results on single-agent approximate RL is put to use =-=[112]-=-, [113], [115]–[119]. A complementary avenue for improving scalability is the discovery and exploitation of the decentralized, modular structure of the multiagent task [45], [48]–[50]. Providing domai   </text>
<query_num> 13730 </query_num>
<text>   m action choices are somehow coordinated or negotiated. Mechanisms for doing so, based on social conventions, roles, and communication, are described next (mainly following the description of Vlassis =-=[2]-=-). The mechanisms here can be used for any type of task (cooperative, competitive, or mixed). Both social conventions and roles restrict the action choices of the agents. An agent role restricts the s  the first joint action in this ordering is selected by all the agents. Communication can be used to negotiate action choices, either alone or in combination with the aforementioned techniques, as in =-=[2]-=- and [81]. When combined with the aforementioned techniques, communication can relax their assumptions and simplify their application. For instance, in social conventions, if only an ordering between   </text>
<query_num> 13731 </query_num>
<text>   model-learning methods that estimate a model, and then learn using model-based techniques [36], [37]. Most MARL algorithms are derived from a model-free algorithm called Q-learning [32], e.g., [13], =-=[38]-=-–[42]. Q-learning [32] turns (2) into an iterative approximation procedure. The current estimate of Q ∗ is updated using estimated samples of the right-hand side of (2). These samples are computed usi l types of MARL goals have been proposed in the literature, which consider stability of the agent’s learning dynamics [54], adaptation to the changing behavior of the other agents [55], or both [13], =-=[38]-=-, [56]–[58]. A detailed analysis of this open problem is given in Section IV. Nonstationarity of the multiagent learning problem arises because all the agents in the system are learning simultaneously ies of learning algorithms discussed in the literature can be identified. For instance, opponent-independent learning is related to stability, whereas opponent-aware learning is related to adaptation =-=[38]-=-, [59]. An opponent-independent algorithm converges to a strategy that is part of an equilibrium solution regardless of what the other agents are doing. An opponentaware algorithm learns models of the ume that agent 1 will take L1, and consequently, takes L2. The resulting joint action (R1,L2) is largely suboptimal, as the agents collide. 1) Coordination-Free Methods: The Team Q-learning algorithm =-=[38]-=- avoids the coordination problem by assuming that the optimal joint actions are unique (which is rarely the case). Then, if all the agents learn the common Q-function in parallel with (7), they can sa ed: maximize one’s benefit under the worst-case assumption that the opponent will always endeavor to minimize it. This principle suggests using opponentindependent algorithms. The minimax-Q algorithm =-=[38]-=-, [39] employs the minimax principle to compute strategies and values for the stage games, and a temporal-difference rule similar to Q-learning to propagate the values across state-action pairs. The a   </text>
<query_num> 13732 </query_num>
<text>   mpler, local maximizations in terms of the local value functions, and aggregating their solutions. Under certain conditions, coordinated selection of an optimal joint action is guaranteed [45], [46], =-=[48]-=-. In general, all the coordination techniques described in Section VI-B next can be applied to the fully cooperative MARL task. For instance, a framework to explicitly reason about possibly costly com ate RL is put to use [112], [113], [115]–[119]. A complementary avenue for improving scalability is the discovery and exploitation of the decentralized, modular structure of the multiagent task [45], =-=[48]-=-–[50]. Providing domain knowledge to the agents can greatly help them in learning solutions to realistic tasks. In contrast, the large size of the state-action space and the delays in receiving inform   </text>
<query_num> 13733 </query_num>
<text>   n VI-B next can be applied to the fully cooperative MARL task. For instance, a framework to explicitly reason about possibly costly communication is the communicative multiagent team decision problem =-=[78]-=-. 3) Indirect Coordination Methods: Indirect coordination methods bias action selection toward actions that are likely to result in good rewards or returns. This steers the agents toward coordinated a   </text>
<query_num> 13734 </query_num>
<text>   n. A. Benefits of MARL A speedup of MARL can be realized thanks to parallel computation when the agents exploit the decentralized structure of the task. This direction has been investigated in, e.g., =-=[45]-=-–[50]. Experience sharing can help agents with similar tasks to learn faster and better. For instance, agents can exchange information using communication [51], skilled agents may serve as teachers fo er the conditions that the reward function is positive and Qi,0 =0∀i, the local policies of the agents provably converge to an optimal joint policy. 2) Coordination-Based Methods: Coordination graphs =-=[45]-=- simplify coordination when the global Q-function can be additively decomposed into local Q-functions that only depend on the actions of a subset of agents. For instance, in an SG with four agents, th y solving simpler, local maximizations in terms of the local value functions, and aggregating their solutions. Under certain conditions, coordinated selection of an optimal joint action is guaranteed =-=[45]-=-, [46], [48]. In general, all the coordination techniques described in Section VI-B next can be applied to the fully cooperative MARL task. For instance, a framework to explicitly reason about possibl proximate RL is put to use [112], [113], [115]–[119]. A complementary avenue for improving scalability is the discovery and exploitation of the decentralized, modular structure of the multiagent task =-=[45]-=-, [48]–[50]. Providing domain knowledge to the agents can greatly help them in learning solutions to realistic tasks. In contrast, the large size of the state-action space and the delays in receiving   </text>
<query_num> 13735 </query_num>
<text>   not review here evolutionary learning techniques. Evolutionary learning, and in general, direct optimization of the agent behaviors, cannot readily benefit from the RL task structure. Panait and Luke =-=[17]-=- offer a comprehensive survey of evolutionary learning, as well as MARL, but only for cooperative agent teams. For the interested reader, examples of coevolution techniques, where the behaviors of the arning template (17)–(18) and in the stagewise win/lose criteria in WoLF algorithms. However, the suitability of such stagewise solutions in the context of the dynamic task is currently unclear [14], =-=[17]-=-. One important research step is understanding the conditions under which single-agent RL works in mixed SGs, especially in light of the preference toward single-agent techniques in practice. This was   </text>
<query_num> 13736 </query_num>
<text>   ns that produced good rewards in the past steers the agent toward coordination. b) Dynamic tasks: In Optimal Adaptive Learning (OAL), virtual games are constructed on top of each stage game of the SG =-=[64]-=-. In these virtual games, optimal joint actions are rewarded with 1, and the rest of the joint actions with 0. An algorithm is introduced that, by biasing the agent toward recently selected optimal ac   </text>
<query_num> 13737 </query_num>
<text>   nts evolve in parallel, can be found in [18]–[20]. Complementary, team learning techniques, where the entire set of agent behaviors is discovered by a single evolution process, can be found, e.g., in =-=[21]-=-–[23]. Evolutionary multiagent learning is a special case of a larger class of techniques originating in optimization theory that explore directly the space of agent behaviors. Other examples in this   </text>
<query_num> 13738 </query_num>
<text>   o reasonable, rather than arbitrary, values. Incomplete, uncertain state measurements could be handled with techniques related to partially observable Markov decision processes [129], as in [130] and =-=[131]-=-. B. Learning Goal The issue of a suitable MARL goal for dynamic tasks with dynamic, learning agents, is a difficult open problem. MARL goals are typically formulated in terms of static games. Their e   </text>
<query_num> 13739 </query_num>
<text>   observations are necessary, and few MARL algorithms exhibit these properties. In real-life applications, more direct derivations of single-agent RL (see Section VI-D1) are preferred [70], [85], [86], =-=[90]-=-. In this section, several representative application domains are reviewed: distributed control, multirobot teams, trading agents, and resource management. A. Distributed Control In distributed contro with the controllers, and the robots’ environment together with their sensors and actuators identify with the process. Particular distributed control domains where MARL is applied are process control =-=[90]-=-, control of traffic signals [91], [92], and control of electrical power networks [93]. B. Robotic Teams Robotic teams (also called multirobot systems) are the most popular application domain of MARL,   </text>
<query_num> 13740 </query_num>
<text>   ong the robots [4]. In resource management, while resources can be managed by a central authority, identifying each resource with an agent may provide a helpful, distributed perspective on the system =-=[5]-=-. Manuscript received November 10, 2006; revised March 7, 2007 and June 18, 2007. This work was supported by the Senter, Ministry of Economic Affairs of The Netherlands, under Grant BSIK03024 within t  parallel with the market [102], [105], [106]. D. Resource Management In resource management, the agents form a cooperative team, and they can be one of the following. 1) Managers of resources, as in =-=[5]-=-. Each agent manages one resource, and the agents learn how to best service requests in order to optimize a given performance measure. 2) Clients of resources, as in [107]. The agents learn how to bes   </text>
<query_num> 13741 </query_num>
<text>   ong which are fully cooperative, repeated games [62]. The MetaStrategy algorithm, introduced in [55], combines modified versions of fictitious play, minimax, and a gametheoretic strategy called Bully =-=[89]-=- to achieve the targeted optimality, compatibility, and safety goals (see Section IV). To compute best responses, the fictitious play and MetaStrategy algorithms require a model of the static task, in   </text>
<query_num> 13742 </query_num>
<text>   or 2) every stage game has a Nash equilibrium that is a saddle point, i.e., not only does the learner not benefit from deviating from this equilibrium, but the other agents do benefit from this [40], =-=[88]-=-. This requirement is satisfied only in a small class of problems. In all other cases, some external mechanism for equilibrium selection is needed for convergence. Instantiations of correlated equilib   </text>
<query_num> 13743 </query_num>
<text>   ots’ environment together with their sensors and actuators identify with the process. Particular distributed control domains where MARL is applied are process control [90], control of traffic signals =-=[91]-=-, [92], and control of electrical power networks [93]. B. Robotic Teams Robotic teams (also called multirobot systems) are the most popular application domain of MARL, encountered under the broadest r   </text>
<query_num> 13744 </query_num>
<text>   programming [29]–[31], model-free methods based on online estimation of value functions [32]–[35], and model-learning methods that estimate a model, and then learn using model-based techniques [36], =-=[37]-=-. Most MARL algorithms are derived from a model-free algorithm called Q-learning [32], e.g., [13], [38]–[42]. Q-learning [32] turns (2) into an iterative approximation procedure. The current estimate   </text>
<query_num> 13745 </query_num>
<text>   ral difference, i.e., the difference between the estimates of Q∗ (xk ,uk) at two successive time steps, k +1 and k. The sequence Qk provably converges to Q∗ under the following conditions [32], [43], =-=[44]-=-. 1) Explicit, distinct values of the Q-function are stored and updated for each state-action pair. 2) The time series of learning rates used for each state-action pair sums to infinity, whereas the s   </text>
<query_num> 13746 </query_num>
<text>   rning, as well as MARL, but only for cooperative agent teams. For the interested reader, examples of coevolution techniques, where the behaviors of the agents evolve in parallel, can be found in [18]–=-=[20]-=-. Complementary, team learning techniques, where the entire set of agent behaviors is discovered by a single evolution process, can be found, e.g., in [21]–[23]. Evolutionary multiagent learning is a   </text>
<query_num> 13747 </query_num>
<text>   s for which convergence is not guaranteed. a) Static tasks: The algorithms presented here assume the availability of a model of the static task, in the form of reward functions. The AWESOME algorithm =-=[60]-=- uses fictitious play, but monitors the other agents and, when it concludes that they are nonstationary, switches from the best response in fictitious play to a centrally precomputed Nash equilibrium  n application domain per se, game-theoretic, stateless tasks are often used to test MARL approaches. Not only algorithms specifically designed for static games are tested on such tasks (e.g., AWESOME =-=[60]-=-, MetaStrategy [55], GIGAWoLF [57]), but also others that can, in principle, handle dynamic SGs (e.g., EXORL [75]). As an avenue for future work, note that distributed control is poorly represented as   </text>
<query_num> 13748 </query_num>
<text>   s of MARL goals have been proposed in the literature, which consider stability of the agent’s learning dynamics [54], adaptation to the changing behavior of the other agents [55], or both [13], [38], =-=[56]-=-–[58]. A detailed analysis of this open problem is given in Section IV. Nonstationarity of the multiagent learning problem arises because all the agents in the system are learning simultaneously. Each ve been voiced regarding their usefulness. For instance, in [14], it is argued that the link between stagewise convergence to Nash equilibria and performance in the dynamic SG is unclear. In [13] and =-=[56]-=-, convergence is required for stability, and rationality is added as an adaptation criterion. For an algorithm to be convergent, the authors of [13] and [56] require that the learner converges to a st have been used in dynamic games in the stagewise fashion explained in the beginning of Section IV, although their extension to dynamic games was not explained in the papers that introduced them [13], =-=[56]-=-. Noregret has not been used in dynamic games, but it could be extended in a similar way. It is unclear how targeted optimality, compatibility, and safety could be extended. Fig. 1. Breakdown of MARL  h equilibrium (hence the name: Adapt When Everyone is Stationary, Otherwise Move to Equilibrium). In repeated games, AWESOME is provably rational and convergent [60] according to the definitions from =-=[56]-=- and [13] given in Section IV. Some methods in the area of direct policy search use gradient update rules that guarantee convergence in specific classes of static games: Infinitesimal Gradient Ascent  s’ best knowledge not made explicit in the literature. This holds for stability requirements, like convergence to equilibria [42], [54], as well as for adaptation requirements, like rationality [13], =-=[56]-=-. Stability of the learning process is needed, because the behavior of stable agents is more amenable to analysis and meaningful performance guarantees. Adaptation to the other agents is needed becaus   </text>
<query_num> 13749 </query_num>
<text>   s the agents’ returns are correlated and cannot be maximized independently. Several types of MARL goals have been proposed in the literature, which consider stability of the agent’s learning dynamics =-=[54]-=-, adaptation to the changing behavior of the other agents [55], or both [13], [38], [56]–[58]. A detailed analysis of this open problem is given in Section IV. Nonstationarity of the multiagent learni e. In this case, the goals are formulated in terms of stage strategies instead of strategies, and expected returns instead of rewards. Convergence to equilibria is a basic stability requirement [42], =-=[54]-=-. It means the agents’ strategies should eventually converge to a coordinated equilibrium. Nash equilibria are most frequently used. However, concerns have been voiced regarding their usefulness. For  ibrium strategies for all agents. This means the equilibrium selection problem arises when the solution of solve is not unique. A particular instance of solve and eval for, e.g., Nash Qlearning [40], =-=[54]-=- is { evali{Q.,k (x, ·)} = Vi(x, NE{Q.,k (x, ·)}) (19) solvei{Q.,k (x, ·)} = NEi{Q.,k (x, ·)} where NE computes a Nash equilibrium (a set of strategies), NEi is agent i’s strategy component of this eq  playing soccer. In navigation, each robot has to find its way from a starting position to a fixed or changing goal position, while avoiding obstacles and harmful interference with other robots [13], =-=[54]-=-. Area sweeping involves navigation through the environment for one of several purposes: retrieval of objects, coverage of as much of the environment surface as possible, and exploration, where the ro nd performance in the dynamic task is not clear, and is to the authors’ best knowledge not made explicit in the literature. This holds for stability requirements, like convergence to equilibria [42], =-=[54]-=-, as well as for adaptation requirements, like rationality [13], [56]. Stability of the learning process is needed, because the behavior of stable agents is more amenable to analysis and meaningful pe   </text>
<query_num> 13750 </query_num>
<text>   search use gradient update rules that guarantee convergence in specific classes of static games: Infinitesimal Gradient Ascent (IGA) [66], Winor-Learn-Fast IGA (WoLF-IGA) [13], Generalized IGA (GIGA) =-=[67]-=-, and GIGA-WoLF [57]. For instance, IGA and WoLF-IGA work in two-agent, two-action games, and use similar gradient update rules ⎧ ⎪⎨ ∂E{r1 | α, β} αk+1 = αk + δ1,k ∂α ⎪⎩ βk+1 = βk + δ2,k ∂E{r2 | α, β}   </text>
<query_num> 13751 </query_num>
<text>   ses temporal-difference reinforcement learning, game theory, and direct policy search techniques. aware algorithms use some form of opponent modeling to keep track of the other agents’ policies [40], =-=[76]-=-, [77]. The field of origin of the algorithms is a taxonomy axis that shows the variety of research inspiration benefiting MARL. MARL can be regarded as a fusion of temporal-difference RL, game theory  learner), and the learner has a model of the opponent’s policy, it might actually do better than the minimax return (15). An opponent model can be learned using, e.g., the M ∗ algorithm described in =-=[76]-=-, or a simple extension of (11) to multiple states ˆh i C j (x, uj )= i j (x, uj ) ∑ ũ j ∈U j Ci j (x, ũj ) (16) where C i j (x, uj ) counts the number of times agent i observed agent j taking action   </text>
<query_num> 13752 </query_num>
<text>   set of problems and are heuristic in nature. Significant advances in approximate MARL can be made if the wealth of theoretical results on single-agent approximate RL is put to use [112], [113], [115]–=-=[119]-=-. A complementary avenue for improving scalability is the discovery and exploitation of the decentralized, modular structure of the multiagent task [45], [48]–[50]. Providing domain knowledge to the a   </text>
<query_num> 13753 </query_num>
<text>   st joint action in this ordering is selected by all the agents. Communication can be used to negotiate action choices, either alone or in combination with the aforementioned techniques, as in [2] and =-=[81]-=-. When combined with the aforementioned techniques, communication can relax their assumptions and simplify their application. For instance, in social conventions, if only an ordering between agents is   </text>
<query_num> 13754 </query_num>
<text>   teach unskilled agents how to solve the task [126]. Shaping is a technique whereby the learning process starts by presenting the agents with simpler tasks, and progressively moves toward complex ones =-=[127]-=-. Preprogrammed reflex behaviors could be built into the agents [70], [86]. Knowledge about the task structure could be used to decompose it into subtasks, and learn a modular solution with, e.g., hie   </text>
<query_num> 13755 </query_num>
<text>   ted by the learner, or statistics of the values observed in the past. a) Static tasks: Joint Action Learners (JAL) learn jointaction values and employ empirical models of the other agents’ strategies =-=[62]-=-. Agent i learns models for all the other agents j ̸= i, using ˆσ i C j (uj )= i j (uj ) ∑ ũ j ∈U j Ci j (ũj (11) ) where ˆσ i j is agent i’s model of agent j’s strategy and Ci j (uj ) counts the numb r of times agent i observed agent j taking action uj . Several heuristics are proposed to increase the learner’s Q-values for the actions with high likelihood of getting good rewards given the models =-=[62]-=-. The Frequency Maximum Q-value (FMQ) heuristic is based on the frequency with which actions yielded good rewards in the past [63]. Agent i uses Boltzmann action selection (5), plugging in modified Q- .,ˆσi n [65]. The models are computed empirically using (11). Fictitious play converges to a Nash equilibrium in certain restricted classes of games, among which are fully cooperative, repeated games =-=[62]-=-. The MetaStrategy algorithm, introduced in [55], combines modified versions of fictitious play, minimax, and a gametheoretic strategy called Bully [89] to achieve the targeted optimality, compatibili   </text>
<query_num> 13756 </query_num>
<text>   thus requiring several robots to coordinate in order to bring about the objective [86]. Robot soccer is a popular, complex test-bed for MARL, that requires most of the skills enumerated earlier [4], =-=[97]-=-–[100]. For instance, intercepting the ball and leading it into the goal involve object retrieval and transportation skills, while the strategic placement of the players in the field is an advanced ve   </text>
<query_num> 13757 </query_num>
<text>   tiagent systems are finding applications in a wide variety of domains including robotic teams, distributed control, resource management, collaborative decision support systems, data mining, etc. [3], =-=[4]-=-. They may arise as the most natural way of looking at the system, or may provide an alternative perspective on systems that are originally regarded as centralized. For instance, in robotic teams, the  programmed with behaviors designed in advance, it is often necessary that they learn new behaviors online, such that the performance of the agent or of the whole multiagent system gradually improves =-=[4]-=-, [6]. This is usually because the complexity of the environment makes the aprioridesign of a good agent behavior difficult, or even, impossible. Moreover, in an environment that changes over time, a  obot, thus requiring several robots to coordinate in order to bring about the objective [86]. Robot soccer is a popular, complex test-bed for MARL, that requires most of the skills enumerated earlier =-=[4]-=-, [97]–[100]. For instance, intercepting the ball and leading it into the goal involve object retrieval and transportation skills, while the strategic placement of the players in the field is an advan   </text>
<query_num> 13758 </query_num>
<text>   unctions to reasonable, rather than arbitrary, values. Incomplete, uncertain state measurements could be handled with techniques related to partially observable Markov decision processes [129], as in =-=[130]-=- and [131]. B. Learning Goal The issue of a suitable MARL goal for dynamic tasks with dynamic, learning agents, is a difficult open problem. MARL goals are typically formulated in terms of static game   </text>
<query_num> 13759 </query_num>
<text>   volve in parallel, can be found in [18]–[20]. Complementary, team learning techniques, where the entire set of agent behaviors is discovered by a single evolution process, can be found, e.g., in [21]–=-=[23]-=-. Evolutionary multiagent learning is a special case of a larger class of techniques originating in optimization theory that explore directly the space of agent behaviors. Other examples in this class   </text>
<query_num> 13760 </query_num>
<text>   y learning, as well as MARL, but only for cooperative agent teams. For the interested reader, examples of coevolution techniques, where the behaviors of the agents evolve in parallel, can be found in =-=[18]-=-–[20]. Complementary, team learning techniques, where the entire set of agent behaviors is discovered by a single evolution process, can be found, e.g., in [21]–[23]. Evolutionary multiagent learning   </text>
<query_num> 13761 </query_num>
<text>   ynamic programming [29]–[31], model-free methods based on online estimation of value functions [32]–[35], and model-learning methods that estimate a model, and then learn using model-based techniques =-=[36]-=-, [37]. Most MARL algorithms are derived from a model-free algorithm called Q-learning [32], e.g., [13], [38]–[42]. Q-learning [32] turns (2) into an iterative approximation procedure. The current est   </text>
</top>
<top>
<paper_num> 138 </paper_num>
<paper_title>   Location coding for mobile image retrieval.  </paper_title>
<paper_abstract>   For mobile image retrieval, efficient data transmission can be achieved by sending only the query features. Each query feature is composed of a descriptor and a location in the image. The former is used to find candidate matching images using a “bag-of-words ” approach while the latter is used in a geometric consistency check to map features in the query image to corresponding features in the database image. We investigate how to compress the location information and how lossy compression affects the geometric consistency check. The location information is converted into a location histogram and a context-based arithmetic coding with location refinement method is then proposed to code the histogram. The effects of lossily compressing the location information are evaluated empirically in terms of the errors in corresponding features and the error of the estimated geometric transformation model. From our experiments, rates at ∼5.1 bits per feature can achieve errors comparable to lossless coding. The proposed scheme achieves a 12.5 × rate reduction compared to the floating point representation, and 2.8 × rate reduction compared to a fixed point representation.  </paper_abstract>
<query_num> 13801 </query_num>
<text>   SURF descriptors [14]. CHOG, a low bit rate descriptor, was proposed in [9]. CHOG, at the rate of 8 bytes per descriptor, can achieve matching capabilities comparable to the uncompressed features. In =-=[15]-=-, the local image patch around the interest point is compressed and sent over the network at low bit rate, which also shifts the workload of descriptor computation to the server side. Unlike the other   </text>
<query_num> 13802 </query_num>
<text>   its Per Feature Figure 7: Feature loss percentage vs. rate of the proposed scheme Compression of Hessian-based Detectors. We experiment with two other types of detectors, the Hessian-Laplace detector =-=[6]-=- and the SURF Fast Hessian detector [7], and show the results in Fig. 8. The parameters of the detectors were tuned to give an average of 750 feature counts per image. For fair comparison, we examine   </text>
<query_num> 13803 </query_num>
<text>   mation: the location of the feature and the descriptor of the local region around a feature. Features are formed at interest points in the image that have specific structures such as corners or blobs =-=[4, 5, 6, 7]-=-. These structures are typically invariant to changes in luminance, scale, rotation, and translation. Descriptors are then calculated from the image intensity values in a local region centered at the   </text>
<query_num> 13804 </query_num>
<text>   mation: the location of the feature and the descriptor of the local region around a feature. Features are formed at interest points in the image that have specific structures such as corners or blobs =-=[4, 5, 6, 7]-=-. These structures are typically invariant to changes in luminance, scale, rotation, and translation. Descriptors are then calculated from the image intensity values in a local region centered at the   A wide range of descriptors were tested in [8] and gradient distribution-based descriptors were shown to perform best. These type of descriptors, such as the Scale Invariant Feature Transform (SIFT) =-=[5]-=-, the Gradient Location and Orientation Histogram (GLOH) [8], the Speeded Up Robust Features (SURF) [7], and Compressed Histogram of Gradients (CHOG) [9], are histograms of the image intensity gradien or re-rank the list of candidate images [1, 12]. This is done by pairwise matching the features extracted from a query image to the features extracted from a database image. Typically, the ratio test =-=[5]-=- is used to predict a set of corresponding features between the two images. Then, RANSAC [10] is applied on the predicted set of features to estimated a geometric transformation model and a set of mat tected in different camera views. Figure 2: Features in the CD Database (left column) and the Zurich Building Database (right column) are randomly scattered but still cluster near structures. In SIFT =-=[5]-=-, Lowe proposed to detect features by finding extrema points in a pyramid of Difference of Gaussian (DoG) images. A 3D quadratic function is fitted to the local samples in the images to find the sub-p rmation causes errors in the GCC. 3.1 Geometric Consistency Check GCC is used to validate the match between the query image and the candidate image. The check consists of two steps: 1) the ratio test =-=[5]-=-, and 2) RANSAC [10]. We describe the process using the following notations. For a query image IQ, a set of features FQ = {lq,i, dq,i}i∈1,2,...N Q is extracted. l denotes the location, d denotes the d   </text>
<query_num> 13805 </query_num>
<text>   mation: the location of the feature and the descriptor of the local region around a feature. Features are formed at interest points in the image that have specific structures such as corners or blobs =-=[4, 5, 6, 7]-=-. These structures are typically invariant to changes in luminance, scale, rotation, and translation. Descriptors are then calculated from the image intensity values in a local region centered at the  to perform best. These type of descriptors, such as the Scale Invariant Feature Transform (SIFT) [5], the Gradient Location and Orientation Histogram (GLOH) [8], the Speeded Up Robust Features (SURF) =-=[7]-=-, and Compressed Histogram of Gradients (CHOG) [9], are histograms of the image intensity gradients of the local region. Fast large-scale image retrieval is enabled by using a Scalable Vocabulary Tree percentage vs. rate of the proposed scheme Compression of Hessian-based Detectors. We experiment with two other types of detectors, the Hessian-Laplace detector [6] and the SURF Fast Hessian detector =-=[7]-=-, and show the results in Fig. 8. The parameters of the detectors were tuned to give an average of 750 feature counts per image. For fair comparison, we examine the results using the location quantiza   </text>
<query_num> 13806 </query_num>
<text>   minance, scale, rotation, and translation. Descriptors are then calculated from the image intensity values in a local region centered at the interest point. A wide range of descriptors were tested in =-=[8]-=- and gradient distribution-based descriptors were shown to perform best. These type of descriptors, such as the Scale Invariant Feature Transform (SIFT) [5], the Gradient Location and Orientation Hist ry image. Then, ground truth correspondences are established if they satisfy the following criteria: 1) the projected database image feature patch and the query image feature patch have overlap error =-=[8]-=- smaller then 40%, 2) the location error between projected database feature location and the query feature location is within 2 pixels, and 3) the orientation difference of the projected database feat   </text>
<query_num> 13807 </query_num>
<text>   of the system critically depends on how much information is transferred in both directions. 1.1 Background on Image Retrieval Image retrieval based on local features has become an attractive approach =-=[3]-=- since they are robust to lighting, rotation, scale, and mild viewpoint changes. A local feature consists of two pieces of information: the location of the feature and the descriptor of the local regi   </text>
<query_num> 13808 </query_num>
<text>   ompressed Histogram of Gradients (CHOG) [9], are histograms of the image intensity gradients of the local region. Fast large-scale image retrieval is enabled by using a Scalable Vocabulary Tree (SVT) =-=[11]-=- for matching. Features are extracted from the database of images to form a set of features. A hierarchical k-means clustering algorithm is applied to the set of features, forming the SVT. This method   </text>
<query_num> 13809 </query_num>
<text>   t-based image retrieval 1. INTRODUCTION Handheld mobile devices, such as camera-phones or PDAs, are expected to become ubiquitous platforms for visual search and mobile augmented reality applications =-=[1, 2]-=-. For mobile image matching, a visual database is typically stored at a server in the network. Hence, for a visual comparison, information must be either uploaded from the mobile to the server, or dow am to the query image histogram. The descriptors are treated as a visual “bag-of-words” in the SVT matching. Geometric consistency check is applied to validate or re-rank the list of candidate images =-=[1, 12]-=-. This is done by pairwise matching the features extracted from a query image to the features extracted from a database image. Typically, the ratio test [5] is used to predict a set of corresponding f   </text>
<query_num> 13810 </query_num>
<text>   terest point is compressed and sent over the network at low bit rate, which also shifts the workload of descriptor computation to the server side. Unlike the other approaches, the retrieval system in =-=[16]-=- sends a tree histogram in place of individual descriptors, which enables significant additional rate reduction. 1.3 Contributions A novel method to compress the location information is proposed. The  ×8 4×4 2×2 Bits/Image 222 807 2014 3730 5552 Bits/Feature 0.247 0.900 2.246 4.160 6.192 1 2 1 1 1 1 3 1 1 1 Figure 3: The histogram of the features contains the histogram map and the histogram count. =-=[16]-=-, it was shown that by discarding the order information of the coded items the bit rate can be reduced. Second, the interest points are structural points in the image, and hence, spatial structure rel formation in a location histogram. The location histogram is an orderless representation in which the information of order is discarded. Therefore, there is a potential bit rate reduction of log2(N!) =-=[16]-=-, where N is the total number of features in the image. The average number of features per query image in the data set is 750. Using the Stirling’s approximation to calculate log2(N!), we find that a   </text>
</top>
<top>
<paper_num> 139 </paper_num>
<paper_title>   Optimal pricing for multiple services in telecommunications networks offering quality-of-service guarantees.  </paper_title>
<paper_abstract>   Abstract — We consider pricing for multiple services offered over a single telecommunications network. Each service has quality of service (QoS) requirements that are guaranteed to users. Service classes may be defined by the type of service, such as voice, video or data, as well as the origin and destination of the connection provided to the user. We formulate the optimal pricing problem as a nonlinear integer expected revenue optimization problem. We simultaneously solve for prices and the resource allocations necessary to provide connections with guaranteed QoS. We derive optimality conditions and a solution method for this class of problems, and apply to a realistic model of a multi-service communications network.  </paper_abstract>
<query_num> 13901 </query_num>
<text>   ach to call admission, it has been suggested that users guarantee their own QoS by purchasing the required bandwidth and buffer resources for their desired QoS directly from the network [19] [21]. In =-=[31]-=-, an analysis of a marketbased methodology is offered as evidence that pricing schemes can offer efficient resource allocations in connectionoriented networks offering QoS guarantees. (See also [33])   </text>
<query_num> 13902 </query_num>
<text>   e have been suggested that would allow multiple service classes, and induce users to behave fairly and efficiently, through simple or randomized packet marking mechanisms in the network [10][13] [14] =-=[20]-=-. Usage based schemes, which charge based on the actual resources used have also been proposed for these types of networks [8] [29]. Priority pricing is another suggestion for allowing multiple servic   </text>
<query_num> 13903 </query_num>
<text>   ernative approach to call admission, it has been suggested that users guarantee their own QoS by purchasing the required bandwidth and buffer resources for their desired QoS directly from the network =-=[19]-=- [21]. In [31], an analysis of a marketbased methodology is offered as evidence that pricing schemes can offer efficient resource allocations in connectionoriented networks offering QoS guarantees. (S   </text>
<query_num> 13904 </query_num>
<text>   ionoriented networks offering QoS guarantees. (See also [33]) These models feature a conventional view of pricing per connection and announcing the price to the public. The paper by de Veciana et al. =-=[32]-=- combines the conventional practice of prices being fixed and announced, with the alternative view of allowing users to guarantee their own QoS by purchasing To Appear in IEEE/ACM Transactions on Netw  , Bijk , PLOSSijk ) ≤ 0 (8) ∂PLOSS ijk An in-depth analysis of equivalent capacity is outside the focus of this paper. Detailed discussions of buffering and multiplexing gains are given in [2][9][28]=-=[32]-=- and [34]. However, it should be noted that our approach remains unchanged for any derivation of equivalent capacity, which satisfies the general properties for equivalent capacity, given in (5) – (8)   </text>
<query_num> 13905 </query_num>
<text>   ive approach to call admission, it has been suggested that users guarantee their own QoS by purchasing the required bandwidth and buffer resources for their desired QoS directly from the network [19] =-=[21]-=-. In [31], an analysis of a marketbased methodology is offered as evidence that pricing schemes can offer efficient resource allocations in connectionoriented networks offering QoS guarantees. (See al   </text>
<query_num> 13906 </query_num>
<text>   n [31], an analysis of a marketbased methodology is offered as evidence that pricing schemes can offer efficient resource allocations in connectionoriented networks offering QoS guarantees. (See also =-=[33]-=-) These models feature a conventional view of pricing per connection and announcing the price to the public. The paper by de Veciana et al. [32] combines the conventional practice of prices being fixe   </text>
<query_num> 13907 </query_num>
<text>   nally, recent results have been offered, based on dynamic programming that suggest a static price schedule results in network performance that is near optimal compared to congestion dependent pricing =-=[27]-=-. 2) Pricing with QoS Guarantees Typically a network with guaranteed QoS, such as the current voice network, must employ a call admission policy in order to satisfy the guarantees to users. In an alte   </text>
<query_num> 13908 </query_num>
<text>   net, offers best effort service, which is prone to unpredictable congestion and delays by definition. The current flow control scheme in the internet is called transmission control protocol (TCP) [2] =-=[34]-=-. Modifications to this scheme have been suggested that would allow multiple service classes, and induce users to behave fairly and efficiently, through simple or randomized packet marking mechanisms   PLOSSijk ) ≤ 0 (8) ∂PLOSS ijk An in-depth analysis of equivalent capacity is outside the focus of this paper. Detailed discussions of buffering and multiplexing gains are given in [2][9][28][32] and =-=[34]-=-. However, it should be noted that our approach remains unchanged for any derivation of equivalent capacity, which satisfies the general properties for equivalent capacity, given in (5) – (8) above. T ntinue for prolonged periods. The peak and mean data rates are defined conservatively, based on MPEG-1 trace data in [29]. The length of the idle/busy period is chosen T i 10sbased on a suggestion in =-=[34]-=- that video sessions see busy periods in the order of 10 seconds. For users with buffer space at the client end, a modest delay in the transmission of packets is acceptable. Some packet loss can be co   </text>
<query_num> 13909 </query_num>
<text>   or randomized packet marking mechanisms in the network [10][13] [14] [20]. Usage based schemes, which charge based on the actual resources used have also been proposed for these types of networks [8] =-=[29]-=-. Priority pricing is another suggestion for allowing multiple services over best effort networks. The most well known work discusses a second bid auction, whereby it is incentive compatible, or in th  of data traffic, where the bursts, such as action scenes can be a high peak and continue for prolonged periods. The peak and mean data rates are defined conservatively, based on MPEG-1 trace data in =-=[29]-=-. The length of the idle/busy period is chosen T i 10sbased on a suggestion in [34] that video sessions see busy periods in the order of 10 seconds. For users with buffer space at the client end, a mo   </text>
<query_num> 13910 </query_num>
<text>   ple or randomized packet marking mechanisms in the network [10][13] [14] [20]. Usage based schemes, which charge based on the actual resources used have also been proposed for these types of networks =-=[8]-=- [29]. Priority pricing is another suggestion for allowing multiple services over best effort networks. The most well known work discusses a second bid auction, whereby it is incentive compatible, or   </text>
<query_num> 13911 </query_num>
<text>   to choose from among several routes, each with its own delay, has been shown to have a stable equilibrium solution where the relative prices induce the desired operating point of the network operator =-=[16]-=-[17]. An interesting problem of regulating the arrival of jobs presented to the network is discussed in [26], and the informational requirements are relaxed, using an adaptive on-line method in [25].   </text>
<query_num> 13912 </query_num>
<text>   to this scheme have been suggested that would allow multiple service classes, and induce users to behave fairly and efficiently, through simple or randomized packet marking mechanisms in the network =-=[10]-=-[13] [14] [20]. Usage based schemes, which charge based on the actual resources used have also been proposed for these types of networks [8] [29]. Priority pricing is another suggestion for allowing m   </text>
</top>
<top>
<paper_num> 140 </paper_num>
<paper_title>   Dynamic multi-resolution data dissemination in storage-centric wireless sensor networks.  </paper_title>
<paper_abstract>   A short version of this paper appeared at MSWiM 2007.  </paper_abstract>
<query_num> 14001 </query_num>
<text>   ; 2) it does not consider the impact of data rate changes on the choice of power-efficient dissemination trees. Energy-efficient unicast/multicast/broadcast for ad hoc networks [2], [27], [22], [19], =-=[35]-=-, [33] has been studied extensively in the last decade. The ETX is first proposed in [2] as a metric to quantify the transmission energy consumption of lossy links. The existing solutions on energy-ef   </text>
<query_num> 14002 </query_num>
<text>   aption heuristics are discussed in Section V and Section VI, respectively. We present the simulation results in Section VII, and the conclusion in Section VIII. II. RELATED WORK A number of solutions =-=[18]-=-, [37], [7], [25], [17] have been proposed to disseminate data from sources to sinks in WSNs. Directed diffusion (DD) [17] is a data-centric communication protocol where user interests are flooded to   case. For instance, flooding or a network-wide grid structure are used by DD and TTDD to facilitate the finding of data sources, which are not necessary when the locations of sources are known. SEAD =-=[18]-=- is designed to disseminate data from a known source to mobile sinks. The paths with different data rates in SEAD can be shared to reduce the number of transmissions. Our work differs from SEAD in sev is either the base station or another node in the network. For instance, remote users may issue queries through the base station while in-network users may issue queries through the nodes around them =-=[18]-=-. A data request is specified by a triple (t, p, [Φ, Ψ]), where t is the sink, p is the temporal resolution, Φ and Ψ are two optional parameters that specify the property of data requested and the dur ange dynamically. In this paper, we assume data sinks are stationary or have low mobility. We leverage the existing solutions to optimize the dissemination topology when sinks are highly mobile [37], =-=[18]-=-. For example, the stationary proxy nodes [37] chosen for the mobile users can serve as data sinks and hence the disruption to the dissemination tree due to mobility can be reduced. B. Network Model W roximation algorithm where the cost of a link is the maximum data rate of the flows on the link. DST is similar to the tree construction algorithm used by an existing data dissemination protocol SEAD =-=[18]-=-. Total Energy Cost(J) 100 90 80 70 60 50 40 30 20 10 MIDT TST DST MTT 0 2 4 6 8 10 Num of Requests Total Energy Cost(J) 140 120 100 80 60 40 20 MIDT TST DST MTT 0 2 4 6 8 10 Num of Requests Control M   </text>
<query_num> 14003 </query_num>
<text>   ast focus on minimizing the total transmission power of nodes on the Steiner/spanning tree that connects the source to multiple sinks. Recently, the problem of minimum-energy reliable multicast [20], =-=[3]-=- has also been studied. Although these algorithms can be used to transmit data from a source to multiple sinks, they are not suitable for dynamic multi-resolution data dissemination because: 1) they d   </text>
<query_num> 14004 </query_num>
<text>   disruption to the dissemination tree due to mobility can be reduced. B. Network Model We assume all nodes in the network are equipped with omnidirectional antennas. A CSMA MAC protocol such as S-MAC =-=[38]-=- is employed. We assume wireless links are lossy, which is consistent with recent empirical findings [41]. Expected transmission count (ETX) [2], [9] of a link is the expected total number of transmis synchronous sleep schedule determined by the temporal resolutions of requests. All other nodes run in a synchronous sleep schedule with a lower duty cycle. Several lowpower MAC protocols (e.g., S-MAC =-=[38]-=- employ synchronous sleep schedules in which all nodes wake up and go to sleep together. In our problem, the total energy consumption of the nodes on the tree during the lifetime of the requests shoul rs that can be chosen according to the temporal resolutions supported by the network. Such a periodic sleeping scheduling mechanism is supported by several power-efficient MAC protocols such as S-MAC =-=[38]-=-. For a data request i specified by (t, pi, [Φi, Ψi]) (see Section III-A), we define the normalized data rate of request i as ri = |Φi| B·pi where |Φi| represents the size of each packet sent by the s   </text>
<query_num> 14005 </query_num>
<text>   e of a wireless channel; 2) it does not consider the impact of data rate changes on the choice of power-efficient dissemination trees. Energy-efficient unicast/multicast/broadcast for ad hoc networks =-=[2]-=-, [27], [22], [19], [35], [33] has been studied extensively in the last decade. The ETX is first proposed in [2] as a metric to quantify the transmission energy consumption of lossy links. The existin irectional antennas. A CSMA MAC protocol such as S-MAC [38] is employed. We assume wireless links are lossy, which is consistent with recent empirical findings [41]. Expected transmission count (ETX) =-=[2]-=-, [9] of a link is the expected total number of transmissions before the sender successfully sends a packet to the receiver through the link. To deal with the link loss, we adopt the per-hop reliabili r problem is a large constant and hence we focus on deriving the average power consumption of nodes. 8power consumption of lossy links is computed in the same way in several recent works [20], [12], =-=[2]-=-. For the rest of time the radio operates in an idle or reception state. Formulation (1) only models the power consumption of data transmissions. The transmission power of acknowledgements can be inco   </text>
<query_num> 14006 </query_num>
<text>   e to the diversity of user requirements, a WSN often needs to collect and report physical information at multiple temporal granularities. For instance, many WSN query processing systems (e.g., TinyDB =-=[26]-=-) support windowed stream query in the form of “report to node i the average temperature of region A once every T seconds”. T is the temporal resolution of the query specified by a user. As a real-wor rt the average temperature of region A to node 5 over the last hour once every minute. We 5note that such a request is also referred to as the sliding-window query in many streaming database systems =-=[26]-=-. The temporal resolution of each request can be different according to the interest of the sink. Moreover, in accordance with several representative WSN applications, we assume the resolution request   </text>
<query_num> 14007 </query_num>
<text>   heuristics are discussed in Section V and Section VI, respectively. We present the simulation results in Section VII, and the conclusion in Section VIII. II. RELATED WORK A number of solutions [18], =-=[37]-=-, [7], [25], [17] have been proposed to disseminate data from sources to sinks in WSNs. Directed diffusion (DD) [17] is a data-centric communication protocol where user interests are flooded to the ne can change dynamically. In this paper, we assume data sinks are stationary or have low mobility. We leverage the existing solutions to optimize the dissemination topology when sinks are highly mobile =-=[37]-=-, [18]. For example, the stationary proxy nodes [37] chosen for the mobile users can serve as data sinks and hence the disruption to the dissemination tree due to mobility can be reduced. B. Network M   </text>
<query_num> 14008 </query_num>
<text>   in our problem is a large constant and hence we focus on deriving the average power consumption of nodes. 8power consumption of lossy links is computed in the same way in several recent works [20], =-=[12]-=-, [2]. For the rest of time the radio operates in an idle or reception state. Formulation (1) only models the power consumption of data transmissions. The transmission power of acknowledgements can be   </text>
<query_num> 14009 </query_num>
<text>   ional antennas. A CSMA MAC protocol such as S-MAC [38] is employed. We assume wireless links are lossy, which is consistent with recent empirical findings [41]. Expected transmission count (ETX) [2], =-=[9]-=- of a link is the expected total number of transmissions before the sender successfully sends a packet to the receiver through the link. To deal with the link loss, we adopt the per-hop reliability mo   </text>
<query_num> 14010 </query_num>
<text>   l, a source node serves the user requests with possibly different data refresh rates. The source node could be a cluster head that stores and aggregates the data from all sensor nodes in its vicinity =-=[13]-=-, or a storage node that caches the network-wide information tagged with certain properties. For instance, a storage node in a surveillance WSN may store and index all information related to a specifi   </text>
<query_num> 14011 </query_num>
<text>   less channel; 2) it does not consider the impact of data rate changes on the choice of power-efficient dissemination trees. Energy-efficient unicast/multicast/broadcast for ad hoc networks [2], [27], =-=[22]-=-, [19], [35], [33] has been studied extensively in the last decade. The ETX is first proposed in [2] as a metric to quantify the transmission energy consumption of lossy links. The existing solutions   </text>
<query_num> 14012 </query_num>
<text>   much smaller. Accurate simulation to the highly probabilistic link characterization [41] of WSNs is the key to evaluating the realistic performance of MIDT. We implemented a link layer model from USC =-=[42]-=- in Prowler. Experimental data shows that the USC model can simulate the highly unreliable links on the Mica2 motes [42]. To improve the link reliability, we implemented an ARQ (Automatic Repeat Reque   </text>
<query_num> 14013 </query_num>
<text>   n the deployments of wireless sensor networks (WSNs) in a variety of dataintensive applications including micro-climate and habitat monitoring [32], precision agriculture, security surveillance [15], =-=[23]-=-, etc. The capability of long-term and dense sensing of WSNs enables the physical environments to be monitored at an unprecedented granularity. One of the major tasks of WSNs is to disseminate useful   </text>
<query_num> 14014 </query_num>
<text>   notation. 1) R(u, v) represents the ETX of link (u, v), which is computed as 1 εu,v , where εu,v denotes the link reception probability of link (u, v) and can be obtained from online link estimators =-=[36]-=-. Note that R(u, v) may not equal R(v, u) due to link asymmetry. 2) G(VG, EG) represents the network. V includes all nodes in the network and E is defined as E = {(u, v)|(u, v ∈ V ) ∧ (R(u, v) ≤ ℜ)} w eme that retransmits a packet if an acknowledgment is not received after a timeout. The maximum number of retransmissions before dropping a packet is 8. A link quality estimator similar to the one in =-=[36]-=- is used by each node to periodically assess the ETXs to its neighbors. 27In our simulations, 200 nodes are deployed in a 150 × 150 m2 region divided into 10 × 10 grids. Each grid contains two nodes   </text>
<query_num> 14015 </query_num>
<text>   ral resolution of the query specified by a user. As a real-world example, an environmental monitoring WSN may receive two types of queries in which a meteorologist requests hourly temperature updates =-=[24]-=- for weather analysis while a biologist 2requests a temperature reading once every several minutes from bird burrows for detailed study of birds’ breeding behavior [32]. Furthermore, the temporal res   </text>
<query_num> 14016 </query_num>
<text>   range. For instance, a surveillance WSN may normally report a reading summary to a user in every a few minutes and must increase the data rate to about a few packets/second when a target is detected =-=[39]-=-. To evaluate MIDT under a wide range of possible scenarios, the data rate in our simulations is varied from 1-200 packets during the 50-second active interval. The size of each packet is 30 bytes, wh   </text>
<query_num> 14017 </query_num>
<text>   roblem of finding the minimum weight Steiner tree in directed graphs. The best centralized algorithm for this problem has an approximation ratio of |D| ǫ where ǫ is a positive constant with 0 &amp;lt; ǫ ≤ 1 =-=[8]-=-. When D = VG \ {s}, all the nodes except 11the source are sinks. As a result, the term |VT |z in (3) becomes a constant for a given network. This special case of the MMDD problem is similar to the o   </text>
<query_num> 14018 </query_num>
<text>   stics are discussed in Section V and Section VI, respectively. We present the simulation results in Section VII, and the conclusion in Section VIII. II. RELATED WORK A number of solutions [18], [37], =-=[7]-=-, [25], [17] have been proposed to disseminate data from sources to sinks in WSNs. Directed diffusion (DD) [17] is a data-centric communication protocol where user interests are flooded to the network data to the user. TTDD [37] is a two-tier dissemination protocol in which each source maintains a network-wide grid structure and sinks can locate the source through local flooding. Cetintemel et al. =-=[7]-=- developed a tree-based dissemination protocol in which nodes change their sleeping schedules dynamically according to event generation rates. Machado et al. [25] proposed to disseminate data to sinks   </text>
<query_num> 14019 </query_num>
<text>   t does not consider the impact of data rate changes on the choice of power-efficient dissemination trees. Energy-efficient unicast/multicast/broadcast for ad hoc networks [2], [27], [22], [19], [35], =-=[33]-=- has been studied extensively in the last decade. The ETX is first proposed in [2] as a metric to quantify the transmission energy consumption of lossy links. The existing solutions on energy-efficien   </text>
<query_num> 14020 </query_num>
<text>   t some new network modules and port our protocols to the Berkeley motes in the future. The MAC layer in Prowler employs a simple CSMA/CA scheme without RTS/CTS, which is similar to the B-MAC protocol =-=[28]-=- in TinyOS. This simple approach is not as effective as the more sophisticated protocols (e.g. IEEE 802.11 [1]) in terms of collision avoidance, but it certainly consumes less energy and the communica   </text>
<query_num> 14021 </query_num>
<text>   that caches the network-wide information tagged with certain properties. For instance, a storage node in a surveillance WSN may store and index all information related to a specific type of intruders =-=[30]-=-. A data request for the source is issued by a sink that is either the base station or another node in the network. For instance, remote users may issue queries through the base station while in-netwo   </text>
<query_num> 14022 </query_num>
<text>   uation and modeling I. INTRODUCTION Recent years have seen the deployments of wireless sensor networks (WSNs) in a variety of dataintensive applications including micro-climate and habitat monitoring =-=[32]-=-, precision agriculture, security surveillance [15], [23], etc. The capability of long-term and dense sensing of WSNs enables the physical environments to be monitored at an unprecedented granularity. sts hourly temperature updates [24] for weather analysis while a biologist 2requests a temperature reading once every several minutes from bird burrows for detailed study of birds’ breeding behavior =-=[32]-=-. Furthermore, the temporal resolution requested by the same user is subject to dynamic changes due to the evolution of environmental activities. For instance, in an intelligent building, a cluster he   </text>
<query_num> 14023 </query_num>
<text>   ve seen the deployments of wireless sensor networks (WSNs) in a variety of dataintensive applications including micro-climate and habitat monitoring [32], precision agriculture, security surveillance =-=[15]-=-, [23], etc. The capability of long-term and dense sensing of WSNs enables the physical environments to be monitored at an unprecedented granularity. One of the major tasks of WSNs is to disseminate u   </text>
<query_num> 14024 </query_num>
<text>   wer consumption of data transmissions. The transmission power of acknowledgements can be incorporated by extending the current formulation. We note that several energy-efficient communication schemes =-=[6]-=- do not use acknowledgement packets. For instance, the sender may overhear data packets forwarded by the next-hop node as implicit acknowledgements. As overhearing assumes the same amount of power as   </text>
</top>
<top>
<paper_num> 141 </paper_num>
<paper_title>   Towards content distribution networks with latency guarantees.  </paper_title>
<paper_abstract>   This paper investigates the performance of a content distribution network designed to provide bounded content access latency. Content can be divided into multiple classes with different configurable per-class delay bounds. The network uses a simple distributed algorithm to dynamically select a subset of its proxy servers for different classes such that a global per-class delay bound is achieved on content access. The content distribution algorithm is implemented and tested on PlanetLab [24], a world-wide distributed Internet testbed. Evaluation results demonstrate that despite Internet delay variability, subsecond delay bounds (of 200-500ms) can be guaranteed with a very high probability at only a moderate content replication cost. The distribution algorithm achieves a 4 to 5 fold reduction in the number of response-time violations compared to prior content distribution approaches that attempt to minimize average latency. To the authors&amp;apos; knowledge, this paper presents the first wide-area performance evaluation of an algorithm designed to bound maximum content access latency, as opposed to optimizing an average performance metric.  </paper_abstract>
<query_num> 14101 </query_num>
<text>   at the above simple delay estimation is sufficient for our purposes, other more sophisticated network measurement/estimation techniques in existing research such as bprobe [4], Pathload [10], PTR/IGI =-=[9]-=-, and Spruce [35] can be plugged into our system, if needed, without any impact on other modules. Latency (ms) 1600 1400 1200 1000 800 600 400 200 Server #1 Actual Server #1 Estimate Server #2 Actual   </text>
<query_num> 14102 </query_num>
<text>   ction for replica placement in CDNs. Our algorithm, being a purely distributed scheme, falls into this category, and can thus address scalability issues of existing replica placement algorithms. SCAN =-=[5]-=- approaches the replica placement problem by minimizing the number of replicas while meeting client latency and server capacity constraints. While it has the same objective as our algorithm, SCAN has   </text>
<query_num> 14103 </query_num>
<text>   d replica placement algorithm and found that with careful design, the router-level fanout-based placement algorithm is almost as good as a greedy algorithm in [25] in most cases. Venkataramani et al. =-=[36]-=- studied the replica placement problem of minimizing overall client access time under servers’ bandwidth constraints. The algorithm approaches the placement problem by hierarchically refining an initi   </text>
<query_num> 14104 </query_num>
<text>   erformance. Existing research on CDNs includes techniques for efficient client request redirection [1, 2, 37, 3], server placement strategies to improve average response time or bandwidth consumption =-=[20, 25, 11, 15, 27]-=-, logical overlay topologies for large-scale content distribution [34, 28, 29], consistency maintenance mechanisms [38, 23], and empirical CDN performance measurement studies [14, 18]. Little attentio evious algorithms. The efficacy of content distribution depends primarily on the placement of replicas. While the replica placement problem in CDNs has been extensively studied in previous literature =-=[20, 25, 11, 15, 27]-=-, the placement objectives 17shave mostly been to optimize some average performance metric such as client-perceived latency, number of ASes traversed, or some notion of cost of link traversal. Li et a  with network traffic conditions. This is different from the degree concept in our distributed replica selection algorithm which is tightly related to network and system load conditions. A later work =-=[27]-=- further evaluated a router fanout-based replica placement algorithm and found that with careful design, the router-level fanout-based placement algorithm is almost as good as a greedy algorithm in [2   </text>
<query_num> 14105 </query_num>
<text>   erformance. Existing research on CDNs includes techniques for efficient client request redirection [1, 2, 37, 3], server placement strategies to improve average response time or bandwidth consumption =-=[20, 25, 11, 15, 27]-=-, logical overlay topologies for large-scale content distribution [34, 28, 29], consistency maintenance mechanisms [38, 23], and empirical CDN performance measurement studies [14, 18]. Little attentio evious algorithms. The efficacy of content distribution depends primarily on the placement of replicas. While the replica placement problem in CDNs has been extensively studied in previous literature =-=[20, 25, 11, 15, 27]-=-, the placement objectives 17shave mostly been to optimize some average performance metric such as client-perceived latency, number of ASes traversed, or some notion of cost of link traversal. Li et a nsitive to imperfect input data. Kangasharju et al. explicitly took storage capacity of each individual CDN server into account and considered each AS as a node in a graph representing one CDN server =-=[15]-=-. The problem of optimizing average number of ASes traversed for client requests was formulated as a combinatorial optimization problem. The study by Jamin et al. [11] investigated the impact of the n In a recent study [16] on general replica placement algorithms, a unified framework was designed to classify them. According to their findings, most existing algorithms for CDNs such as algorithms in =-=[25, 11, 15]-=- do not scale for systems with more than ����� nodes because of the large number of content objects a CDN could host. They identified decentralized algorithms as an important direction for replica pla   </text>
<query_num> 14106 </query_num>
<text>   erformance. Existing research on CDNs includes techniques for efficient client request redirection [1, 2, 37, 3], server placement strategies to improve average response time or bandwidth consumption =-=[20, 25, 11, 15, 27]-=-, logical overlay topologies for large-scale content distribution [34, 28, 29], consistency maintenance mechanisms [38, 23], and empirical CDN performance measurement studies [14, 18]. Little attentio ple replicas from the enhancement that results from their non-random proper placement by our algorithm. Average Latency Greedy: A greedy network performance optimization algorithm has been studied in =-=[25, 11]-=-. Although its authors mentioned that it can essentially be applied to optimize any metric, it is applicable to optimizing only overall or average performance metrics, which is not the same as guarant evious algorithms. The efficacy of content distribution depends primarily on the placement of replicas. While the replica placement problem in CDNs has been extensively studied in previous literature =-=[20, 25, 11, 15, 27]-=-, the placement objectives 17shave mostly been to optimize some average performance metric such as client-perceived latency, number of ASes traversed, or some notion of cost of link traversal. Li et a raph representing one CDN server [15]. The problem of optimizing average number of ASes traversed for client requests was formulated as a combinatorial optimization problem. The study by Jamin et al. =-=[11]-=- investigated the impact of the number of replicas on the performance of various replica placement methods. Their major finding was that increasing the number of replicas is effective in reducing clie In a recent study [16] on general replica placement algorithms, a unified framework was designed to classify them. According to their findings, most existing algorithms for CDNs such as algorithms in =-=[25, 11, 15]-=- do not scale for systems with more than ����� nodes because of the large number of content objects a CDN could host. They identified decentralized algorithms as an important direction for replica pla   </text>
<query_num> 14107 </query_num>
<text>   erformance. Existing research on CDNs includes techniques for efficient client request redirection [1, 2, 37, 3], server placement strategies to improve average response time or bandwidth consumption =-=[20, 25, 11, 15, 27]-=-, logical overlay topologies for large-scale content distribution [34, 28, 29], consistency maintenance mechanisms [38, 23], and empirical CDN performance measurement studies [14, 18]. Little attentio ple replicas from the enhancement that results from their non-random proper placement by our algorithm. Average Latency Greedy: A greedy network performance optimization algorithm has been studied in =-=[25, 11]-=-. Although its authors mentioned that it can essentially be applied to optimize any metric, it is applicable to optimizing only overall or average performance metrics, which is not the same as guarant s that the client requests received by a CDN non-replica server can always be directed to the nearest replica and takes the corresponding latency value as the cost. According to simulation results of =-=[25]-=-, ALG performs very close to optimal in minimizing the average client-perceived latency in a CDN. In addition to the centralized algorithms mentioned above, we also introduce a variant of our original evious algorithms. The efficacy of content distribution depends primarily on the placement of replicas. While the replica placement problem in CDNs has been extensively studied in previous literature =-=[20, 25, 11, 15, 27]-=-, the placement objectives 17shave mostly been to optimize some average performance metric such as client-perceived latency, number of ASes traversed, or some notion of cost of link traversal. Li et a e authors did not evaluate the performance of their dynamic programming algorithm for a realistic Internet topology. Besides, the algorithm has a relatively high computational complexity � Qiu et al. =-=[25]-=- formulated the replica placement problem as that of choosing a fixed number of replicas among a given set of locations to minimize the overall client request latencies. This problem is mapped to a K-   </text>
<query_num> 14108 </query_num>
<text>   ontent via a self-organizing application-level multicast tree to find a server that meets the client perceived latency requirement and server resource constraints to be the replica. Rabinovich et al. =-=[26]-=- proposed a protocol suite for dynamic replication and migration of Internet objects. It has algorithms for deciding on the number and placement of replicas and an algorithm for distributing requests   </text>
<query_num> 14109 </query_num>
<text>   ple delay estimation is sufficient for our purposes, other more sophisticated network measurement/estimation techniques in existing research such as bprobe [4], Pathload [10], PTR/IGI [9], and Spruce =-=[35]-=- can be plugged into our system, if needed, without any impact on other modules. Latency (ms) 1600 1400 1200 1000 800 600 400 200 Server #1 Actual Server #1 Estimate Server #2 Actual Server #2 Estimat   </text>
<query_num> 14110 </query_num>
<text>   predominant use of the Internet for Web content delivery has recently spurred much research on CDN performance. Existing research on CDNs includes techniques for efficient client request redirection =-=[1, 2, 37, 3]-=-, server placement strategies to improve average response time or bandwidth consumption [20, 25, 11, 15, 27], logical overlay topologies for large-scale content distribution [34, 28, 29], consistency   To handle these situations, we need to make sure that when a server is overloaded, the content objects it hosts will be reduced and therefore its workload can be reduced. Redirecting client requests =-=[3, 37]-=- is one important technique to protect servers from flash events. While this technique alleviates overload, it does not solve the delay bound problem, as the delay elapsed to reach the overloaded serv   </text>
<query_num> 14111 </query_num>
<text>   predominant use of the Internet for Web content delivery has recently spurred much research on CDN performance. Existing research on CDNs includes techniques for efficient client request redirection =-=[1, 2, 37, 3]-=-, server placement strategies to improve average response time or bandwidth consumption [20, 25, 11, 15, 27], logical overlay topologies for large-scale content distribution [34, 28, 29], consistency  essens temporal locality of content access requests. As a result, many client requests have to be served from servers’ disks instead of memory. This phenomenon has been confirmed in previous research =-=[37]-=-. Finally, depending on content size, storage cost may or may not be an issue. For example, consider a future Starbucks chain that entertains its customers by offering wireless access to real-time (st  To handle these situations, we need to make sure that when a server is overloaded, the content objects it hosts will be reduced and therefore its workload can be reduced. Redirecting client requests =-=[3, 37]-=- is one important technique to protect servers from flash events. While this technique alleviates overload, it does not solve the delay bound problem, as the delay elapsed to reach the overloaded serv   </text>
<query_num> 14112 </query_num>
<text>   redirection [1, 2, 37, 3], server placement strategies to improve average response time or bandwidth consumption [20, 25, 11, 15, 27], logical overlay topologies for large-scale content distribution =-=[34, 28, 29]-=-, consistency maintenance mechanisms [38, 23], and empirical CDN performance measurement studies [14, 18]. Little attention has been given to QoS guarantee issues in CDNs. In this paper, we address th   </text>
<query_num> 14113 </query_num>
<text>   rements show that the average network latency of downloading a file is roughly proportional to its size when the file size is between 1KB and 100KB. This range encompasses the majority of web objects =-=[7, 30]-=- and is the range our system is targeted for. Hence: ¢¤£¦¥¨§�©���¥���� . We leverage this property to quickly 12sCDF CDF CDF 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 Span First Latency First 0.1 ALG Random S   </text>
<query_num> 14114 </query_num>
<text>   rform close to the centralized one. 4.2 Distributed Algorithm A few distributed algorithms for constructing dominating sets have been proposed. Liang and Haas used a distributed algorithm called DDCH =-=[21]-=- to generate virtual backbones for Ad Hoc networks. When synchronously executed, DDCH achieves the same approximation ratio as the greedy algorithm (Figure 3) £���§ but � there exist networks for whic information it has at the instant. This guarantees that a dominating set will be generated but the cardinality of the dominating set could be larger than that generated in a synchronous execution. In =-=[21]-=- the authors discussed the asynchronous operation of DDCH, which uses a periodic mode too, except that their algorithm usually takes multiple rounds to finish. The mode can also be applied to LRG [12] erformance in terms of minimizing NOR, we compare our distributed algorithm with the centralized greedy heuristic listed in Figure 3 and two distributed algorithms mentioned in Section 4, namely DDCH =-=[21]-=- and LRG [12]. We did not include Kuhn’s algorithm [19] in our comparison study because firstly 9sCentralized DDCH LRG DG 30 Nodes 200ms 300ms 400ms 200ms 300ms 400ms 200ms 300ms 400ms 200ms 300ms 400  size. This analysis can be used to determine the cost of QoS contracts. 7 Related Work There are many theoretic results on graph domination problem and its variants. The three distributed algorithms =-=[21, 12, 19]-=- we studied are the most related ones to our problem. In this paper, we gave our own distributed algorithm which is very simple yet performs well in a highly asynchronous environment and terminates fa   </text>
<query_num> 14115 </query_num>
<text>   s for deciding on the number and placement of replicas and an algorithm for distributing requests among available replicas. However their work does not address QoS guarantee issues. Ko and Rubenstein =-=[17]-=- investigated an abstracted problem of placing replicas of different types of resources 18 £������ � § .sin large-scale network systems such as P2P networks and wireless sensor networks. They consider   </text>
<query_num> 14116 </query_num>
<text>   tegies to improve average response time or bandwidth consumption [20, 25, 11, 15, 27], logical overlay topologies for large-scale content distribution [34, 28, 29], consistency maintenance mechanisms =-=[38, 23]-=-, and empirical CDN performance measurement studies [14, 18]. Little attention has been given to QoS guarantee issues in CDNs. In this paper, we address the specific problem of providing subsecond gua   </text>
<query_num> 14117 </query_num>
<text>   tion [20, 25, 11, 15, 27], logical overlay topologies for large-scale content distribution [34, 28, 29], consistency maintenance mechanisms [38, 23], and empirical CDN performance measurement studies =-=[14, 18]-=-. Little attention has been given to QoS guarantee issues in CDNs. In this paper, we address the specific problem of providing subsecond guarantees on access delay. The paper evaluates the feasibility   </text>
<query_num> 14118 </query_num>
<text>   vers then chose this server as their replica. Figure 10 plots the latencies of client requests and probings of one of the servers that chose � as its replica. At time 115, we started multiple httperf =-=[22]-=- clients in our LAN, sending requests at very high rates to � � to overload it. detected that latency to � became very high. Client requests directed to � � continuously exceeded the 300ms latency bou   </text>
<query_num> 14119 </query_num>
<text>   we believe that the above simple delay estimation is sufficient for our purposes, other more sophisticated network measurement/estimation techniques in existing research such as bprobe [4], Pathload =-=[10]-=-, PTR/IGI [9], and Spruce [35] can be plugged into our system, if needed, without any impact on other modules. Latency (ms) 1600 1400 1200 1000 800 600 400 200 Server #1 Actual Server #1 Estimate Serv   </text>
<query_num> 14120 </query_num>
<text>   with high probability, where is the maximum span of the graph and yields an approximation � � ratio with � expectation � of . Kuhn and Wattenhofer designed a more sophisticated distributed algorithm =-=[19]-=- based on £ LP ����� relaxation § tech� niques. Given an arbitrary constant , the algorithm computes a � ������� � dominating £ set § of expected size times optimal in rounds. Though theoretically app  the performance of these algorithms when � used in asynchronous systems was not studied. Moreover, these algorithms share a common disadvantage: they need multiple rounds to finish. Kuhn’s algorithm =-=[19]-=- can be made to finish in one round by setting � � � the performance will be far from optimal. In practice, more rounds translates into more overhead and a longer termination time, which may be proble istributed algorithm with the centralized greedy heuristic listed in Figure 3 and two distributed algorithms mentioned in Section 4, namely DDCH [21] and LRG [12]. We did not include Kuhn’s algorithm =-=[19]-=- in our comparison study because firstly 9sCentralized DDCH LRG DG 30 Nodes 200ms 300ms 400ms 200ms 300ms 400ms 200ms 300ms 400ms 200ms 300ms 400ms set 1 8 7 6 8.8 7.7 6.3 9 7.8 6.5 8.6 7.5 6.4 set 2   size. This analysis can be used to determine the cost of QoS contracts. 7 Related Work There are many theoretic results on graph domination problem and its variants. The three distributed algorithms =-=[21, 12, 19]-=- we studied are the most related ones to our problem. In this paper, we gave our own distributed algorithm which is very simple yet performs well in a highly asynchronous environment and terminates fa   </text>
</top>
<top>
<paper_num> 142 </paper_num>
<paper_title>   Rethinking the library OS from the top down.  </paper_title>
<paper_abstract>   “There is nothing new under the sun, but there are a lot of old things we don’t know.” – Ambrose Bierce, The Devil’s Dictionary This paper revisits an old approach to operating system construction, the library OS, in a new context. The idea of the library OS is that the personality of the OS on which an application depends runs in the address space of the application. A small, fixed set of abstractions connects the library OS to the host OS kernel, offering the promise of better system security and more rapid independent evolution of OS components. We describe a working prototype of a Windows 7 library OS that runs the latest releases of major applications such as Microsoft Excel, PowerPoint, and Internet Explorer. We demonstrate that desktop sharing across independent, securely isolated, library OS instances can be achieved through the pragmatic reuse of networking protocols. Each instance has significantly lower overhead than a full VM bundled with an application: a typical application adds just 16MB of working set and 64MB of disk footprint. We contribute a new ABI below the library OS that enables application mobility. We also show that our library OS can address many of the current uses of hardware virtual machines at a fraction of the overheads. This paper describes the first working prototype of a full commercial OS redesigned as a library OS capable of running significant applications. Our experience shows that the longpromised benefits of the library OS approach—better protection of system integrity and rapid system evolution—are readily obtainable.  </paper_abstract>
<query_num> 14201 </query_num>
<text>   Design. General Terms Experimentation, Performance. 1. Introduction D.4 [Operating Systems]: The library OS approach to OS construction was championed by several operating system designs in the 1990s =-=[3, 10, 13, 21]-=-. The idea of the library OS is that the entire personality of the OS on which an application depends runs in its address space as a library. An OS personality is the implementation of the OS’s applic plications. Cache Kernel, Exokernel, and Nemesis innovated by providing applications with fine-grained, customized control of hardware resources, such as page tables, network packets, and disk blocks =-=[10, 13, 21]-=-. In contrast, Drawbridge’s differing goals (security, host independence, and migration) free it to offer higher-level abstractions. These higher-level abstractions make it easier to share underlying   is useful in the common case (sequential file access), but can harm performance of an application that accesses its data randomly. This idea was brought to fruition in the Cache kernel [10], Nemesis =-=[21]-=-, and Exokernel [13]. The Cache kernel design focused on allowing applications fine-grained control over thread scheduling and memory swapping. The Nemesis OS was particularly focused on eliminating s   </text>
<query_num> 14202 </query_num>
<text>   Design. General Terms Experimentation, Performance. 1. Introduction D.4 [Operating Systems]: The library OS approach to OS construction was championed by several operating system designs in the 1990s =-=[3, 10, 13, 21]-=-. The idea of the library OS is that the entire personality of the OS on which an application depends runs in its address space as a library. An OS personality is the implementation of the OS’s applic plications. Cache Kernel, Exokernel, and Nemesis innovated by providing applications with fine-grained, customized control of hardware resources, such as page tables, network packets, and disk blocks =-=[10, 13, 21]-=-. In contrast, Drawbridge’s differing goals (security, host independence, and migration) free it to offer higher-level abstractions. These higher-level abstractions make it easier to share underlying  imization that is useful in the common case (sequential file access), but can harm performance of an application that accesses its data randomly. This idea was brought to fruition in the Cache kernel =-=[10]-=-, Nemesis [21], and Exokernel [13]. The Cache kernel design focused on allowing applications fine-grained control over thread scheduling and memory swapping. The Nemesis OS was particularly focused on   </text>
<query_num> 14203 </query_num>
<text>   Design. General Terms Experimentation, Performance. 1. Introduction D.4 [Operating Systems]: The library OS approach to OS construction was championed by several operating system designs in the 1990s =-=[3, 10, 13, 21]-=-. The idea of the library OS is that the entire personality of the OS on which an application depends runs in its address space as a library. An OS personality is the implementation of the OS’s applic plications. Cache Kernel, Exokernel, and Nemesis innovated by providing applications with fine-grained, customized control of hardware resources, such as page tables, network packets, and disk blocks =-=[10, 13, 21]-=-. In contrast, Drawbridge’s differing goals (security, host independence, and migration) free it to offer higher-level abstractions. These higher-level abstractions make it easier to share underlying  mmon case (sequential file access), but can harm performance of an application that accesses its data randomly. This idea was brought to fruition in the Cache kernel [10], Nemesis [21], and Exokernel =-=[13]-=-. The Cache kernel design focused on allowing applications fine-grained control over thread scheduling and memory swapping. The Nemesis OS was particularly focused on eliminating scheduling and perfor   </text>
<query_num> 14204 </query_num>
<text>   d. New paravirtualization approaches, which warp the VM interface gradually farther away from a raw hardware interface, expose instead resources that a guest operating system can use more efficiently =-=[5, 12, 17, 38, 39]-=-. Despite significant research effort, however, running a full legacy operating system in a VM still incurs substantial overhead, as the legacy operating system brings with it system management proces   </text>
<query_num> 14205 </query_num>
<text>   ess. All kernel data, however, reside within a shared, monolithic kernel address space. This approach of adding security checks and/or replicated kernel data structures is also exemplified by SELinux =-=[23]-=-, zones [29], jails [34] and containers [7], however it suffers several drawbacks. First, it is difficult to know that the job of inserting additional checks is done, or that the set of hooks is compl   </text>
<query_num> 14206 </query_num>
<text>   irtual machines [35] are the primary success story for packaging an application and its dependencies on a fully-featured OS into a single self-contained unit. Research systems, such as The Collective =-=[31]-=-, use a single virtual machine per application to encapsulate an application and its OS. Treating a complete legacy operating system, including the kernel, as a library for a single application wastes   </text>
<query_num> 14207 </query_num>
<text>   licy such as ―Do not allow the contents of file X to leave the machine‖ must be translated into a perilous series of decisions on whether to allow innocuous-looking accesses to local system resources =-=[15, 41]-=-. It is tempting to try to paint a layer of isolation functionality onto an existing monolithic system, but the result is hard to trust since it requires maintaining properties across a large interfac   </text>
<query_num> 14208 </query_num>
<text>   ode, albeit not 301entire interactive GUI applications [11]. The isolation technique proposed in Xax, namely hardware memory protection and a limited system call table, is shared by Drawbridge; NaCl =-=[40]-=- uses alternate techniques based on software fault isolation. Stronger isolation in monolithic kernels The monolithic organization of conventional operating systems makes them difficult to secure. In   </text>
<query_num> 14209 </query_num>
<text>   security monitor. In providing the ABI, the security monitor enforces a set of external policies governing the host OS resources available to the application. Inspired by previous work in Singularity =-=[33]-=-, we encode policy in manifest files associated with the application. The manifest whitelists the host OS resources that an application may access, identified by URI path. We also use the manifest as   </text>
<query_num> 14210 </query_num>
<text>   ss of functionality and performance, a number of systems have attempted to augment monolithic systems with additional security checks in a minimally disruptive manner. For instance, the Linux VServer =-=[32]-=- adds additional access control list checks and imposes a separate namespace for kernel objects used by an isolated process. All kernel data, however, reside within a shared, monolithic kernel address   </text>
<query_num> 14211 </query_num>
<text>   uses four host OS calls to issue requests over an anonymous named pipe to dkmon; the second, which replaces the NT system-call service table on a per-process basis using techniques developed for Xax =-=[11]-=-; and the third, which makes Hyper-V hypercalls. dkmon services ABI requests by modifying the address space and host OS handle table of the calling process with standard Windows cross-process manipula  isolating untrusted web applications demonstrated that a limited set of OS abstractions were sufficient for porting large libraries of legacy code, albeit not 301entire interactive GUI applications =-=[11]-=-. The isolation technique proposed in Xax, namely hardware memory protection and a limited system call table, is shared by Drawbridge; NaCl [40] uses alternate techniques based on software fault isola   </text>
<query_num> 14212 </query_num>
<text>   using the default sequential prefetching heuristics. Like many of its contemporaries, the library OS approach is largely forgotten, a casualty of the rise of the modern virtual machine monitor (VMM) =-=[8]-=-. While most new OS designs of the time— including library OS designs—ran only a handful of custom applications on small research prototypes, VMM systems proliferated because they could run major appl   </text>
</top>
<top>
<paper_num> 143 </paper_num>
<paper_title>   Supporting Multiple-Keyword Search in A Hybrid Structured Peer-to-Peer Network.  </paper_title>
<paper_abstract>   structured peer-to-peer (P2P) networks only support singlekeyword exact-match lookups. In practice, however, users often have fuzzy information for identifying these items and tend to submit broad queries. The support of searching based on multiple  </paper_abstract>
<query_num> 14301 </query_num>
<text>   . The desired results are all the items whose storage keywords contains the keywords in the query. Most proposed approaches supporting multiple-keyword search in DHTs use data replication method [10]–=-=[13]-=-. Generally, an item with multiple keywords is stored multiple times. Each time one of the keywords is used as the key to insert it into DHT nodes. When looking up some file, the query is resolved int mbers) Structured Overlay Backbone node (cluster leader) Fig. 1. The network structure of MKey Unstructured Overlays result. Observing that the search result of a single keyword may be of large size, =-=[13]-=- proposes to incrementally compute the intersection at different nodes and only return the final result to the user. Edge bandwidth to users is hence greatly saved. However, a problem in all the above gal MKey IDs. In summary, MKey is most applicable to items/queries with a large number of keywords. If most items and queries only contain a few keywords, simple data replication methods such as [10]–=-=[13]-=- are enough. With the increase of keywords, however, MKey shows a significant improvement as compared to these approaches, in both storage cost and search cost. MKey can be extended to support full-te ers and take disjoint groups of bits from MD5 signature as independent hash functions [15]. The total number of nodes is 100, 000. We compare MKey with the simple data replication method used in [10]–=-=[13]-=-. This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE ICC 2006 proceedings.sFig. 5. Average number of copies in st   </text>
<query_num> 14302 </query_num>
<text>   In fact, researchers have considered many ways to improve search functions in DHTs. For example, [7] uses latent semantic indexing technique to build up full-text search in structured overlays; [8], =-=[9]-=- combine structured and unstructured overlays to reduce search cost. Different from these works, we consider multiple-keyword search in structured networks. We assume a data item shared by some peer c   </text>
<query_num> 14303 </query_num>
<text>   alues of the elements in a set evenly distribute throughout the range of the function. The question of what hash function to use in practice remains an open problem; currently MD5 is a popular choice =-=[15]-=-. Suppose the Bloom filters are of m-bit length and use k hash functions. The target set A contains n elements. After all the elements of A are hashed into the Bloom filter, the probability that a spe the performance of MKey. We build MKey on top of a public Chord implementation [17]. We set m = 128 for Bloom filters and take disjoint groups of bits from MD5 signature as independent hash functions =-=[15]-=-. The total number of nodes is 100, 000. We compare MKey with the simple data replication method used in [10]–[13]. This full text paper was peer reviewed at the direction of IEEE Communications Socie   </text>
<query_num> 14304 </query_num>
<text>   an easily conduct search. However, in structured peer-to-peer networks it becomes much more difficult. In fact, researchers have considered many ways to improve search functions in DHTs. For example, =-=[7]-=- uses latent semantic indexing technique to build up full-text search in structured overlays; [8], [9] combine structured and unstructured overlays to reduce search cost. Different from these works, w   </text>
<query_num> 14305 </query_num>
<text>   cult. In fact, researchers have considered many ways to improve search functions in DHTs. For example, [7] uses latent semantic indexing technique to build up full-text search in structured overlays; =-=[8]-=-, [9] combine structured and unstructured overlays to reduce search cost. Different from these works, we consider multiple-keyword search in structured networks. We assume a data item shared by some p   </text>
<query_num> 14306 </query_num>
<text>   f Computer Science The Hong Kong University of Science and Technology Clear Water Bay, Kowloon, Hong Kong Email: {csvenus, kenyiu, gchan}@cs.ust.hk 1-4244-0355-3/06/$20.00 (c) 2006 IEEE Tables (DHTs) =-=[3]-=-–[6]. These systems are efficient because data items can be located in a small number of hops. However, a major limitation of DHT systems is that they only support lookups with exact match, i.e., a se stem Architecture II. SYSTEM DESIGN As shown above, MKey consists of a structured network as its backbone. Any existing DHT systems such as Chord, Pastry and Tapestry can be modified to underlay MKey =-=[3]-=-– [6]. Each backbone node is also the leader of a cluster. Cluster members form an unstructured overlay and cooperate to share the loads (storage and search) assigned to their leader. There are two re filter. We call this Bloom filter an item descriptor. In almost all existing DHT systems, node IDs are computed by a hash function based on nodespecific information such as IP address and private key =-=[3]-=-– [6]. However, in our system, node IDs have strict formation rules. First of all, all the node IDs are m-bit. We define two classes of IDs: the first class consists of all the IDs that have only one   return to that in traditional DHT systems. Any existing DHT system such as Chord, Pastry and Tapestry may underlay MKey to provide these basic functions. We take Chord in our current implementations =-=[3]-=-. Chord has no restrictions on node IDs since it only orderly organizes nodes in a ring, while prefix routing in Pastry or postfix routing in Tapestry is more applicable to a large number of diverse n   </text>
<query_num> 14307 </query_num>
<text>   h large storage loads tends to form a large cluster. The whole system can hence achieve load balancing. In MKey, a data item’s keywords are represented by a Bloom filter, which is a m-bit vector [14]–=-=[16]-=-. This vector directs the storage locations of the item index. A query is also mapped to a m-bit vector according to its keywords, from which a search range is determined. We carefully control the map contains n elements. After all the elements of A are hashed into the Bloom filter, the probability that a specific bit is still 0 can be computed as [14]: � P0(n) = 1 − 1 �kn ≈ e m −kn/m . Similar to =-=[16]-=-, we assume the bits in the Bloom filter are independently set to 0 with probability P0(n) and set to 1 with probability (1 − P0(n)). Therefore, the average number of bit-1 in the Bloom filter is m(1   </text>
<query_num> 14308 </query_num>
<text>   leader. The union of all the search results serves as the final search result. In more details, we use a Bloom filter to summarize an item’s storage keywords, which is a bit vector of certain length =-=[14]-=-. The Bloom filter is split into several bit vectors of the same length, each corresponding to a backbone node ID. Each of the nodes needs to store one copy of the item index within its clusters. In a e with large storage loads tends to form a large cluster. The whole system can hence achieve load balancing. In MKey, a data item’s keywords are represented by a Bloom filter, which is a m-bit vector =-=[14]-=-–[16]. This vector directs the storage locations of the item index. A query is also mapped to a m-bit vector according to its keywords, from which a search range is determined. We carefully control th  length and use k hash functions. The target set A contains n elements. After all the elements of A are hashed into the Bloom filter, the probability that a specific bit is still 0 can be computed as =-=[14]-=-: � P0(n) = 1 − 1 �kn ≈ e m −kn/m . Similar to [16], we assume the bits in the Bloom filter are independently set to 0 with probability P0(n) and set to 1 with probability (1 − P0(n)). Therefore, the   </text>
<query_num> 14309 </query_num>
<text>   mputer Science The Hong Kong University of Science and Technology Clear Water Bay, Kowloon, Hong Kong Email: {csvenus, kenyiu, gchan}@cs.ust.hk 1-4244-0355-3/06/$20.00 (c) 2006 IEEE Tables (DHTs) [3]–=-=[6]-=-. These systems are efficient because data items can be located in a small number of hops. However, a major limitation of DHT systems is that they only support lookups with exact match, i.e., a search Architecture II. SYSTEM DESIGN As shown above, MKey consists of a structured network as its backbone. Any existing DHT systems such as Chord, Pastry and Tapestry can be modified to underlay MKey [3]– =-=[6]-=-. Each backbone node is also the leader of a cluster. Cluster members form an unstructured overlay and cooperate to share the loads (storage and search) assigned to their leader. There are two reasons r. We call this Bloom filter an item descriptor. In almost all existing DHT systems, node IDs are computed by a hash function based on nodespecific information such as IP address and private key [3]– =-=[6]-=-. However, in our system, node IDs have strict formation rules. First of all, all the node IDs are m-bit. We define two classes of IDs: the first class consists of all the IDs that have only one bit o   </text>
<query_num> 14310 </query_num>
<text>   words. The desired results are all the items whose storage keywords contains the keywords in the query. Most proposed approaches supporting multiple-keyword search in DHTs use data replication method =-=[10]-=-–[13]. Generally, an item with multiple keywords is stored multiple times. Each time one of the keywords is used as the key to insert it into DHT nodes. When looking up some file, the query is resolve in legal MKey IDs. In summary, MKey is most applicable to items/queries with a large number of keywords. If most items and queries only contain a few keywords, simple data replication methods such as =-=[10]-=-–[13] are enough. With the increase of keywords, however, MKey shows a significant improvement as compared to these approaches, in both storage cost and search cost. MKey can be extended to support fu  filters and take disjoint groups of bits from MD5 signature as independent hash functions [15]. The total number of nodes is 100, 000. We compare MKey with the simple data replication method used in =-=[10]-=-–[13]. This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE ICC 2006 proceedings.sFig. 5. Average number of copies   </text>
</top>
<top>
<paper_num> 144 </paper_num>
<paper_title>   Incorporating prior knowledge in support vector machines for classification: A review.  </paper_title>
<paper_abstract>   For classification, support vector machines (SVMs) have recently been introduced and quickly became the state of the art. Now, the incorporation of prior knowledge into SVMs is the key element that allows to increase the performance in many applications. This paper gives a review of the current state of research regarding the incorporation of two general types of prior knowledge into SVMs for classification. The particular forms of prior knowledge considered here are presented in two main groups: class-invariance and knowledge on the data. The first one includes invariances to transformations, to permutations and in domains of input space, whereas the second one contains knowledge on unlabeled data, the imbalance of the training set or the quality of the data. The methods are then described and classified in the three categories that have been used in literature: sample methods based on the modification of the training data, kernel methods based on the modification of the kernel and optimization methods based on the modification of the problem formulation. A recent method, developed for support vector regression, considers prior knowledge on arbitrary regions of the input space. It is exposed here when applied to the classification case. A discussion is then conducted to regroup sample and optimization methods under a regularization framework.  </paper_abstract>
<query_num> 14401 </query_num>
<text>   (Tx, T ′ z) is equivalent to computing the inner product between the averages Φ(x) and Φ(z) of the sets of transformed samples {Φ(Tx)| T ∈ T } and {Φ(T ′ z)| T ′ ∈ T }. 4.2.4 Kernels between sets In =-=[31]-=-, a kernel between sets of vectors is proposed. The idea is to classify the samples defined as sets of d-dimensional vectors xi and now written hal-00021555, version 2 - 24 Apr 2007 χ = {x1, . . .,xi, mples from unknown distributions p and p ′ from a parametric family P. The kernel requires to fit the distributions p and p ′ to the sets as an intermediate step, which ensures permutation-invariance =-=[31]-=-. When P is chosen as the family of multivariate normal distributions N(µ,Σ), p and p ′ are fitted by setting µ and Σ to their maximum likelihood estimates given by the sample mean and the empirical c  proposed to compute the principal angles in feature space using only inner products between columns of the input matrices. This extension allows to introduce an additional kernel as in the method of =-=[31]-=- in order to deal with non-linear cases without requiring to explicitly compute the feature map. 4.2.5 Knowledge-driven kernel selection All previously described methods involving a modification of th . to build a classifier separating sets of vectors (here in matrix form) that incorporates permutationinvariance (here between rows of matrices). But the approach is rather different from the ones of =-=[31]-=- or [70]. Here, the permutation invariance is not incorporated in the kernel but in the learning machine itself. To do so, a SVM with matrix inputs Z i ∈ R m×d instead of vectors xi is considered by d , on a smaller scale, other forms of invariances are also studied when, for instance, considering structured inputs such as matrices. In this setting invariance to permutations of rows as proposed by =-=[31]-=-, [70] or [56] can be crucial for the problem. The methods proposed by [17], [18] and their extension based on [38] allow to include some class-invariance knowledge on regions of the input space, whic   </text>
<query_num> 14402 </query_num>
<text>   . For multi-class applications, there also exists learning machines that can tackle directly multi-class applications such as neural networks [4,33] or even multi-class support vector machines (MSVM) =-=[69,21,8]-=-. Nonetheless, a very common approach consists in building a set of binary classifiers, each one either trained to separate one class from the others (the one-against-all method) or only to distinguis   </text>
<query_num> 14403 </query_num>
<text>   In this framework, class-invariance inside polyhedral regions has been 22introduced for linear classification in [17] and was then extended to the nonlinear case via a reformulation of the kernel in =-=[18]-=-. These two methods are regrouped under the name Knowledge-Based Linear Programming (KBLP). The learning machine considered here uses the linear programming formulation (13). Assuming as prior knowled training on the data and the learning of the prior knowledge. hal-00021555, version 2 - 24 Apr 2007 Similar constraints can be derived for prior knowledge on negative samples and added to the problem =-=[18]-=-. As prior knowledge on a polyhedral set only requires the addition of a set of linear constraints, knowledge on many regions for the two classes can be easily combined and included to the problem. It nce, considering structured inputs such as matrices. In this setting invariance to permutations of rows as proposed by [31], [70] or [56] can be crucial for the problem. The methods proposed by [17], =-=[18]-=- and their extension based on [38] allow to include some class-invariance knowledge on regions of the input space, which might be interesting if, for instance, these regions lack training samples. hal   </text>
<query_num> 14404 </query_num>
<text>   a similarity measure between a sample and such an object. Here, tangent vectors can be used to locally approximate the transformations [45] and lead to the Tangent Vector Kernels (TVK) introduced in =-=[44]-=-. 4.2.3 Haar-integration kernels Haar-integration has been introduced in [54] for the construction of invariant features. In a similar approach, Haar-integration has been used to generate invariant ke   </text>
<query_num> 14405 </query_num>
<text>   allowing to incorporate prior knowledge on a property of the function to estimate (usually the smoothness). But the implicit inclusion of regularization in SVMs, equivalent to regularization networks =-=[59,14]-=-, might explain why few articles studied the application of regularization techniques for the incorporation of other forms of prior knowledge into SVMs. However, in the case of classification, the add   </text>
<query_num> 14406 </query_num>
<text>   also result in a linear program. This is now presented, since some of the methods exposed in the next section for the incorporation of prior knowledge use this form of SVMs. Following the approach of =-=[37]-=-, the ℓ1-norm of the parameters α in (12) is minimized instead of the ℓ2-norm of the weights w as in (8). In practice, to yield a linear program, a new set of variables a bounding the ℓ1-norm of the p t margin SVM is s.t. min (α,b,ξ,a) 1T a + C1 T ξ D(KDα + b1) ≥ 1 − ξ −a ≤ α ≤ a ξ ≥ 0 (13) In this formulation, no assumption on the symmetry or positive definiteness of the kernel matrix K is needed =-=[37]-=-. The form of the resulting output function (12) remains unchanged. Here, the sparsity is enforced by the minimization of the ℓ1-norm of the parameters α which makes some αi vanish to zero. It has als   </text>
<query_num> 14407 </query_num>
<text>   ductive learning”. A transductive learning for SVMs, in which the prior knowledge takes the form of the number num+ of positive samples in the test set, has also been proposed for text classification =-=[26]-=-. In this scheme, the test samples are assigned a misclassification cost C ∗ which is different of the one used for the training samples and is further refined for each class as C ∗ + and C∗ − in acco   </text>
<query_num> 14408 </query_num>
<text>   e kernel amongst admissible kernels with respect to prior knowledge on the imbalance of the training set. 4.2.1 Jittering kernels Jittering kernels were first developed for kernel k-Nearest-Neighbors =-=[10]-=- and then presented for the incorporation of transformation-invariance into SVMs [11]. This approach is related to the virtual SV (VSV) method (see Sect. 4.1.1). Instead of considering an extended tra   </text>
<query_num> 14409 </query_num>
<text>   efined by minimizing a distance between sets of transformed samples. When the transformation is approximated by a Taylor expansion, this method is analogous to the tangent distance method of [57]. In =-=[45]-=-, these concepts are considered under the name of object-to-object distance, where an object corresponds to the set of all transformed samples Px or Pz. Sample-to-object distance is also considered, i  kernels, these latter requiring to test every jittered form of the center. Therefore, the sample-to-object method can be faster but introduces restrictions on the class of admissible transformations =-=[45]-=-. Objects can also be coded by local distributions centered at the samples. In this case they are called soft-objects. One has then to define a similarity measure between a sample and such an object.   </text>
<query_num> 14410 </query_num>
<text>   en proposed [30]. Many softwares, usually available on the Internet, have been developed in the last years to speed up the training time or to deal with large data sets such as SVM light [25], libSVM =-=[6,15]-=- and libsvmTL [47], HeroSVM [12,13] or Core Vector Machine (CVM) [61]. 63 Prior knowledge for classification This Section starts by giving a definition of prior knowledge as considered in this paper.   </text>
<query_num> 14411 </query_num>
<text>   nction for this purpose has been considered. Nonetheless, it results in a modification of the kernel rather than a modification of the optimization problem as exposed in the following. The authors of =-=[51]-=- incorporated local invariance in the sense of (15) and proposed invariant kernels. Defining the tangent vectors by dxi = ∂ ∣ ∂θ ∣ θ=0 Tθxi (36) hal-00021555, version 2 - 24 Apr 2007 allows to include ariance matrix of the tangent vectors i=1 ( N∑ Cγ = (1 − γ)I + γ dxidx i=1 T i ) 1 2 (39) Then, minimizing the regularized cost (38) under the original constraints (4) leads to a standard SVM problem =-=[51]-=-, yielding the output function N∑ f(x) = αiyi〈C i=1 −1 γ xi, C −1 γ x〉 + b (40) Thus a linear invariant SVM is equivalent to a standard SVM where the input is first transformed via the linear mapping   </text>
<query_num> 14412 </query_num>
<text>   o support vector learning by the addition of constraints to the optimization problem. In this framework, class-invariance inside polyhedral regions has been 22introduced for linear classification in =-=[17]-=- and was then extended to the nonlinear case via a reformulation of the kernel in [18]. These two methods are regrouped under the name Knowledge-Based Linear Programming (KBLP). The learning machine c h, for a domain of dimension n, B+ ∈ R n×d , x ∈ R d and d+ ∈ R n . The implication (46) can be transformed for a linear SVM in an equivalent system of linear inequalities having the solution u ∈ R n =-=[17]-=- B T +u + w = 0, d T +u − b + 1 ≤ 0, u ≥ 0 (47) hal-00021555, version 2 - 24 Apr 2007 For prior knowledge on negative samples (y = −1), we have which is equivalent to B−x ≤ d− ⇒ w T x + b ≤ −1 (48) B   instance, considering structured inputs such as matrices. In this setting invariance to permutations of rows as proposed by [31], [70] or [56] can be crucial for the problem. The methods proposed by =-=[17]-=-, [18] and their extension based on [38] allow to include some class-invariance knowledge on regions of the input space, which might be interesting if, for instance, these regions lack training sample   </text>
<query_num> 14413 </query_num>
<text>   og and quadprog are capable of solving the linear and quadratic programs derived from SVMs. Nonetheless, the scale of the problems led to develop specific methods such as chunking [29], decomposition =-=[41]-=- or its extreme case known as the Sequential Minimal Optimization (SMO) algorithm [42], of which modifications have been proposed [30]. Many softwares, usually available on the Internet, have been dev   </text>
<query_num> 14414 </query_num>
<text>   only used in radial basis kernels. For the Gaussian kernel, this yields ( 2 −ρ(x, z) k(x, z) = exp 2σ2 ) (24) The use of another distance for transformation-invariance has been extensively studied in =-=[57]-=- for neural networks under the name of tangent distance (TD) and originally incorporated in SVMs as TD kernels by [23]. The main idea is to measure the distance not between the samples x and z but bet nce was defined by minimizing a distance between sets of transformed samples. When the transformation is approximated by a Taylor expansion, this method is analogous to the tangent distance method of =-=[57]-=-. In [45], these concepts are considered under the name of object-to-object distance, where an object corresponds to the set of all transformed samples Px or Pz. Sample-to-object distance is also cons   </text>
<query_num> 14415 </query_num>
<text>   support vector learning. For each region a set of constraints is added to the linear program. The method is thus only limited by computational power. 4.3.6 Permutation-invariant SVM (π-SVM) The paper =-=[56]-=- focuses on the issue of Section 4.2.4, i.e. to build a classifier separating sets of vectors (here in matrix form) that incorporates permutationinvariance (here between rows of matrices). But the app (Z i , W)+b) over the training set, since maximizing M would lead to a better separation 25of the training data. The main steps of the proposed procedure to train a permutation-invariant SVM (π-SVM) =-=[56]-=- are as follows: (1) compute the radius R and centroid of the smallest hypersphere enclosing all training data; (2) solve the SVM on the training data composed of matrices; (3) find the permutation of ing matrices accordingly; (5) repeat from step 1 until some criterion is met. hal-00021555, version 2 - 24 Apr 2007 For all these steps, efficient algorithms, of which the description can be found in =-=[56]-=-, exist. The idea here is to permute the rows of training matrices so as to minimize the bound on the generalization error based on the ratio between the radius of the data and the margin. The class o  scale, other forms of invariances are also studied when, for instance, considering structured inputs such as matrices. In this setting invariance to permutations of rows as proposed by [31], [70] or =-=[56]-=- can be crucial for the problem. The methods proposed by [17], [18] and their extension based on [38] allow to include some class-invariance knowledge on regions of the input space, which might be int   </text>
<query_num> 14416 </query_num>
<text>   ture 17space induced by κ. In this case, a regularized estimate of the covariance matrix, involving the computation of eigenvectors in the feature space by Kernel Principal Component Analysis (KPCA) =-=[52]-=-, is used. Another similar approach, independently developed at the same time, can be found in [70], where a positive definite kernel is defined over sets of vectors represented as matrices. Consideri  current research aim at building invariant kernels. One reason is that these kernels may also be directly applied to other kernel-based classifiers such as Kernel Principal Component Analysis (KPCA) =-=[52]-=- or Kernel Fisher Discriminant Analysis (KFDA) [39]. 26Also, most of the work focus on one particular type of prior knowledge: classinvariance to a transformation of the input. This can be explained   </text>
</top>
<top>
<paper_num> 145 </paper_num>
<paper_title>   Improved Competitive Ratios for Submodular Secretary Problems (Extended Abstract).  </paper_title>
<paper_abstract>   Abstract. The Classical Secretary Problem was introduced during the 60’s of the 20 th century, nobody is sure exactly when. Since its introduction, many variants of the problem have been proposed and researched. In the classical secretary problem, and many of its variant, the input (which is a set of secretaries, or elements) arrives in a random order. In this paper we apply to the secretary problem a simple observation which states that the random order of the input can be generated by independently choosing a random continuous arrival time for each secretary. Surprisingly, this simple observation enables us to improve the competitive ratio of several known and studied variants of the secretary problem. In addition, in some cases the proofs we provide assuming random arrival times are shorter and simpler in comparison to existing proofs. In this work we consider three variants of the secretary problem, all of which have the same objective of maximizing the value of the chosen set of secretaries given a monotone submodular function f. In the first variant we are allowed to hire a set of secretaries only if it is an independent set of a given partition matroid. The second variant allows us to choose any set of up to k secretaries. In the last and third variant, we can hire any set of secretaries satisfying a given knapsack constraint. 1  </paper_abstract>
<query_num> 14501 </query_num>
<text>   . Table 1 summarizes the above results. 1.2 Related Work Many variants of CS have been considered throughout the years and we shall mention here only those most relevant to this work. Babaioff et al. =-=[4]-=- considered the case where the chosen subset of secretaries needs to be an independent set of a given matroid, and the objective function f is linear. They provided a competitive ratio of Ω(log −1 r)  0903 [5] 0.368 [3] SKS 0.0184 0.0104 [5] 0.0368 [3] 1 For linear objective functions one can apply the algorithm for the classical secretary problem to each subset of secretaries independently.known =-=[4, 9, 18, 20]-=-. The special variant of SCS where the objective function is linear has also been studied. Two incomparable competitive ratios were obtained by Babaioff et al. [3] and Kleinberg [19], achieving compet   </text>
<query_num> 14502 </query_num>
<text>   fined above. Remark: The probability that two secretaries arrive at the same time is 0, thus we ignore this event. The following theorem is used occasionally in our proofs. Similar theorems appear in =-=[12, 11]-=-. 4 If the input is given as a random permutation, the size n of S is assumed to be known. Note that in order to generate the random arrival times of the secretaries and assign them upon arrival, n ha   </text>
<query_num> 14503 </query_num>
<text>   gning and analyzing algorithms for submodular variants of CS. For SCS we present a competitive ratio of (e − 1)/(e 2 + e) ≈ 0.170, and the current best result for this problem is due to Bateni et al. =-=[5]-=- who provided a (1 − e −1 )/7 ≈ 0.0903 competitive ratio. There are two points to notice when comparing our result and that of [5]. First, [5] did not optimize the competitive ratio analysis of their   still obtain improved competitive ratios, though the improvement is smaller than stated above. Second, the algorithm presented in this paper for SCS can be seen as a continuous time “counterpart” of =-=[5]-=-’s algorithm. However, our analysis is simpler than the analysis presented in [5], and also enables us to provide improved competitive ratios. For SKS we provide a competitive ratio of (20e) −1 ≈ 0.01  the best known competitive ratio for the linear version of SKS is only 10e −1 ≈ 0.0368 [3]. As before, the algorithm presented in this paper for SKS can be seen as a continuous time “counterpart” of =-=[5]-=-’s algorithm. However, our analysis is simpler than the analysis presented in [5], enabling us to provide improved competitive ratios. Table 1 summarizes the above results. 1.2 Related Work Many varia sults for the monotone submodular and linear variants of the problems we consider. Problem Our Result Previous Result Best Result for Linear Variant SPMS 0.153 0.0000555 [17] 0.368 1 SCS 0.170 0.0903 =-=[5]-=- 0.368 [3] SKS 0.0184 0.0104 [5] 0.0368 [3] 1 For linear objective functions one can apply the algorithm for the classical secretary problem to each subset of secretaries independently.known [4, 9, 1 7] provide a competitive ratio of Ω(log −1 r) (where r is the rank of the matroid). If the constraint is that the chosen subset of secretaries belongs to the intersection of ℓ matroids, Bateni et al. =-=[5]-=- provide a competitive ratio of Ω(ℓ −1 log −2 r). If the objective function is submodular and monotone, the special case of a partition matroid is exactly SPMS, and the special case of a uniform matro   </text>
<query_num> 14504 </query_num>
<text>   ion is submodular. In recent years many new results in this field have been achieved. The most basic problem, in this field, is that of unconstrained maximization of a nonmonotone submodular function =-=[12, 15]-=-. Other works consider the maximization of a nonmonotone submodular function under various combinatorial constraints [17, 22, 25]. The maximization of a monotone submodular function under various comb   </text>
<query_num> 14505 </query_num>
<text>   ion is submodular. In recent years many new results in this field have been achieved. The most basic problem, in this field, is that of unconstrained maximization of a nonmonotone submodular function =-=[12, 15]-=-. Other works consider the maximization of a nonmonotone submodular function under various combinatorial constraints [17, 22, 25]. The maximization of a monotone submodular function under various comb fined above. Remark: The probability that two secretaries arrive at the same time is 0, thus we ignore this event. The following theorem is used occasionally in our proofs. Similar theorems appear in =-=[12, 11]-=-. 4 If the input is given as a random permutation, the size n of S is assumed to be known. Note that in order to generate the random arrival times of the secretaries and assign them upon arrival, n ha   </text>
<query_num> 14506 </query_num>
<text>   is that of unconstrained maximization of a nonmonotone submodular function [12, 15]. Other works consider the maximization of a nonmonotone submodular function under various combinatorial constraints =-=[17, 22, 25]-=-. The maximization of a monotone submodular function under various combinatorial constraints (such as a matroid, the intersection of several matroids and knapsack constraints) has also been widely stu   </text>
<query_num> 14507 </query_num>
<text>   maximization of a monotone submodular function under various combinatorial constraints (such as a matroid, the intersection of several matroids and knapsack constraints) has also been widely studied =-=[7, 8, 21, 23, 25]-=-. Thus, it comes as no surprise that recent works have combined the secretary problem with submodular optimization. Gupta et al. [17] were the first to consider this combination. For the variant where   </text>
<query_num> 14508 </query_num>
<text>   oducts of a slots’ weights with the values of the secretaries assigned to them. Babaioff et al. [2] provide a competitive ratio of 1/4 for this special variant. Additional variants of CS can found in =-=[1, 6, 13, 14, 16]-=-. Another rich field of study is that of submodular optimization, namely optimization problems in which the objective function is submodular. In recent years many new results in this field have been a   </text>
<query_num> 14509 </query_num>
<text>   of secretaries a different weight. The value of the objective function in this case is the sum of the products of a slots’ weights with the values of the secretaries assigned to them. Babaioff et al. =-=[2]-=- provide a competitive ratio of 1/4 for this special variant. Additional variants of CS can found in [1, 6, 13, 14, 16]. Another rich field of study is that of submodular optimization, namely optimiza   </text>
</top>
<top>
<paper_num> 146 </paper_num>
<paper_title>   HPF-2 Support for Dynamic Sparse Computations.  </paper_title>
<paper_abstract>   . There is a class of sparse matrix computations, such as direct solvers of systems of linear equations, that change the fill-in (nonzero entries) of the coefficient matrix, and involve row/column operations (pivoting) . This paper addresses the problem of the parallelization of these sparse computations from the point of view of the parallel language and the compiler. Dynamic data structures for sparse matrix storage are analyzed, permitting to efficiently deal with fill-in and pivoting issues. Any of the data representations considered enforces the handling of indirections for data accesses, pointer referencing and dynamic data creation. All of these elements go beyond current data-parallel compilation technology. Our solution is to propose a small set of new extensions to HPF-2 to parallelize these codes, and to support part of the new capabilities on a runtime library. This approach has been evaluated on a Cray T3E, implementing, in particular, the sparse LU factorization. 1 Introd...  </paper_abstract>
<query_num> 14601 </query_num>
<text>   DDLY calls). Fig. 10 shows execution times and speed-up for the parallel LU algorithm. Test sparse matrices were taken from the Harwell-Boeing suite and University of Florida Sparse Matrix Collection =-=[9]-=- (see Table 1). The efficiency of the parallel code is high when the size of the input matrix is significantly large. We also carried out experiments considering meshes of processors instead of linear   </text>
<query_num> 14602 </query_num>
<text>   e independences on the loops traversing rows and columns [3]. Coarser parallelism level can be exploited thanks to the elimination tree, which can be used to schedule parallel tasks in a multifrontal =-=[14]-=- code. It is also possible to use a coarse matrix decomposition to obtain an ordering to bordered block triangular form, as is done in the MCSPARSE package [17]. The supernodal [11] approach is also a   </text>
<query_num> 14603 </query_num>
<text>   e programmer-defined data distributions, partition loop iterations, remap data and generate optimized communication schedules. Most of these solutions are based on the inspector-executor paradigm [22]=-=[8]-=-. Current language constructs and the supportive runtime libraries are insufficiently developed, leading to low efficiencies when they are applied to a wide set of irregular codes. In the context of s   </text>
<query_num> 14604 </query_num>
<text>   extensively tested a number of pseudo-regular distribution schemes for sparse problems, which combines natural extensions of regular data distributions with compressed data storages [2] [4] [23] [25] =-=[26]-=-. These distribution schemes can be incorporated to a data-parallel language (HPF) in a simple way. The programmer can use them easily and obtain high efficiencies from the parallelization of irregula   </text>
<query_num> 14605 </query_num>
<text>   hedule parallel tasks in a multifrontal [14] code. It is also possible to use a coarse matrix decomposition to obtain an ordering to bordered block triangular form, as is done in the MCSPARSE package =-=[17]-=-. The supernodal [11] approach is also a parallelizable code [16]. Some of the above parallel solutions can be implemented using the approach described in this paper. Loop-level LU approaches can be i   </text>
<query_num> 14606 </query_num>
<text>   loped and extensively tested a number of pseudo-regular distribution schemes for sparse problems, which combines natural extensions of regular data distributions with compressed data storages [2] [4] =-=[23]-=- [25] [26]. These distribution schemes can be incorporated to a data-parallel language (HPF) in a simple way. The programmer can use them easily and obtain high efficiencies from the parallelization o we can use the SPARSE directive to specify that a sparse matrix (or sparse array) is stored using a particular linked list scheme. This directive was previously introduced, for instance in [2] and in =-=[23]-=-, in the context of static sparse applications. Fig. 5 shows the BNF syntax for the dynamic SPARSE directive. The first two data structures, LLRS and LLCS, are defined by two arrays of pointers (!poin ach, however, is not suitable to the linked list sparse directive, due to the use of different data storage schemes. However, they could be implemented using the basic BCS or BRS sparse distributions =-=[2, 23]-=-. The implementation of the supernodal code in [11] uses some sort of column compressed storage, but it would be necessary to simplify the memory management and the data access patterns to consider a   </text>
<query_num> 14607 </query_num>
<text>   ng on the type of data accesses we have to deal with. To simplify the discussion, let us take our working example application, the LU factorization of a sparse matrix, computed using a general method =-=[1, 13]-=-. These methods solve directly the sparse problem and shares the same loop structure of the corresponding dense code. Fig. 1 shows an in-place code for the direct right-looking LU algorithm, where an  1). The sequential efficiency of the Fortran 90 implementation of the sparse LU algorithm was also tested. Table 2 presents comparison results from this implementation and the Fortran 77 MA48 routine =-=[13]-=-. We observe that the MA48 routine is significantly faster than our algorithm for many matrices, but it should be considered the fact that the Cray Fortran 90 compiler is not efficient generating code   </text>
<query_num> 14608 </query_num>
<text>   sible to use a coarse matrix decomposition to obtain an ordering to bordered block triangular form, as is done in the MCSPARSE package [17]. The supernodal [11] approach is also a parallelizable code =-=[16]-=-. Some of the above parallel solutions can be implemented using the approach described in this paper. Loop-level LU approaches can be implemented using the LLRCS data storage (in addition to LLCS), an   </text>
<query_num> 14609 </query_num>
<text>   tion (the feasibility of its design is not clear), as well as the possibility of mixing in the same code place holders (dense notations) with real arrays. Bik and Wijshoff [7] and Kotlyar and Pingaly =-=[21]-=- propose a similar approach, based on the automatic transformation of a dense program, annotated with sparse directives, into a semantically equivalent sparse code. However, the design of such compile   </text>
</top>
<top>
<paper_num> 147 </paper_num>
<paper_title>   Linearly Bounded Reformulations of Conjunctive Databases.  </paper_title>
<paper_abstract>   Database reformulation is the process of rewriting the data  and rules of a deductive database in a functionally equivalent manner.  </paper_abstract>
<query_num> 14701 </query_num>
<text>   eries using materialized views and conrains references to major results in the areas of query containment and view materialization. Papers [16,17, 21, 30] describe approaches to view materialization. =-=[3, 12, 13, 23]-=- treat the problem of using available materialized views for query evaluation. Nearly all results described in the literature concern rewriting of single queries; notice that the database reformulatio   </text>
<query_num> 14702 </query_num>
<text>   ets of queries. Transformations of database schemas and queries can be considered together as reformulations of logical theories. [31] provides a theoretical foundation for theory reformulations, and =-=[15, 22]-=- contain work on general transformations of logical theories. Descriptions of basic methods used in this paper can be found, e.g., in [14]. 8 Conclusions and Future Work The first contribution of this   </text>
<query_num> 14703 </query_num>
<text>   ies using views. [1] discusses the complexity of answering queries using materialized views and conrains references to major results in the areas of query containment and view materialization. Papers =-=[16,17, 21, 30]-=- describe approaches to view materialization. [3, 12, 13, 23] treat the problem of using available materialized views for query evaluation. Nearly all results described in the literature concern rewri   </text>
<query_num> 14704 </query_num>
<text>   ion capacity was introduced in [19] as a fundamental theoretical concept which encompasses schema equivalence and dominance. Other work on relative information capacity includes [5, 25, 26]. Tutorial =-=[18]-=- surveys a number of frameworks -- relative information capacity among others -- for dealing with the issue of semantic heterogeneity arising in database integration. In practical database systems, da   </text>
<query_num> 14705 </query_num>
<text>   ions such as data warehousing and multidatabase integration have promoted the study of views in databases. The paper [32] is a survey of containment and rewriting/optimization of queries using views. =-=[1]-=- discusses the complexity of answering queries using materialized views and conrains references to major results in the areas of query containment and view materialization. Papers [16,17, 21, 30] desc   </text>
<query_num> 14706 </query_num>
<text>   ods complement schema transformation methods in that they are applied to databases that are already operational. Query rewriting is important for query optimization, especially in deductive databases =-=[27]-=- where queries can be complex and the amount of data accessed can be overwhelming. [28] is a survey on implementation techniques and implemented projects in deductive databases. There is an extensive   </text>
<query_num> 14707 </query_num>
<text>   riting methods in datalog and its extensions see [2, 33, 34]. In addition, applications such as data warehousing and multidatabase integration have promoted the study of views in databases. The paper =-=[32]-=- is a survey of containment and rewriting/optimization of queries using views. [1] discusses the complexity of answering queries using materialized views and conrains references to major results in th   </text>
<query_num> 14708 </query_num>
<text>   semantic heterogeneity arising in database integration. In practical database systems, database design frequently uses normalization, first introduced in [10] and described in detail in [33]. Papers =-=[7, 20]-=- survey methods and issues in multidatabase integration. Query transformation is another aspect of database transformation tasks; query rewriting methods complement schema transformation methods in th   </text>
<query_num> 14709 </query_num>
<text>   ter, relative information capacity was introduced in [19] as a fundamental theoretical concept which encompasses schema equivalence and dominance. Other work on relative information capacity includes =-=[5, 25, 26]-=-. Tutorial [18] surveys a number of frameworks -- relative information capacity among others -- for dealing with the issue of semantic heterogeneity arising in database integration. In practical datab   </text>
</top>
<top>
<paper_num> 148 </paper_num>
<paper_title>   Predicate Abstraction via Symbolic Decision Procedures  </paper_title>
<paper_abstract>   Vol. 3 (2:1) 2007, pp. 1–1–20 www.lmcs-online.org  </paper_abstract>
<query_num> 14801 </query_num>
<text>   etails. The techniques that aim to reduce the number of calls to the theorem prover or decision procedure are mostly based on enumerating cubes over P in an increasing order of their size. Das et al. =-=[DDP99]-=- enumerates cubes over a tree, after fixing the order of predicates 1The dual of this problem, which is to compute the strongest Boolean formula GP(e) that is implied by e, can be expressed as ¬FP(¬e)   </text>
<query_num> 14802 </query_num>
<text>   he CNF representation of the formula can be exponentially large compared to the original formula. However, we can use recent techniques to obtain the CNF form lazily, by a method proposed by McMillan =-=[McM02]-=-. For the rest of hte paepr, we focus here on computing FP( � ei∈E ei) when ei is a predicate. Unless specified otherwise, we always use e to denote ( � ei∈E ei), a disjunction of predicates in the se   </text>
<query_num> 14803 </query_num>
<text>   more times. Consider the set G . = {x &amp;lt; y + 1,y &amp;lt; x − 2,x &amp;lt; 4 Constraints like x ⊲⊳ c are handled by adding a special variable x0 to denote the constant 0, and rewriting the constraint as x ⊲⊳ x0 + c =-=[SSB02]-=-. (d) (e)sPREDICATE ABSTRACTION VIA SYMBOLIC DECISION PROCEDURES ∗ 15 z − 1,...}. In this case we can produce the fact y &amp;lt; z − 3 from y &amp;lt; x − 2,x &amp;lt; z − 1 and then x &amp;lt; z − 2 from y &amp;lt; z − 3,x &amp;lt; y + 1. T   </text>
<query_num> 14804 </query_num>
<text>   od to construct SDP for a combination of two simple theories T1 ∪ T2 (including EUF + DIF), by using an extension of the Nelson-Oppen combination [NO80] method. We use Binary Decision Diagrams (BDDs) =-=[Bry86]-=- to construct FP(e) from the shared representations efficiently in practice. We present a preliminary evaluation of our procedure on predicate abstraction benchmarks from device driver verification in   </text>
<query_num> 14805 </query_num>
<text>   on failed proof attempts in the verification tool. Das and Dill [DD01] and subsequently Ball et al. [BCDR04] use counterexamples to refine the predicate abstraction incrementally. Jhala and McMillan =-=[JM05]-=- use interpolants to refine the predicate abstraction. It is not clear if it is always preferable to compute the abstraction incrementally. But, we have observed that the refinement loop can often bec   </text>
<query_num> 14806 </query_num>
<text>   on is a technique for automatically creating finite abstract models of finite and infinite state systems [GS97]. The method has been widely used in abstracting finite-state models of programs in SLAM =-=[BMMR01]-=- and numerous other software verification projects [HJMS02, CCG + 04]. It has also been used for synthesizing loop invariants [FQ02] and verifying distributed protocols [DDP99, LBC03]. The fundamental  by the formula e. Their technique requires |P |.2 |P | theorem prover calls in the worst case. Other techniques sacrifice precision to gain efficiency, by only considering cubes of some fixed length =-=[BMMR01]-=-. All these techniques may require an exponential number of theorem prover calls in the worst case, and demonstrate worst case behavior in practice. However, more importantly, since these queries are  set of software verification benchmarks. The benchmarks are generated from the predicate abstraction step for constructing Boolean Programs from C programs of Microsoft Windows device drivers in SLAM =-=[BMMR01]-=-. We compare our method with two other methods for performing predicate abstraction: : DP-based: This method uses the decision procedure zapato [BCLZ04] to enumerate the set of cubes that imply e. Var   </text>
<query_num> 14807 </query_num>
<text>   produce compact shared expressions. We provide a method to construct SDP for a combination of two simple theories T1 ∪ T2 (including EUF + DIF), by using an extension of the Nelson-Oppen combination =-=[NO80]-=- method. We use Binary Decision Diagrams (BDDs) [Bry86] to construct FP(e) from the shared representations efficiently in practice. We present a preliminary evaluation of our procedure on predicate ab (f(h(x))) is {x,h(x),f(h(x))}. For a set of predicates G, terms(G) denotes the union of the set of terms in any g ∈ G. A decision procedure for EUF can be obtained by the congruence closure algorithm =-=[NO80]-=-, described in Figure 6. For a set of predicates G, let m = |terms(G)|. We can show that if we iterate the loop in step (2) of DPT(G) (shown in Figure 3) for at least 3m steps, then DPT can implement   </text>
<query_num> 14808 </query_num>
<text>   raction benchmarks from device driver verification in SLAM. 1. Introduction Predicate abstraction is a technique for automatically creating finite abstract models of finite and infinite state systems =-=[GS97]-=-. The method has been widely used in abstracting finite-state models of programs in SLAM [BMMR01] and numerous other software verification projects [HJMS02, CCG + 04]. It has also been used for synthe   </text>
<query_num> 14809 </query_num>
<text>   s to be reset across each call, precluding any learning across calls. Alternately, predicate abstraction can be formulated as a quantifier elimination problem. Lahiri et al. [LBC03] and Clarke et al. =-=[CKSY04]-=- perform predicate abstraction by reducing the problem of computing FP(e) to Boolean quantifier elimination. The former method first transforms a first-order quantifier elimination problem into Boolea   </text>
<query_num> 14810 </query_num>
<text>   set of techniques to avoid computing the most precise abstraction upfront, and refine it only based on failed proof attempts in the verification tool. Das and Dill [DD01] and subsequently Ball et al. =-=[BCDR04]-=- use counterexamples to refine the predicate abstraction incrementally. Jhala and McMillan [JM05] use interpolants to refine the predicate abstraction. It is not clear if it is always preferable to co raction incrementally. But, we have observed that the refinement loop can often becomes the main bottleneck in these techniques (for example in SLAM), and limits the scalability of the overall system =-=[BCDR04]-=-. 2. Setup Figure 1 defines the syntax of a quantifier-free fragment of first-order logic. An expression in the logic can either be a term or a formula. A term can either be a variable or ans4 S. K. L   </text>
<query_num> 14811 </query_num>
<text>   the state of the prover has to be reset across each call, precluding any learning across calls. Alternately, predicate abstraction can be formulated as a quantifier elimination problem. Lahiri et al. =-=[LBC03]-=- and Clarke et al. [CKSY04] perform predicate abstraction by reducing the problem of computing FP(e) to Boolean quantifier elimination. The former method first transforms a first-order quantifier elim conds. increasing order of size) are used to prevent enumerating exponential number of cubes in practice. : UCLID-based: This method performs quantifier-elimination using incremental SATbased methods =-=[LBC03]-=-. The procedure works by first converting the problem into an existential quantifier elimination problem in first-order logic and then reducing it to Boolean quantifier elimination by using an encodin  UCLID-based tool is no longer actively maintained, and we had trouble translating these SLAM benchmarks to input of UCLID. From our earlier experience of using UCLID on similar benchmarks (Fig. 3 in =-=[LBC03]-=-), we believe that most of these benchmarks can be solved within a few seconds, and the total runtime would not differ by more than 2–3X (in favor of the current technique). To compare with UCLID-base   </text>
<query_num> 14812 </query_num>
<text>   to construct a symbolic decision procedure for the combination of saturation theories T1 and T2, given SDP for T1 and T2. The combination is based on an extension of the Nelson-Oppen (N-O) framework =-=[NO79]-=- that constructs a decision procedure for the theory T1 ∪ T2 using the decision procedures of T1 and T2.s16 S. K. LAHIRI, T. BALL, AND B. COOK We assume that the theories T1 and T2 have disjoint signa  are communicated to the other theory. This process is continued until the set ∆ does not change. In this case, the method returns satisfiable. Let us denote this algorithm as DPT1 ∪T2 . Theorem 4.1 (=-=[NO79]-=-). For convex, stably infinite and signature-disjoint theories T1 and T2, DPT1 ∪T2 is a decision procedure for T1 ∪ T2. There can be at most |Vsh| irredundant equalities over Vsh, therefore the N-O lo omputing the (shared) expression t[e] as before, SDPTi also 5We need these restrictions only to exploit the N-O combination result. The definition of convexity and stably infiniteness can be found in =-=[NO79]-=-. 6We assume that each theory has an inference rule for deriving equality between variables in the theory, and DPT also returns a set of equality over variables.sPREDICATE ABSTRACTION VIA SYMBOLIC DEC   </text>
</top>
<top>
<paper_num> 149 </paper_num>
<paper_title>   Prospectus for the Next LAPACK and ScaLAPACK Libraries.  </paper_title>
<paper_abstract>   Dense linear algebra (DLA) forms the core of many scientific computing applications. Consequently, there is continuous interest and demand for the development of increasingly better algorithms in the field. Here ’better ’ has a broad meaning, and includes improved reliability, accuracy, robustness, ease of use, and  </paper_abstract>
<query_num> 14901 </query_num>
<text>   by a ratio of up to .75, but at the possible cost of numerical stability for some singular vectors. 6. When few or no vectors are desired, the bottleneck shifts entirely to phase 1. Bischof and Lang =-=[40]-=- have proposed a Successive Band Reduction (SBR) algorithm that will asymptotically (for large dimension n) change most of the Level 2 BLAS operations in phase 1 to Level 3 BLAS operations. They repor   </text>
<query_num> 14902 </query_num>
<text>   e. 2. There is extensive research in other high performance programming languages funded by DOE, NSA and the DARPA HPCS program. UPC [95], Titanium [96], CAF [97], Fortress [98], X10 [99] and Cascade =-=[100]-=- are all languages under active development, and are being designed for high productivity SWE on high peformance machines. It is natural to ask if these are appropriate programming languages for ScaLA   </text>
<query_num> 14903 </query_num>
<text>   ent-wise, relative error bound, i.e. a bound on the number of correct digits in each component. Both error bounds can be computed for a tiny O(n 2 ) extra cost after the initial O(n 3 ) factorization =-=[16]-=-. 2. As mentioned in sec. 3, there can be a large speed difference between different floating point units on the same processor, with single precision running 10x faster than double precision on an IB 1). 5. The new Sca/LAPACK routines will be converted to use the latest BLAS standard [88, 4, 12], which also provides new high precision functionality necessary for new routines required in section 4 =-=[12, 16]-=-, systematically ensure thread-safety, and deprecate superseded routines. 6. Appropriate tools will be used (eg., autoconf, bugzilla, svn, automatic overnight build and test, etc.) to streamline insta   </text>
<query_num> 14904 </query_num>
<text>   gress (what fraction of peak is reached?) and limit the search space. 2. There is extensive research in other high performance programming languages funded by DOE, NSA and the DARPA HPCS program. UPC =-=[95]-=-, Titanium [96], CAF [97], Fortress [98], X10 [99] and Cascade [100] are all languages under active development, and are being designed for high productivity SWE on high peformance machines. It is nat   </text>
<query_num> 14905 </query_num>
<text>   he development of out-of-core versions of matrix factorizations. ScaLAPACK prototypes [49] are under development that implement out-of-core data management for the LU, QR, and Cholesky factorizations =-=[77, 78]-=-. Users have asked for two kinds of parallel I/O: to a single file from a sequential LAPACK program (possible with sequential I/O in the reference implementation), and to a single file from MPI-based   </text>
<query_num> 14906 </query_num>
<text>   iagonal has a rank bounded by a small constant. Recent research has focused on methods exploiting semiseparability, or being a sum of a banded matrix and a semiseparable matrix, for better efficiency =-=[52, 53]-=-. The development of such algorithms is being considered in a future release. Most exciting are the recent observations of Gu, Bini and others [54] that a companion matrix is banded plus semiseparable   </text>
<query_num> 14907 </query_num>
<text>   languages will be provided, depending instead on interoperability of these languages with the F95 subset used in (1). 5. The new Sca/LAPACK routines will be converted to use the latest BLAS standard =-=[88, 4, 12]-=-, which also provides new high precision functionality necessary for new routines required in section 4 [12, 16], systematically ensure thread-safety, and deprecate superseded routines. 6. Appropriate   </text>
<query_num> 14908 </query_num>
<text>   o exhaustive and then choose different levels of search effort for different routines depending on which are more important. Sophisticated data modeling techniques may be used to minimize search time =-=[94]-=-. Beyond these development activities, the following research tasks are being performed, which should influence and improve the development. 1. As discussed in section 3 emerging architectures offer p l parameter values in a large search space. A number of users have already expressed concern about installation time and difficulty, when this additional tuning may occur. Based on earlier experience =-=[94]-=-, it is possible to use statisticalsProspectus for LAPACK and ScaLAPACK 19 models to both limit the search space and more accurately determine the optimal parameters at run time. 5. In a dynamically c   </text>
<query_num> 14909 </query_num>
<text>   quad, ...} to one or two. 7. Performance tuning will be done systematically. Initial experiments are showing up to 10x speedups using different communication schemes that can be applied in the BLACS =-=[89, 90]-=-. In addition to tuning the BLAS [91, 92] and BLACS, there are over 1300 calls to the ILAENV routine in LAPACK, many of which return tuning parameters that have never been systematically tuned. Some o   </text>
<query_num> 14910 </query_num>
<text>   ra searching and pivoting may involve nonnegligible communication costs. 6. Progress has been made in the development of new algorithms for computing or estimating the condition number of tridiagonal =-=[24, 25]-=- or triangular matrices [26]. These algorithms play an important role in obtaining error bounds in matrix factorizations, and the most promising algorithms should be evaluated and incorporated in the   </text>
<query_num> 14911 </query_num>
<text>   t the most effective optimizing compilers still work best on Fortran (even when they share “back ends” with the C compiler, because of the added difficulty of discerning the absence of aliasing in C) =-=[87]-=-. The F95 features adopted include recursion (to support new matrix data structures and associated algorithms discussed in section 4), modules (to support production of versions for different precisio   </text>
<query_num> 14912 </query_num>
<text>   tion. LU and other matrix factorization have left-looking and right-looking formulations [101]. It has even been observed that transition between the two can be done by automatic code transformations =-=[102]-=-, although more powerful methods than simple dependency analysis is necessary. It is known that lookahead can be used to improve performance, by performing panel factorizations in parallel with the up   </text>
<query_num> 14913 </query_num>
<text>   tuning will be done systematically. Initial experiments are showing up to 10x speedups using different communication schemes that can be applied in the BLACS [89, 90]. In addition to tuning the BLAS =-=[91, 92]-=- and BLACS, there are over 1300 calls to the ILAENV routine in LAPACK, many of which return tuning parameters that have never been systematically tuned. Some of these parameters are block sizes for bl   </text>
<query_num> 14914 </query_num>
<text>   volve nonnegligible communication costs. 6. Progress has been made in the development of new algorithms for computing or estimating the condition number of tridiagonal [24, 25] or triangular matrices =-=[26]-=-. These algorithms play an important role in obtaining error bounds in matrix factorizations, and the most promising algorithms should be evaluated and incorporated in the future release. 4.2 Algorith   </text>
</top>
<top>
<paper_num> 150 </paper_num>
<paper_title>   A general agnostic active learning algorithm.  </paper_title>
<paper_abstract>   We present a simple, agnostic active learning algorithm that works for any hypothesis class of bounded VC dimension, and any data distribution. Our algorithm extends a scheme of Cohn, Atlas, and Ladner to the agnostic setting, by (1) reformulating it using a reduction to supervised learning and (2) showing how to apply generalization bounds even for the non-i.i.d. samples that result from selective sampling. We provide a general characterization of the label complexity of our algorithm. This quantity is never more than the usual PAC sample complexity of supervised learning, and is exponentially smaller for some hypothesis classes and distributions. We also demonstrate improvements experimentally.  </paper_abstract>
<query_num> 15001 </query_num>
<text>   ask. In such cases, reduction-based active learning algorithms can be relatively efficient. On the other hand, agnostic supervised learning is computationally intractable for many hypothesis classes (=-=Guruswami and Raghavendra, 2006-=-), and of course, agnostic active learning is at least as hard in the worst case. Our reduction to supervised learning is benign in the sense that the learning problems we need to solve are over sampl   </text>
<query_num> 15002 </query_num>
<text>   m 1 only needs Õ(θd log2 (1/ε)) labels to achieve error ε ≈ ν and Õ(θd(log 2 (1/ε) + (ν/ε) 2 )) labels to achieve error ε ≪ ν. The latter matches the dependence on ν/ε in the Ω((ν/ε) 2 ) lower bound (=-=Kääriäinen, 2006-=-). The linear dependence on θ improves on the quadratic dependence shown for A 2 (Hanneke, 2007) 4 . For an illustrative consequence of this, suppose DX is the uniform distribution on the sphere in R   </text>
<query_num> 15003 </query_num>
<text>   mension vcdim(H) = d &amp;lt; ∞. Recall that the nth shattering coefficient S(H,n) is defined as the maximum number of ways in which H can label a set of n points; by Sauer’s lemma, this is at most O(n d ) (=-=Bousquet, Boucheron, and Lugosi, 2004, p.175-=-). We denote by DX the marginal of D over X . In our active learning model, the learner receives unlabeled data sampled from DX ; for any sampled point x, it can optionally request the label y  ction. We assume for simplicity that the minimal error ν = inf{errD(h) : h ∈ H} is achieved by a hypothesis h ∗ ∈ H. Our algorithm and analysis use the following normalized uniform convergence bound (=-=Bousquet, Boucheron, and Lugosi, 2004, p.200-=-). Lemma 1 (=-=Vapnik and Chervonenkis (1971)-=-). Let F be a family of measurable functions f : Z → {0,1} over a space Z. Denote by EZf the empirical average of f over a subset Z ⊂ Z. Let αn = √ (4/   </text>
<query_num> 15004 </query_num>
<text>   ound. We use a normalized bound that takes into account the empirical error (computed on ˆ S ∪ T —again, not an i.i.d. sample) of the hypothesis in question. Earlier work on agnostic active learning (=-=Balcan, Beygelzimer, and Langford, 2006; Hanneke, 2007-=-) has been able to upper bound label complexity in terms of a parameter of the hypothesis class (and data distribution) called the disagreement coefficient. We give label complexity bou essive: rather than seeking out points that are maximally informative, it queries every point that it is somewhat unsure about. The early work of Cohn-Atlas-Ladner (1994) and the recent A2 algorithm (=-=Balcan, Beygelzimer, and Langford, 2006-=-) are similarly mellow in their querying strategy. The label complexity improvements achievable by such algorithms are nicely captured by a parameter called the disagreement coefficient, introduced re   </text>
</top>
<top>
<paper_num> 151 </paper_num>
<paper_title>   Multi-Label Classification: An Overview.  </paper_title>
<paper_abstract>   Nowadays, multi-label classification methods are increasingly required by modern applications, such as protein function classification, music categorization and semantic scene classification. This paper introduces the task of multi-label classification, organizes the sparse related literature into a structured presentation and performs comparative experimental results of certain multi-label classification methods. It also contributes the definition of concepts for the quantification of the multi-label nature of a data set.  </paper_abstract>
<query_num> 15101 </query_num>
<text>   ( p( c ) log p( c ) + q( c ) log q( c ) ) i where p(ci) = relative frequency of class ci and q(ci) = 1−p(ci). They also allowed multiple labels in the leaves of the tree. Adaboost.MH and Adaboost.MR (=-=Schapire &amp; Singer, 2000-=-) are two extensions of AdaBoost (=-=Freund &amp; Schapire, 1997-=-) for multi-label classification. They both apply AdaBoost on weak classifiers of the form H:X ×L → R. In AdaBoost.MH if the sign of the output   </text>
<query_num> 15102 </query_num>
<text>   are increasingly required by modern applications, such as protein function classification (=-=Zhang &amp; Zincir-Heywood, 2005-=-), musicscategorization (=-=Li &amp; Ogihara, 2003-=-) and semantic scene classification (=-=Boutell et al., 2004-=-). In semantic scene classification, a photograph can belong to more than one conceptual class, such as sunsets and beaches at the same time. Similarly, in music categorization a song may belong to mo Ex. Sports Religion Science Politics 1 X X 2 X X 3 X 4 X X There exist two straightforward problem transformation methods that force the learning problem into traditional single-label classification (=-=Boutell et al., 2004-=-). The first one (dubbed PT1) subjectively or randomly selects one of the multiple labels of each multi-label instance and discards the rest, while the second one (dubbed PT2) simply discards every mu data set of Table 1 using this method. One of the negative aspects of PT3 is that it may lead to data sets with a large number of classes and few examples per class. PT3 has been used in the past in (=-=Boutell et al., 2004; Diplaris et al., 2005-=-). Table 4: Transformed data set using PT3 Ex. Sports (Sports ∧ Politics) (Science ∧ Politics) (Science ∧ Religion) 1 X 2 X 3 X 4 X The most common problem transformation method y the |L| classifiers: ( x) { l} : H ( x) = l H PT 4 = U l l∈L Figure 1 shows the four data sets that are constructed by PT4 when applied to the data set of Table 1. PT4 has been used in the past in (=-=Boutell et al., 2004; Goncalves &amp; Quaresma, 2003; Lauser &amp; Hotho, 2003; Li &amp; Ogihara, 2003-=-).sFigure 1: The four data sets that are constructed by PT4 A straightforward, yet undocumented, problem transformation method is   =-=et al., 2005-=-) and yeast (=-=Elisseeff &amp; Weston, 2002-=-) are biological data sets that are concerned with protein function classification and gene function classification respectively. The scene data set (=-=Boutell et al., 2004-=-) contains data related to a scene classification problem. These data sets were retrieved from the site of the Support Vector Classification library LIBSVM (=-=Chang &amp; Lin, 2001-=-), and transformed to a sp   </text>
<query_num> 15103 </query_num>
<text>   hrough this extension the approach takes into consideration the potential dependencies among the different labels. Note here that this improvement is actually a specialized case of applying Stacking (=-=Wolpert, 1992-=-) (a method for the combination of multiple classifiers) on top of PT4. The second improvement of (=-=Godbole &amp; Sarawagi, 2004-=-) is SVM-specific and concerns the margin of SVMs in multi-label classificati   </text>
<query_num> 15104 </query_num>
<text>   relative frequency of class ci and q(ci) = 1−p(ci). They also allowed multiple labels in the leaves of the tree. Adaboost.MH and Adaboost.MR (=-=Schapire &amp; Singer, 2000-=-) are two extensions of AdaBoost (=-=Freund &amp; Schapire, 1997-=-) for multi-label classification. They both apply AdaBoost on weak classifiers of the form H:X ×L → R. In AdaBoost.MH if the sign of the output of the weak classifiers is positive for a new example x   </text>
<query_num> 15105 </query_num>
<text>   s. Note here that this improvement is actually a specialized case of applying Stacking (=-=Wolpert, 1992-=-) (a method for the combination of multiple classifiers) on top of PT4. The second improvement of (=-=Godbole &amp; Sarawagi, 2004-=-) is SVM-specific and concerns the margin of SVMs in multi-label classification problems. They improve the margin by a) removing very similar negative training instances which are within a threshold d  defined as: 1 HammingLoss(H, D) = ∑ = D D i 1 YiΔZ L Where Δ stands for the symmetric difference of two sets and corresponds to the XOR operation in Boolean logic. The following metrics are used in (=-=Godbole &amp; Sarawagi, 2004-=-) for the evaluation of H on D: 1 Accuracy(H, D) = ∑ = 1 Precision(H, D) = ∑ = 1 Recall(H, D) = ∑ = Y I Z D i i D i 1 Yi U Z i Y I Z D i D i 1 Z i Y I Z D i D i 1 Yi i i isBoutell et al. (2004) give a   </text>
<query_num> 15106 </query_num>
<text>   sd.auth.gr/multilabel.html). We experimented with the three PT methods in conjunction with the following classifier learning algorithms: kNN (=-=Aha, Kibler &amp; Albert-=-), C4.5 (=-=Quinlan, 1993-=-), Naive Bayes (=-=John &amp; Langley, 1995-=-) and SMO (=-=Platt, 1998-=-). For performance evaluation, we used the HammingLoss, Accuracy, Precision and Recall metrics that were presented in the previous section. We experimented on the following multi   </text>
</top>
<top>
<paper_num> 152 </paper_num>
<paper_title>   Kernel Methods for Deep Learning.  </paper_title>
<paper_abstract>   We introduce a new family of positive-definite kernel functions that mimic the computation in large, multilayer neural nets. These kernel functions can be used in shallow architectures, such as support vector machines (SVMs), or in deep kernel-based architectures that we call multilayer kernel machines (MKMs). We evaluate SVMs and MKMs with these kernel functions on problems designed to illustrate the advantages of deep architectures. On several problems, we obtain better results than previous, leading benchmarks from both SVMs with Gaussian kernels as well as deep belief nets. 1  </paper_abstract>
<query_num> 15201 </query_num>
<text>   (MKMs) perform very competitively on multiclass data sets designed to foil shallow architectures [11]. 53.1 Multilayer kernel machines We explored how to train MKMs in stages that involve kernel PCA =-=[13]-=- and feature selection [14] at intermediate hidden layers and large-margin nearest neighbor classification [15] at the final output layer. Specifically, for ℓ-layer MKMs, we considered the following t rom the above procedure provide a first proof-of-concept. We discuss each of these steps in greater detail below. Kernel PCA. Deep learning in MKMs is achieved by iterative applications of kernel PCA =-=[13]-=-. This use of kernel PCA was suggested over a decade ago [16] and more recently inspired by the pretraining of deep belief nets by unsupervised methods. In MKMs, the outputs (or features) from kernel   </text>
<query_num> 15202 </query_num>
<text>   ) n The integral representation makes it straightforward to show that these kernel functions are positivesemidefinite. The kernel function in eq. (1) has interesting connections to neural computation =-=[8]-=- that we explore further in sections 2.2–2.3. However, we begin by elucidating its basic properties. 2.1 Basic properties We show how to evaluate the integral in eq. (1) analytically in the appendix.  (x) · f(y) = m∑ Θ(wi · x)Θ(wi · y)(wi · x) n (wi · y) n , (8) i=1 where m is the number of output units. The connection with the arc-cosine kernel function emerges in the limit of very large networks =-=[10, 8]-=-. Imagine that the network has an infinite number of output units, and that the weights Wij are Gaussian distributed with zero mean and unit variance. In this limit, we see that eq. (8) reduces to eq. (x) · f(y) = kn(x, y). Thus the arc-cosine kernel function in eq. (1) can be viewed as the inner product between feature vectors derived from the mapping of an infinite single-layer threshold network =-=[8]-=-. Many researchers have noted the general connection between kernel machines and neural networks with one layer of hidden units [1]. The n = 0 arc-cosine kernel in eq. (1) can also be derived from an   our expression for the angular part of the kernel function in eq. (16), we recover our earlier claim that J0(θ) = π −θ. Related integrals for the special case n = 0 can also be found in earlier work =-=[8]-=-.For the case n&amp;gt;0, the integral in eq. (16) can be performed by the method of differentiating under the integral sign. In particular, we note that: ∫ π 2 ∫ π/2 cos dψ 0 n ψ 1 ∂ = (1 − cos θ cos ψ) n+1   </text>
<query_num> 15203 </query_num>
<text>   Mahalanobis distance metric for these outputs, though other methods are equally viable [17]. The use of LMNN is inspired by the supervised fine-tuning of weights in the training of deep architectures =-=[18]-=-. In MKMs, however, this supervised training only occurs at the final layer (which underscores the importance of feature selection in earlier layers). LMNN learns a distance metric by solving a proble   </text>
<query_num> 15204 </query_num>
<text>   al nets, over shallow architectures, such as support vector machines (SVMs) [1]. Deep architectures learn complex mappings by transforming their inputs through multiple layers of nonlinear processing =-=[2]-=-. Researchers have advanced several motivations for deep architectures: the wide range of functions that can be parameterized by composing weakly nonlinear transformations, the appeal of hierarchical  els have error rates from 22.36–25.64%. Results are s kernels of varying degree (n) and levels of recursion (ℓ). The best previous results are 24 SVMs with RBF kernels and 22.50% for deep belief nets =-=[2]-=-. See text for details. 26 24 22 Test error rate (%) 21 20 19 18 17 1 2 3 4 5 6 Step (n=0) Test error rate (%) 1 2 3 4 5 6 SVM−RBF Ramp (n=1) 1 2 3 4 5 6 Quarter!pipe (n=2) Figure 3: Left: examples fr  varying 1 2 degree 3 4 (n) 5 6and levels 1 2 of3 recursion 4 5 6 (ℓ). The best previous results are 19.13% f Step (n=0) with RBF kernels Ramp (n=1) and 18.63% Quarter−pipe for deep belief (n=2) nets =-=[2]-=-. See text for details. Figure 2: Left: examples from the rectangles-image data set. Right: classification error rates on the test set. SVMs with arc-cosine kernels have error rates from 22.36–25.64%. els have error rates from 22.36–25.64%. Results are s kernels of varying degree (n) and levels of recursion (ℓ). The best previous results are 24 SVMs with RBF kernels and 22.50% for deep belief nets =-=[2]-=-. See text for details. 21 20 19 18 17 Test error rate (%) 21 20 19 18 17 1 2 3 4 5 6 Step (n=0) Test error rate (%) SVM−RBF 1 2 3 4 5 6 Ramp (n=1) DBN−3 1 2 3 4 5 6 Quarter!pipe (n=2) Figure 3: Left:  varying 1 2 degree 3 4 (n) 5 6and levels 1 2 of3 recursion 4 5 6 (ℓ). The best previous results are 19.13% f Step (n=0) with RBF kernels Ramp (n=1) and 18.63% Quarter−pipe for deep belief (n=2) nets =-=[2]-=-. See text for details. Figure 3: Left: examples from the convex data set. Right: classification error rates on the test set. SVMs with arc-cosine kernels have error rates from 17.15–20.51%. Results a   </text>
<query_num> 15205 </query_num>
<text>   ation functions For n = 0, the activation function is a step function, and the network is an array of perceptrons. For n = 1, the activation function is a ramp function (or rectification nonlinearity =-=[9]-=-), and the mapping f(x) is piecewise linear. More generally, the nonlinear (non-polynomial) behavior of these networks is induced by thresholding on weighted sums. We refer to networks with these acti   </text>
<query_num> 15206 </query_num>
<text>   deep architectures yet drawn to the elegance of kernel methods. In this paper, we explore the possibility of deep learning in kernel machines. Though we share a similar motivation as previous authors =-=[7]-=-, our approach is very different. Our paper makes two main contributions. First, we develop a new family of kernel functions that mimic the computation in large neural nets. Second, using these kernel arc-cosine kernels may be yielding some of the advantages typically associated with deep architectures. S D 3 Deep learning In this section, we explore how to use kernel methods in deep architectures =-=[7]-=-. We show how to train deep kernel-based architectures by a simple combination of supervised and unsupervised methods. Using the arc-cosine kernels in the previous section, these multilayer kernel mac sic intuitions behind deep learning in the altogether different context of kernel-based architectures. A similar validation was provided by recent work on kernel methods for semi-supervised embedding =-=[7]-=-. We hope that our results inspire more work on kernel methods for deep learning. There are many possible directions for future work. For SVMs, we are currently experimenting with arc-cosine kernel fu   </text>
<query_num> 15207 </query_num>
<text>   discuss each of these steps in greater detail below. Kernel PCA. Deep learning in MKMs is achieved by iterative applications of kernel PCA [13]. This use of kernel PCA was suggested over a decade ago =-=[16]-=- and more recently inspired by the pretraining of deep belief nets by unsupervised methods. In MKMs, the outputs (or features) from kernel PCA at one layer are the inputs to kernel PCA at the next lay   </text>
<query_num> 15208 </query_num>
<text>   ltilayer kernel machines We explored how to train MKMs in stages that involve kernel PCA [13] and feature selection [14] at intermediate hidden layers and large-margin nearest neighbor classification =-=[15]-=- at the final output layer. Specifically, for ℓ-layer MKMs, we considered the following training procedure: 1. Prune uninformative features from the input space. 2. Repeat ℓ times: (a) Compute princip ce metric learning. Test examples in MKMs are classified by a variant of kNN classification on the outputs of the final layer. Specifically, we use large margin nearest neighbor (LMNN) classification =-=[15]-=- to learn a Mahalanobis distance metric for these outputs, though other methods are equally viable [17]. The use of LMNN is inspired by the supervised fine-tuning of weights in the training of deep ar nce metric by solving a problem in semidefinite programming; one advantage of LMNN is that the required optimization is convex. Test examples are classified by the energy-based decision rule for LMNN =-=[15]-=-, which was itself inspired by earlier work on multilayer neural nets [19]. 3.2 Experiments on multiway classification We evaluated MKMs on the two multiclass data sets from previous benchmarks [11] t y on the mnist-backrandom data set. Finally, LMNN classification in the output layer yielded consistent improvements over basic kNN classification provided that we used the energy-based decision rule =-=[15]-=-. 4 Discussion In this paper, we have developed a new family of kernel functions that mimic the computation in large, multilayer neural nets. On challenging data sets, we have obtained results that ou   </text>
<query_num> 15209 </query_num>
<text>   of LMNN is that the required optimization is convex. Test examples are classified by the energy-based decision rule for LMNN [15], which was itself inspired by earlier work on multilayer neural nets =-=[19]-=-. 3.2 Experiments on multiway classification We evaluated MKMs on the two multiclass data sets from previous benchmarks [11] that exhibited the largest performance gap between deep and shallow archite   </text>
<query_num> 15210 </query_num>
<text>   rchical distributed representations, and the potential for combining unsupervised and supervised methods. Experiments have also shown the benefits of deep learning in several interesting applications =-=[3, 4, 5]-=-. Many issues surround the ongoing debate over deep versus shallow architectures [1, 6]. Deep architectures are generally more difficult to train than shallow ones. They involve difficult nonlinear op   </text>
<query_num> 15211 </query_num>
<text>   rchical distributed representations, and the potential for combining unsupervised and supervised methods. Experiments have also shown the benefits of deep learning in several interesting applications =-=[3, 4, 5]-=-. Many issues surround the ongoing debate over deep versus shallow architectures [1, 6]. Deep architectures are generally more difficult to train than shallow ones. They involve difficult nonlinear op plore better schemes for feature selection [21, 22] and kernel selection [23]. Also, it would be desirable to incorporate prior knowledge, such as the invariances modeled by convolutional neural nets =-=[24, 4]-=-, though it is not obvious how to do so. These issues and others are left for future work. A Derivation of kernel function In this appendix, we show how to evaluate the multidimensional integral in eq   </text>
<query_num> 15212 </query_num>
<text>   rently experimenting with arc-cosine kernel functions of fractional and (even negative) degree n. For MKMs, we are hoping to explore better schemes for feature selection [21, 22] and kernel selection =-=[23]-=-. Also, it would be desirable to incorporate prior knowledge, such as the invariances modeled by convolutional neural nets [24, 4], though it is not obvious how to do so. These issues and others are l   </text>
<query_num> 15213 </query_num>
<text>   tively on multiclass data sets designed to foil shallow architectures [11]. 53.1 Multilayer kernel machines We explored how to train MKMs in stages that involve kernel PCA [13] and feature selection =-=[14]-=- at intermediate hidden layers and large-margin nearest neighbor classification [15] at the final output layer. Specifically, for ℓ-layer MKMs, we considered the following training procedure: 1. Prune l and marginal histograms of its discretized values; then, using these histograms, we estimate each feature’s mutual information with the class label and sort the features in order of these estimates =-=[14]-=-. In the second step, considering only the first w features in this ordering, we compute the error rates of a basic kNN classifier using Euclidean distances in feature space. We compute these error ra   </text>
<query_num> 15214 </query_num>
<text>   ure work. For SVMs, we are currently experimenting with arc-cosine kernel functions of fractional and (even negative) degree n. For MKMs, we are hoping to explore better schemes for feature selection =-=[21, 22]-=- and kernel selection [23]. Also, it would be desirable to incorporate prior knowledge, such as the invariances modeled by convolutional neural nets [24, 4], though it is not obvious how to do so. The   </text>
<query_num> 15215 </query_num>
<text>   uts of the final layer. Specifically, we use large margin nearest neighbor (LMNN) classification [15] to learn a Mahalanobis distance metric for these outputs, though other methods are equally viable =-=[17]-=-. The use of LMNN is inspired by the supervised fine-tuning of weights in the training of deep architectures [18]. In MKMs, however, this supervised training only occurs at the final layer (which unde   </text>
</top>
<top>
<paper_num> 153 </paper_num>
<paper_title>   Dex: A Semantic-Graph Differencing Tool for Studying Changes in Large Code Bases.  </paper_title>
<paper_abstract>   This paper describes an automated tool called Dex (Difference extractor) for analyzing syntactic and semantic changes in large C-language code bases. It is applied to patches obtained from a source code repository, each of which comprises the code changes made to accomplish a particular task. Dex produces summary statistics characterizing these changes for all of the patches that are analyzed. Dex applies a graph differencing algorithm to abstract semantic graphs (ASGs) representing each version. The differences are then analyzed to identify higher-level program changes. We describe the design of Dex, its potential applications, and the results of applying it to analyze bug fixes from the Apache and GCC projects. The results include detailed information about the nature and frequency of missing condition defects in these projects. 1.  </paper_abstract>
<query_num> 15301 </query_num>
<text>   based on whether other parts match. Chawathe and Garcia-Molina presented an algorithm to find changes in structured data modeled by unordered trees, which also includes move, copy and glue operations =-=[6]-=-. Their algorithm is a heuristic, iterative update mechanism like ours, but only accommodates tree edges. Note that the edit-distance problem between unordered trees is, in general, NPcomplete [29]. C   </text>
<query_num> 15302 </query_num>
<text>   de repositories is aided by the use of text differencing tools such as Gnu diff [11]. However, because such tools do not understand program syntax and semantics, they can provide only limited help [4]=-=[14]-=-[27]. This paper describes an automated tool called Dex (Difference extractor) for analyzing syntactic and semantic changes in large C-language code bases. Dex is meant to be useful both to software d rwitz proposed a technique for identifying semantic and textual differences between program versions that is based on partitioning program components into sets of components with equivalent behaviors =-=[14]-=-. The partitioning algorithm represents programs using program representation graphs, which combine aspects of program dependence graphs and static single assignment forms but do not contain the kind   </text>
<query_num> 15303 </query_num>
<text>   epositories is aided by the use of text differencing tools such as Gnu diff [11]. However, because such tools do not understand program syntax and semantics, they can provide only limited help [4][14]=-=[27]-=-. This paper describes an automated tool called Dex (Difference extractor) for analyzing syntactic and semantic changes in large C-language code bases. Dex is meant to be useful both to software devel lse positives and has high accuracy. Yang described an algorithm for finding syntactic differences for use with version control software between two programs that works by analyzing their parse trees =-=[27]-=-. Unlike our algorithm, Yang’s parse trees do not contain the semantic information present in ASTs and his algorithm places restrictions on admissible matchings that our algorithm does not: (1) two no   </text>
<query_num> 15304 </query_num>
<text>   iginal reason for creating Dex was to analyze bug fixes in large code bases in order to determine the kinds of execution profiling that should be used in conjunction with observation-based testing [9]=-=[18]-=-. 2 We are particularly interested in revealing missing condition defects, which involve omitted conditional code and are notoriously difficult to expose. Hence, many of the statistics Dex currently g   </text>
<query_num> 15305 </query_num>
<text>   matchings that our algorithm does not: (1) two nodes may only match if their parents match and (2) the order of sibling nodes must be preserved. Wang et al describe a binary matching tool called BMAT =-=[25]-=-. This tool works on program binaries, by matching basic blocks from the original binary to basic blocks in the new one. Matching is done based on similarity of basic blocks and limited control flow i   </text>
<query_num> 15306 </query_num>
<text>   of the merged program is given by Binkley et al in [5]. Mens surveys current results on software merging, including an overview of several differencing algorithms currently used for software merging =-=[19]-=-, most of which are presented in this section. Westfechtel presents an approach for merging of revisions where the editor is aware of the AST of the document and automatically assigns tags to new and   </text>
<query_num> 15307 </query_num>
<text>   port from the editing tool. Berlage and Genau propose representing differences by the sequence of commands executed by the user to make the change, which are automatically recorded by the application =-=[2]-=-. They also study how to merge two command sequences. A well-studied application of software differencing is software merging, where two sets of changes to a program have to be merged into a single fi   </text>
<query_num> 15308 </query_num>
<text>   re two sets of changes to a program have to be merged into a single final version. One of the first algorithms to provide guarantees on the behavior of the merged program is given by Binkley et al in =-=[5]-=-. Mens surveys current results on software merging, including an overview of several differencing algorithms currently used for software merging [19], most of which are presented in this section. West   </text>
<query_num> 15309 </query_num>
<text>   the possibility of determining effort spent on different changes by analyzing size and type of each change, together with reported total monthly effort for each developer from an accounting database =-=[12]-=-. Using this technique they looked for modules that were becoming harder to maintain, and also compared the amount of effort necessary for bug fixes versus new features. 9. Conclusion Dex provides an   </text>
<query_num> 15310 </query_num>
<text>   two sets of changes. Krinke presents a technique for detecting duplicate code in a program based on detecting pairs subgraphs with identical length-k paths in a fine-grained program dependence graph =-=[17]-=-,. A fine-grained PDG has nodes similar to those on an AST, and edges corresponding to those in the AST plus control and data dependence edges. Analyzing this graph allows the algorithm to find simila   </text>
</top>
<top>
<paper_num> 154 </paper_num>
<paper_title>   Convergence of Independent Adaptive Learners.  </paper_title>
<paper_abstract>   In this paper we analyze the convergence of independent adaptive learners in repeated games. We show that, in this class of games, independent adaptive learners converge to pure Nash equilibria, if they exist. We discuss the relation between our result and convergence results of adaptive play. The importance of our result stems from the fact that, unlike adaptive play, no communication/action observability is assumed. We also relate this result to recent results on the convergence of weakened ficticious play processes for independent learners [11, 19]. Finally we present some experimental results to illustrate the main ideas of the paper.  </paper_abstract>
<query_num> 15401 </query_num>
<text>   e authors discuss the use of optimistic and pessimistic assumptions on the other agents’ behavior and show their algorithm to converge in behavior to an optimal decision-rule. Kapetanakis and Kudenko =-=[7, 8]-=- an improvement is proposed that deals with more non-deterministic settings. Recent results have established the convergence of a variation of fictitious play for independent learners [11], first intr   </text>
<query_num> 15402 </query_num>
<text>   ents and are capable of perceiving (a posteriori) their actions and rewards. Learning algorithms considering JALs are easily implementable from standard single-agent reinforcement learning algorithms =-=[6, 12, 13]-=-. Action observability allows a learning agent to build statistics on the other agents’ behavior-rules and act in a best-response sense. This is the underlying principle of standard methods such as fi   </text>
<query_num> 15403 </query_num>
<text>   ents and are capable of perceiving (a posteriori) their actions and rewards. Learning algorithms considering JALs are easily implementable from standard single-agent reinforcement learning algorithms =-=[6, 12, 13]-=-. Action observability allows a learning agent to build statistics on the other agents’ behavior-rules and act in a best-response sense. This is the underlying principle of standard methods such as fi ents. Furthermore, no a priori knowledge of the payoff function is required. 1 For a discussion on the differences between convergence in beliefs and convergence in behavior, see the works by Littman =-=[12]-=-, Young [22].s6 Institute for Systems and Robotics 3.1 Independent adaptive learning process Let Γ = � N, (Ak), (rk) � be a repeated game played at discrete instants of time t = 1, 2, . . .. At each p   </text>
<query_num> 15404 </query_num>
<text>   n be naturally captured using a game theoretic model and their observed behavior suitably interpreted using game theoretic concepts. When addressing game theory from a learning perspective, Boutilier =-=[1]-=- distinguishes two fundamental classes of learning agents: independent learners (IL) and joint-action learners (JAL). The former have no knowledge on other agents, interacting with the environment as   </text>
<query_num> 15405 </query_num>
<text>   on is often far from trivial. With no knowledge on the other agents’ actions and 2sTechnical Report RT-603-07, May 2007 3 payoffs, the problem becomes more difficult. Tan [18] and Claus and Boutilier =-=[3]-=- some empirical evidence is gathered that describes the convergence properties of reinforcement learning methods in multi-agent settings. In these works, the experimental performance of ILs is compare   </text>
<query_num> 15406 </query_num>
<text>   ormation and action recognition is often far from trivial. With no knowledge on the other agents’ actions and 2sTechnical Report RT-603-07, May 2007 3 payoffs, the problem becomes more difficult. Tan =-=[18]-=- and Claus and Boutilier [3] some empirical evidence is gathered that describes the convergence properties of reinforcement learning methods in multi-agent settings. In these works, the experimental p   </text>
<query_num> 15407 </query_num>
<text>   performance of ILs is compared with that of JALs (using fictitious play). Wang and de Silva [21] and Crites and Barto [4] a similar comparison is performed for specific problems. Lauer and Riedmiller =-=[10]-=- study independent learners in deterministic settings. They provide a learning algorithm that relies on strict assumptions on the other agents’ behavior. In particular, the authors discuss the use of   </text>
<query_num> 15408 </query_num>
<text>   r players. Instead, we rely on the sampling process to implicitly provide this information. Before introducing our main result, we need the following definition, adapted from the work by Singh et al. =-=[17]-=-. Definition 3.1 (GLIE strategy). A strategy σi is greedy in the limit with infinite exploration (GLIE) if it verifies the following conditions: • each action is visited infinitely often; • in the lim well-known example of GLIE policy is Boltzmann exploration: P [At = a | r] = e r(a)/Tt � u∈A er(u)/Tt where Tt is a temperature parameter that decays at an adequate rate (see the work by Singh et al. =-=[17]-=- for further details). Theorem 3.1. Let Γ = � N, (Ak), (rk) � be a weakly acyclic N-player game. If K ≤ m L(Γ) + 2 , then every independent adaptive learner following a GLIE policy will converge to a   </text>
<query_num> 15409 </query_num>
<text>   scribe interactions of economical agents. Recent years have witnessed an increasing interest from the computer science and robotic communities in applying game theoretic models to multi-agent systems =-=[4, 5, 21]-=-. For example, the interaction of a group of robots moving in a common environment can be naturally captured using a game theoretic model and their observed behavior suitably interpreted using game th   </text>
<query_num> 15410 </query_num>
<text>   scribe interactions of economical agents. Recent years have witnessed an increasing interest from the computer science and robotic communities in applying game theoretic models to multi-agent systems =-=[4, 5, 21]-=-. For example, the interaction of a group of robots moving in a common environment can be naturally captured using a game theoretic model and their observed behavior suitably interpreted using game th cement learning methods in multi-agent settings. In these works, the experimental performance of ILs is compared with that of JALs (using fictitious play). Wang and de Silva [21] and Crites and Barto =-=[4]-=- a similar comparison is performed for specific problems. Lauer and Riedmiller [10] study independent learners in deterministic settings. They provide a learning algorithm that relies on strict assump   </text>
<query_num> 15411 </query_num>
<text>   unlike adaptive play, no communication/action observability is assumed. We also relate this result to recent results on the convergence of weakened ficticious play processes for independent learners =-=[11, 19]-=-. Finally we present some experimental results to illustrate the main ideas of the paper. 1 Introduction Game theory provides a mathematical framework to model situations in which several decisionmake d Kudenko [7, 8] an improvement is proposed that deals with more non-deterministic settings. Recent results have established the convergence of a variation of fictitious play for independent learners =-=[11]-=-, first introduced by Van der Genugten [19]. In this paper, we propose independent adaptive learning, a variation of adaptive play for independent learners. This algorithm has an obvious advantage ove   </text>
<query_num> 15412 </query_num>
<text>   ween the optimal and the suboptimal equilibria is much less significant. 4.0.5 3-Player game We now consider a fully cooperative 3-player game with multiple equilibria introduced by Wang and Sandholm =-=[20]-=-. In this game, 3 players have available 3 possible actions, α, β and γ. The players are rewarded maximum payoff if all 3 coordinate in the same individual action; they are rewarded a small payoff if  d (player 3) αα αβ αγ βα ββ βγ γα γβ γγ α 10 -20 -20 -20 -20 5 -20 5 -20 β -20 -20 5 -20 10 -20 5 -20 -20 γ -20 5 -20 5 -20 -20 -20 -20 10 Figure 14: Payoff for the 3-player game by Wang and Sandholm =-=[20]-=-. (α, α, α) (α, α, β) (α, α, γ) (α, β, α) (α, β, β) (α, β, γ) (β, α, α) (β, α, β) (β, α, γ) (β, β, α) (β, β, β) (β, β, γ) (γ, α, α) (γ, α, β) (γ, α, γ) (γ, β, α) (γ, β, β) (γ, β, γ) 40 20 0 (α, γ, α)   700 800 900 Time (steps) 40 20 0 1 Player 1 2 3 1 2 3 4 5 Player 2 &amp; 3 a) Evolution of expected payoff; b) Limit strategies. Figure 16: Learning performance in the 3-player game by Wang and Sandholm =-=[20]-=-. leads to a slight increase in the number of runs converging to the optimal equilibria and consequent decrease in the number of runs converging to the suboptimal equilibria (Figure 16.b)). This is al   </text>
</top>
<top>
<paper_num> 155 </paper_num>
<paper_title>   On Preemptive Resource Constrained Scheduling: Polynomial-Time Approximation Schemes.  </paper_title>
<paper_abstract>   We study resource constrained scheduling problems where the objective is to compute feasible  preemptive schedules minimizing the makespan and using no more resources than what are available.  </paper_abstract>
<query_num> 15501 </query_num>
<text>   )PTAS for the computation of the weighted independent set in a graph G 2 G and weights w, then we obtain a (F)PTAS for the fractional weighted coloring problem for graphs G 2 G. Using a recent result =-=[17]-=- for computing the maximum weighted independent set in intersection graphs of disks in the plane, we obtain the following: Corollary 6.2 There is a PTAS for the computation of the fractional weighted   </text>
<query_num> 15502 </query_num>
<text>   2 T , is processed by one machine requiring p j units of time and r ij units of resource i, i = 1; : : : ; s, from which only c i units are available at each time. One may assume w.l.o.g. that r ij 2 =-=[0; 1-=-] and c i  1. The objective is to compute a preemptive schedule of the tasks minimizing the maximum completion time C max . The three dots in the notation indicate that there are no restrictions on th s some simple inapproximability results. 4.1.1 Approximation Algorithms It is known [45, 49] that for general s, sD-KP or equivalently PIP has a 4 =s 1=c min ) approximation algorithm when all r ij 2 =-=[0; -=-1], c min = min i c i  1 and y j p j  0. This implies the following result: Theorem 4.1 For any number s of resources, there is a polynomial-time approximation algorithm with performance ratio O(s 1 c th our approach presented in the previous section and extending the algorithm to arbitrary OPT , we obtain the following. Theorem 4.2 For any number s of resources, if c i  12  2 log(2s) and r ij 2 [0=-=; 1-=-], for each i and j, there is a polynomial time approximation algorithm that computes a (1 + )-approximate solution for the preemptive resource constrained scheduling problem. 10 The last two results   </text>
<query_num> 15503 </query_num>
<text>   asks with running time polynomial in m and n. These results are based on methods by Grotschel et al. [26] and use the ellipsoid method. Another related problem is fractional graph coloring, see e.g. [=-=18, 25, 39, 41, 43, 47, 48-=-]. Grotschel et al. [25] proved that the weighted fractional coloring problem is NP-hard for general graphs, but can be solved in polynomial time for perfect graphs. They have proved the following int by Lund and Yannakakis [39] who proved that there exists as&amp;gt; 0 such that there is no polynomial-time approximation algorithm for the problem with approximation ratio n , unless P=NP. Feige and Kilian =-=[-=-18] showed that the fractional chromatic number  f (G) cannot be approximated within jV j 1  ) for any  &amp;gt; 0, unless ZPP=NP. Recently, the authors [29] proved that fractional coloring is NP-hard even f gether at the same time, therefore the (fractional) coloring problem for graphs can be viewed as a special case of (preemptive) resource constrained scheduling. Hence the inapproximability results in =-=[39, 18]-=- imply the following: Theorem 4.3 For anys&amp;gt; 0, the preemptive resource constrained scheduling problem with n tasks and s resources has no polynomial-time approximation algorithm with approximation rat   </text>
<query_num> 15504 </query_num>
<text>   asks with running time polynomial in m and n. These results are based on methods by Grotschel et al. [26] and use the ellipsoid method. Another related problem is fractional graph coloring, see e.g. [=-=18, 25, 39, 41, 43, 47, 48-=-]. Grotschel et al. [25] proved that the weighted fractional coloring problem is NP-hard for general graphs, but can be solved in polynomial time for perfect graphs. They have proved the following int t degree 4. Similarly, as it was shown by Gerke and McDiarmid [22], the problem remains NP-hard even for triangle-free graphs. Regarding the approximability of the fractional chromatic number, Matsui =-=[41]-=- gave a polynomial-time 2-approximation algorithm for unit disk graphs. 1.2 New Results The results presented in this paper are based on linear programming formulations. They are typically of the foll s in G. By applying this general result for intersection graphs of disks in the plane, we also obtain a PTAS for the fractional coloring problem providing a substantial improvement on Matsui&amp;apos;s result =-=[41]-=-. 2 Approximate Max-Min Resource Sharing In this section we will follow the presentation of [24] and use the notation introduced there. Let f : B ! IR M + be a vector with M non-negative, continuous,  graph class contains planar graphs and unit disk graphs, Corollary 6.1 implies the following result which also provides a substantial improvement on Matsui&amp;apos;s polynomial-time 2-approximation algorithm =-=[41]-=- for unit disk graphs: Corollary 6.3 There is a PTAS for the computation of the fractional weighted chromatic number for planar and unit disk graphs. 7 Conclusion In this paper we have studied preempt   </text>
<query_num> 15505 </query_num>
<text>   ax x2P p T f(x). Based on the paper of Grigoriadis et al. [24], we derive the following result extending some of the previous works on computing approximate solutions for fractional covering problems =-=[24, 44, 5-=-4]: If there exists a polynomial-time approximation algorithm with approximation ratio c for the subproblem, i.e. forsnding a vector ^ x(p) 2 P such that p T f(^x)  1 c (p)), then there is also a poly nction value 1  c   . Interestingly, the number of iterations (hence also the number of calls to the solver for the subproblem) is bounded by O(M(lnM + ln c 3 +  2 )), independently of the width [44] of P and the number of variables. If - in particular - there is a (fully) polynomial time approximation scheme ((F)PTAS) for the subproblem, one also gets a (F)PTAS for the original problem and the n s bounded by N 0 +O( M ln c  dln( 1 t )e X `=1 2 2` + M  dln( 1 t )e X `=1 2 ` )  O( M ln(cM)  + M ln c  3 + M  2 ): The total number of iterations can be improved by the scaling method used in [44, 24]. The idea is to reduce the parameter t step by step to the desired accuracy. In the s-th scaling phase we set  s =  s 1 =2 and t s =  s =5 and use the current approximate point x s 1 as its initia plex, the optimum of this linear program is also attained at a vertex ~ x of P corresponding to a (single) conguration ~ f . A similar argument was used for the bin packing problem by Plotkin et al. [=-=4-=-4]. At this vertex ~ x ~ f = r and ~ x f = 0 for f 6= ~ f . Therefore, it suces tosnd a subset ~ f of tasks that can be executed in parallel and has the largest associated prot value c ~ f in the prot ly relatively weak polynomial-time approximation results (i.e. with constant, logarithmic, or with even worse approximation ratios). To handle these cases too, we have extended some of the methods in =-=[24, 44, 54]-=- to the case where the subproblem can be solved only approximatively. The underlying algorithm is independent from the width [44] and the number of variables. We note that by using other techniques [3   </text>
<query_num> 15506 </query_num>
<text>   ctively. De la Vega and Lueker [12] gave a linear-time algorithm with asymptotic approximation ratio s +  for eachsxed  &amp;gt; 0. Further results and improvements for non-preemptive variant are given in [1=-=0, 50, 51]-=-. For the preemptive variant substantially less results are known: Blazewicz et al. [6] proved that when m issxed, the problem Pmjpmtn; res:::jC max (with identical machines) and even the variant Rmjp   </text>
<query_num> 15507 </query_num>
<text>   ctively. De la Vega and Lueker [12] gave a linear-time algorithm with asymptotic approximation ratio s +  for eachsxed  &amp;gt; 0. Further results and improvements for non-preemptive variant are given in [1=-=0, 50, 51]-=-. For the preemptive variant substantially less results are known: Blazewicz et al. [6] proved that when m issxed, the problem Pmjpmtn; res:::jC max (with identical machines) and even the variant Rmjp ance ratio O(s 1 c min ) for the preemptive resource constrained scheduling problem. This result can be further improved by using the algorithm by Srinivasan [49]. Furthermore, Srivastav and Stangier =-=[50-=-, 51] showed that if c min  16  2 log(2s) and OPT  12= 2 (where OPT is the optimum value of the linear relaxation of (9)), an -approximate solution for sD-KP can be computed in polynomial-time. The ru   </text>
<query_num> 15508 </query_num>
<text>   ection 2 we describe the methodology used for solving the max-min resource sharing problem. Let f(x) = (f 1 (x); : : : ; f M (x)) and (p) = max x2P p T f(x). Based on the paper of Grigoriadis et al. [=-=24]-=-, we derive the following result extending some of the previous works on computing approximate solutions for fractional covering problems [24, 44, 54]: If there exists a polynomial-time approximation  cular - there is a (fully) polynomial time approximation scheme ((F)PTAS) for the subproblem, one also gets a (F)PTAS for the original problem and the number of iterations is at most O(M(lnM +  2 )) [=-=24]-=-. This fact can be particularly useful for models with exponentially many variables. In Section 3 we describe a linear programming approach for the preemptive resource constrained scheduling problem,  btain a PTAS for the fractional coloring problem providing a substantial improvement on Matsui&amp;apos;s result [41]. 2 Approximate Max-Min Resource Sharing In this section we will follow the presentation of =-=[24]-=- and use the notation introduced there. Let f : B ! IR M + be a vector with M non-negative, continuous, concave functions f m , block B a non-empty, convex, compact set and e T = (1; : : : ; 1) 2 IR M  1  . In particular, if there is a (F)PTAS for the block problem computing an ^ x 2 B such that p T f(^x)  (1 )(p) for any constant  &amp;gt; 0, then there is a (F)PTAS for the resource sharing problem [24]. The algorithm uses the logarithmic potential function  t (; f) = ln  + t M M X m=1 ln(f m ); 4 where  2 IR, f = (f 1 ; : : : ; f M ) are variables associated with the coupling constraints f m  strictly increasing function of . The logarithmic dual vector p = p(f) for asxed f is dened by p m (f) = t M (f) f m (f) : (4) By (3), we have p(f) 2 P . We will also use the following properties [24]: Proposition 2.1 (a) p(f) T f = (1 + t)(f): (b) (f) 1+t  (f)  (f) 1+t=M : Now dene parameter v = v(x; ^ x) by v(x; ^ x) = p T ^ f p T f p T ^ f + p T f ; (5) where p 2 P , f = f(x), ^ f = f(^x   </text>
<query_num> 15509 </query_num>
<text>   el (non-malleable) model P jsize j jC max , there is a value size j 2 M = f1; : : : ; mg given for each task T j indicating that T j can be processed on any subset of processors of cardinality size j =-=[4, 14, 15, 28, 33, 52]-=-. In the malleable variant P jfctnjC max , each task can be executed on an arbitrary subset of processors, and the execution time p j (`) depends on the number ` of processors assigned to it [38, 42,   </text>
<query_num> 15510 </query_num>
<text>   i.e. it is part of the input. First we give the presentation of our approximation algorithms, then we brie  </text>
<query_num> 15511 </query_num>
<text>   mate solution for the sD-KP. Using that s is constant, the running time of our scheduling algorithm is bounded by O((K(n; s; c) +n ln ln(n 1 ))n ln( 1 )( 2 + ln n)). The currently known best bound [7] for K(n; s; ()) is O(n b s  cs ) = n O( s  ) . By using this bound and the above argument, we obtain the following result. Theorem 4.4 For anysxed number s of resources, there is a PTAS for the p task j and capacity c 2 = m. It is easy to check, that the subproblem in this case is the cardinality constrained ( P n j=1 x j  m) knapsack problem, which has a FPTAS with running time O(nm 2  1 ) [7]. In addition, the initial interval for the binary search on r can be bounded as for s = 2 resources. Hence the following holds: Theorem 4.7 There is a FPTAS of running time O(n 2 ln( 1 ) max(m 2  1   </text>
<query_num> 15512 </query_num>
<text>   oximation ratio n , unless P=NP. Feige and Kilian [18] showed that the fractional chromatic number  f (G) cannot be approximated within jV j 1  ) for any  &amp;gt; 0, unless ZPP=NP. Recently, the authors [29=-=-=-=-] proved that fractional coloring is NP-hard even for graphs with  f (G) = 3 and constant degree 4. Similarly, as it was shown by Gerke and McDiarmid [22], the problem remains NP-hard even for triangl   </text>
<query_num> 15513 </query_num>
<text>   ssical (1-dimensional) Knapsack Problem (instead of the s-dimensional variant). This can be solved approximately with any () accuracy in O(nmin(ln n; ln(1=)) + 1= 2 min(n; 1= ln(1=))) = O(n 2 ) time [=-=32-=-]. In addition, we have to count the overhead of O(n ln ln(n 1 )) operations in each iteration (i.e. the computation of the root and the new price vector). Hence the previous bound can be substituted   </text>
</top>
<top>
<paper_num> 156 </paper_num>
<paper_title>   Alibi Framework for Identifying Reactive Jamming Nodes in Wireless LAN.  </paper_title>
<paper_abstract>   Reactive jamming nodes are the nodes of the network that get compromised and become the source of jamming attacks. They assume to know any shared secrets and protocols used in the networks. Thus, they can jam very effectively and are very stealthy. We propose a novel approach to identifying the reactive jamming nodes in wireless LAN (WLAN). We rely on the half-duplex nature of nodes: they cannot transmit and receive at the same time. Thus, if a compromised node jams a packet, it cannot guess the content of the jammed packet. More importantly, if an honest node receives a jammed packet, it can prove that it cannot be the one jamming the packet by showing the content of the packet. Such proofs of jammed packets are called “alibis ”-the key concept of our approach. In this paper, we present an alibi framework to deal with reactive jamming nodes in WLAN. We propose a concept of alibi-safe topologies on which our proposed identification algorithms are proved to correctly identify the attackers. We further propose a realistic protocol to implement the identification algorithm. The protocol includes a BBC-based timing channel for information exchange under the jamming situation and a similarity hashing technique to reduce the storage and network overhead. The framework is evaluated in a realistic TOSSIM simulation where the simulation characteristics and parameters are based on real traces on our small-scale MICAz test-bed. The results show that in reasonable dense networks, the alibi framework can accurately identify both non-colluding and colluding reactive jamming nodes. Therefore, the alibi approach is a very promising approach to deal with reactive jamming nodes. This material is based upon work supported by the  </paper_abstract>
<query_num> 15601 </query_num>
<text>   and can easily be detected by the BS. 4.5 Similarity Hashing (SimHash) &amp; Alibi Locality Sensitive Hashing (LSH) is a popular technique used in information retrieval to detect near-duplicate documents =-=[6]-=-. Essentially, LSH is a method of performing probabilistic dimension reduction of high-dimensional data. The basic idea is to hash high-dimension input objects so that similar objects are mapped into  to hash the packet content. That means, the reception similarity of two packet content B1, B2 will be calculated as sim(B1, B2) = 1 − h(B1),h(B2) l . While there are several techniques to implement F =-=[6]-=-[10][3][8], we choose the random projection technique, also referred to as simhash, proposed by Charikar [6] due to its simple and efficient implementation [18]. The random projection method of the si  t times uniformly, we will have a t-bit output of vector ⃗v. For any two vectors ⃗u, ⃗v, the bits of two t-bit outputs match with probability proportional to the cosine of the angle between them. In =-=[6]-=-, the author proves that Pr[hr(⃗u) = hr(⃗v)] = 1 − θ(⃗u,⃗v) , where θ is the angle of π vectors ⃗u and ⃗v. The point of using simhash is to reduce the amount of data needed for storing and transmittin   </text>
<query_num> 15602 </query_num>
<text>   ckers leave no identity information in the jammed packets (e.g. by corrupting the sender field), detection systems relying on identity clues to infer nodes causing the jammed packet do not work (e.g. =-=[24]-=-[12]) We propose a novel framework to this challenging problem. The framework relies on “alibi” concept and thus is named as alibi framework. Alibi is“a form of defense whereby a defendant attempts to h jamming attacks to the network and then deliver the new shared secrets to un-compromised nodes only. Researchers have been looking into the problem of identifying mis-behaving/compromised nodes. In =-=[24]-=-[12], the authors propose the detection schemes to identify misbehaving nodes that greedily consume the bandwidth by modifying its MAC parameters. However, these detection schemes will fail to detect  approaches. It needs a jamming-resistant communication like BBC[9]. In terms of detection, it collects the proofs showing good behaviors of nodes instead of collecting proofs of bad behaviors of nodes=-=[24]-=-[12]. 7. CONCLUSIONS We have presented a design and implementation of the alibi framework to deal with reactive jamming nodes. The framework relies on the novel concept of “alibi” in which the detecto   </text>
<query_num> 15603 </query_num>
<text>   e packet content. That means, the reception similarity of two packet content B1, B2 will be calculated as sim(B1, B2) = 1 − h(B1),h(B2) l . While there are several techniques to implement F [6][10][3]=-=[8]-=-, we choose the random projection technique, also referred to as simhash, proposed by Charikar [6] due to its simple and efficient implementation [18]. The random projection method of the similarity p   </text>
<query_num> 15604 </query_num>
<text>   ed as proactive or reactive. In the proactive jamming strategy, the attacker jams the channel without caring about the on-going communication. A typical example of this type is the continuous jamming =-=[30]-=-[28]. This strategy is the 9Detection Probability 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 ABS OMS 2 3 4 5 6 7 8 9 #Jammers #False alarms 10 5 ABS OMS 0 2 3 4 5 6 7 8 9 #Jammers Detection Probability   </text>
<query_num> 15605 </query_num>
<text>   fective 802.11 jammer [28]. Therefore, there is a strong motivation to avoid using one single shared secret. One approach is to use multiple shared secrets rather just one single shared secret. In [5]=-=[7]-=-[14], the common idea is to divide nodes into multiple groups and assign one unique shared secret to each group. The assigned shared secret of a group is used to derive the hopping-pattern for that gr   </text>
<query_num> 15606 </query_num>
<text>   k can deal with this type of attackers in the context of WLAN (see Section 6) briefly due to two reasons. First, many approaches are only concerned about how to build jamming-resistant communications =-=[27]-=-[26][22][9][11] without identifying the source of jamming. Jamming-resistant communications are necessary but not sufficient because as long as the jamming nodes are not identified, they always have e d secret of traditional spread spectrum technologies. Recently, researchers have been proposing spread spectrum communication without any pre-shared secrets. Uncoordinated Frequency Hopping (UFH) [26]=-=[27]-=-[11] allows two nodes that do not any common secret to establish a secret key for future FHSS communication. Uncoordinated Direct Sequence Spread Spectrum (UDSSS) [22] avoids jamming by randomly selec   </text>
<query_num> 15607 </query_num>
<text>   lest way to perform a jamming attack. However, it is not energy-efficient due to the continuous jamming activity which also makes the attacker easy to detect. Reactive jamming strategy [2][19][13][17]=-=[15]-=-[20][4][12][28][23][21], in contrast, avoids these drawbacks by intelligently listening and jamming the channel. Thus, reactive jamming attacks are more difficult to detect and are more energy-efficie   </text>
<query_num> 15608 </query_num>
<text>   n deal with this type of attackers in the context of WLAN (see Section 6) briefly due to two reasons. First, many approaches are only concerned about how to build jamming-resistant communications [27]=-=[26]-=-[22][9][11] without identifying the source of jamming. Jamming-resistant communications are necessary but not sufficient because as long as the jamming nodes are not identified, they always have effec hared secret of traditional spread spectrum technologies. Recently, researchers have been proposing spread spectrum communication without any pre-shared secrets. Uncoordinated Frequency Hopping (UFH) =-=[26]-=-[27][11] allows two nodes that do not any common secret to establish a secret key for future FHSS communication. Uncoordinated Direct Sequence Spread Spectrum (UDSSS) [22] avoids jamming by randomly s   </text>
<query_num> 15609 </query_num>
<text>   rect Sequence Spread Spectrum (UDSSS) [22] avoids jamming by randomly selecting a spread code sequence from a pool of code sequences. However, UDSSS is vulnerable to reactive jamming attacks. RD-DSSS =-=[16]-=- also proposes a similar technique that can be resistant to reactive jamming attacks. BBC [9] proposes a coding approach to encode data to be transmitted into“indelible”marks that can be decoded witho   </text>
<query_num> 15610 </query_num>
<text>   rotocol simplest way to perform a jamming attack. However, it is not energy-efficient due to the continuous jamming activity which also makes the attacker easy to detect. Reactive jamming strategy [2]=-=[19]-=-[13][17][15][20][4][12][28][23][21], in contrast, avoids these drawbacks by intelligently listening and jamming the channel. Thus, reactive jamming attacks are more difficult to detect and are more en   </text>
<query_num> 15611 </query_num>
<text>   s leave no identity information in the jammed packets (e.g. by corrupting the sender field), detection systems relying on identity clues to infer nodes causing the jammed packet do not work (e.g. [24]=-=[12]-=-) We propose a novel framework to this challenging problem. The framework relies on “alibi” concept and thus is named as alibi framework. Alibi is“a form of defense whereby a defendant attempts to pro  perform a jamming attack. However, it is not energy-efficient due to the continuous jamming activity which also makes the attacker easy to detect. Reactive jamming strategy [2][19][13][17][15][20][4]=-=[12]-=-[28][23][21], in contrast, avoids these drawbacks by intelligently listening and jamming the channel. Thus, reactive jamming attacks are more difficult to detect and are more energy-efficient. Due to  mming attacks to the network and then deliver the new shared secrets to un-compromised nodes only. Researchers have been looking into the problem of identifying mis-behaving/compromised nodes. In [24]=-=[12]-=-, the authors propose the detection schemes to identify misbehaving nodes that greedily consume the bandwidth by modifying its MAC parameters. However, these detection schemes will fail to detect stea oaches. It needs a jamming-resistant communication like BBC[9]. In terms of detection, it collects the proofs showing good behaviors of nodes instead of collecting proofs of bad behaviors of nodes[24]=-=[12]-=-. 7. CONCLUSIONS We have presented a design and implementation of the alibi framework to deal with reactive jamming nodes. The framework relies on the novel concept of “alibi” in which the detector co   </text>
<query_num> 15612 </query_num>
<text>   this type of attackers in the context of WLAN (see Section 6) briefly due to two reasons. First, many approaches are only concerned about how to build jamming-resistant communications [27][26][22][9]=-=[11]-=- without identifying the source of jamming. Jamming-resistant communications are necessary but not sufficient because as long as the jamming nodes are not identified, they always have effective jammin cret of traditional spread spectrum technologies. Recently, researchers have been proposing spread spectrum communication without any pre-shared secrets. Uncoordinated Frequency Hopping (UFH) [26][27]=-=[11]-=- allows two nodes that do not any common secret to establish a secret key for future FHSS communication. Uncoordinated Direct Sequence Spread Spectrum (UDSSS) [22] avoids jamming by randomly selecting   </text>
<query_num> 15613 </query_num>
<text>   to perform a jamming attack. However, it is not energy-efficient due to the continuous jamming activity which also makes the attacker easy to detect. Reactive jamming strategy [2][19][13][17][15][20]=-=[4]-=-[12][28][23][21], in contrast, avoids these drawbacks by intelligently listening and jamming the channel. Thus, reactive jamming attacks are more difficult to detect and are more energy-efficient. Due   </text>
<query_num> 15614 </query_num>
<text>   way to perform a jamming attack. However, it is not energy-efficient due to the continuous jamming activity which also makes the attacker easy to detect. Reactive jamming strategy [2][19][13][17][15]=-=[20]-=-[4][12][28][23][21], in contrast, avoids these drawbacks by intelligently listening and jamming the channel. Thus, reactive jamming attacks are more difficult to detect and are more energy-efficient.   </text>
<query_num> 15615 </query_num>
<text>   which significantly reduces the signal to noise and interference ratio (SINR) and reducing probability of successful message receptions. There are two types of jamming strategies in wireless networks =-=[31]-=-: proactive and reactive jamming strategies. In proactive jamming strategies, attackers jam channels regardless of whether there are on-going communication activities on the channels. Typical examples ing proofs incurs overhead to nodes, the base station only collects proofs when there is a jamming attack. To detect the presence of jamming attacks, we use a similar detection techniques proposed in =-=[31]-=-. For the uplink traffic (i.e. from nodes to the base station), a jamming attack is declared if the base station receives a significant number of corrupted packets with strong received signal strength   </text>
</top>
<top>
<paper_num> 157 </paper_num>
<paper_title>   Trade-Offs in Protecting Storage: A Meta-Data Comparison of Cryptographic, Backup/Versioning, Immutable/Tamper-Proof, and Redundant Storage Solutions.  </paper_title>
<paper_abstract>   Modern storage systems are responsible for increasing amounts of data and the value of the data itself is growing in importance. Several primary storage system solutions have emerged for the protection of data: (1) Secure Storage through Cryptography, (2) Backup and Versioning Systems, (3) Immutable and Tamper-Proof Storage, and (4) Redundant Storage. Using results from published studies, we compare these four solutions against different requirements highlighting trade-offs in performance, space, attack resistance, and cost. We also present a case study of applying these solutions based on design work at NCSA. Lastly, we conclude that while different storage protection solutions may be appropriate for different requirements, some general conclusions can be made about current state-ofthe -art storage protection solutions as well as directions for future research.  </paper_abstract>
<query_num> 15701 </query_num>
<text>   ) Redundancy improves availability and integrity in the space dimension - data is replicated in different space so it can be recovered instantly. The best example of storage system redundancy is RAID =-=[4]-=-. Either through mirroring (RAID 1), parity (RAID 5), or error-correcting codes (RAID 3), RAID can recover from hardware failures. Erasure codes [12] are another important storage system protection te  performance improvements, if the additional data can be used to provide benefits to read performance. Penalties are generally paid in write performance, but these penalties again are well understood =-=[4]-=-. 3.2. Space Cryptographic and tamper-proof systems incur similar sorts of overhead in the form of additional metadata. In cryptographic systems, this is key management data. Tamper-proofing requires   </text>
<query_num> 15702 </query_num>
<text>   a they require may be missing. Additionally, the extra copies of commonly read 106 data can improve performance. This automatic balancing between replication levels is the key idea of the HP AutoRAID =-=[29]-=- system. AutoRAID puts newly written data in a RAID-1 style redundancy pattern. As space is filled, older, unaccessed data is moved to a more efficient RAID5 style redundancy. This allows better stora nal parity and other “RAID6” schemes use (n + 2) disks. Erasure codes allow an arbitrarily flexible choice of space overhead, although with added complexity. Finally, both D-GRAD [26] and HP AutoRAID =-=[29]-=- impose variable amounts of overhead, adjust107 ing to disk usage patterns to provide better protection when the space is available. When disk space is not available, they approach RAID-5 efficiency.   </text>
<query_num> 15703 </query_num>
<text>   aphy is implemented, the magnitude of the costs can vary greatly. One important factor is the choice of cipher. 3DES can result in throughput as low as 10MB/sec, while Blowfish can approach 53 MB/sec =-=[32]-=-. Performance comparisons of different ciphers can be found in [8, 27, 18]. The new AES cipher, Rijndael, can be very high performing if properly implemented. On the other hand, public key cryptograph   </text>
<query_num> 15704 </query_num>
<text>   e 2. Relationship between tamperresistant and immutable techniques Cryptographic systems may provide tamper-proofing in addition to confidentiality [34, 3, 15, 11]. Contentaddressable storage systems =-=[9, 10]-=- inherently include tamper-proofing. Two general techniques exist: cryptographic hashes and cryptographic signatures. In a hashing system, a summary of the data is made with a hash function. If a copy   </text>
<query_num> 15705 </query_num>
<text>   e best example of storage system redundancy is RAID [4]. Either through mirroring (RAID 1), parity (RAID 5), or error-correcting codes (RAID 3), RAID can recover from hardware failures. Erasure codes =-=[12]-=- are another important storage system protection technique based on redundancy. Erasure codes allow a more flexible allotment between failure resistance and space usage than ordinary parity or mirrori   </text>
<query_num> 15706 </query_num>
<text>   ectiveness (see Figure 4) of backup management software, and necessitates different strategies. Figure 4. Effectiveness of backup as a function of mobility. Notable versioning file systems include S4 =-=[28]-=-, the Repairable File Service (RFS) [35], and the Elephant file system [23]. S4 and RFS are both comprehensive versioning systems—all writes are logged up to a specified age. S4 [28] operates under th ive versioning impose similar space overheads, in that space requirements continually grow. Depending on the application, the rate of growth can vary a lot, although for many systems it is reasonable =-=[28]-=-. Workstation disks typically experience no more than 200MB a day of write traffic, and so can easily be comprehensively versioned, perhaps to an immutable system. Redundancy achieves its gains primar . Since versioning is often quite space efficient, a large window of changes can be kept. An attacker may make many changes to a system to perform a denial of service against the versioning system–S4 =-=[28]-=- addresses those attacks by rate limiting changes if malicious behavior is detected. Together, these qualities allow versioning to greatly increase availability. It is important to realize that all of   </text>
<query_num> 15707 </query_num>
<text>   ed such that any two disk failure can be recovered. Write performance is somewhat worse than RAID-5, but read performance is equivalent. Some data is of course more important than other data. D-GRAID =-=[26]-=- recognizes this. D-GRAID uses the extra free space generally available in a disk array to store additional copies of more critical data, such as file-system metadata. Under failure, D-GRAID allows th able space. Row-diagonal parity and other “RAID6” schemes use (n + 2) disks. Erasure codes allow an arbitrarily flexible choice of space overhead, although with added complexity. Finally, both D-GRAD =-=[26]-=- and HP AutoRAID [29] impose variable amounts of overhead, adjust107 ing to disk usage patterns to provide better protection when the space is available. When disk space is not available, they approac   </text>
<query_num> 15708 </query_num>
<text>   gement software, and necessitates different strategies. Figure 4. Effectiveness of backup as a function of mobility. Notable versioning file systems include S4 [28], the Repairable File Service (RFS) =-=[35]-=-, and the Elephant file system [23]. S4 and RFS are both comprehensive versioning systems—all writes are logged up to a specified age. S4 [28] operates under the assumption that the host system is unt d through the host system, using strong cryptography to defeat a potentially compromised host. Studies of daily write traffic show that total versioning is reasonable given usual disk capacities. RFS =-=[35]-=- extends the S4 concept to include logging to isolate damage caused by a malicious modification. Given the identification of a malicious process, RFS uses a dependency tree of operationssTable 2. Comp   </text>
<query_num> 15709 </query_num>
<text>   ifferent strategies. Figure 4. Effectiveness of backup as a function of mobility. Notable versioning file systems include S4 [28], the Repairable File Service (RFS) [35], and the Elephant file system =-=[23]-=-. S4 and RFS are both comprehensive versioning systems—all writes are logged up to a specified age. S4 [28] operates under the assumption that the host system is untrustworthy. Every write up to the a  Very Low High High High Security High High Medium/Low Medium Very Low Low Management Overhead High Medium High/Medium Low Low Low to automatically identify a recovery state. The Elephant file system =-=[23]-=- keeps “landmark” versions, representative of the file system state at increasing intervals in time. Individual files may be assigned different versioning levels: no versioning, complete versioning, w   </text>
<query_num> 15710 </query_num>
<text>   ling cost of disk drives, improves performance, especially for restores. However, tape is still the cheapest media for bulk backup, especially given it’s low power costs. Massive Arrays of Idle Disks =-=[6]-=-, or MAID, examines the potential of idling many of these disks to save power. Power usage can be reduced by 90% with only marginal reductions in performance compared to an active array. Recent work h ces, floorspace, power, etc. For example, to store 1 PB in a tape library for a year, one may consume $9,400 worth of power. Under the same conditions, a disk array may consume $91,500 worth of power =-=[6]-=-. The specific additional resources imposed can vary greatly from technique to technique and are also de108 pendent on the specific implementation details of a particular environment. Although it may  atencies can exceed an hour. Although it is uncertain to us at this time the specific causes of these peak latencies, in the long term, a move away from tape and towards spinning disk may be feasible =-=[6]-=-. 4.5. Remarks The storage environment at NCSA is divided into two sub-environments, the high-performance systems and the mass storage system. Both are large and heavily used, but have unique characte   </text>
<query_num> 15711 </query_num>
<text>   lties on write performance. Double failures are not uncommon due to common environmental conditions or malicious attack. To survive double failures, a redundancy technique such as Row-Diagonal Parity =-=[7]-=- may be used. Data loss from double failures generally comes from the failure of individual blocks on other disks during the rebuild of a failed disk. These failed blocks which could ordinarily be scr   </text>
<query_num> 15712 </query_num>
<text>   m which they employ. To demonstrate the significance of cryptographic operations, Table 1 offers a performance comparison of select systems. The Transparent Cryptographic File System [3] and NCryptFS =-=[33]-=- are two examples of file system level systems that perform all cryptographic operations within the operating system prior to writing data to disk. Both are designed to be layered on top of another fi onship between tamperresistant, immutable, and cryptographic systems. Tamperresistant systems use cryptography to provide integrity [14], while some cryptographic systems provide only confidentiality =-=[33]-=-. Some tamper-resistant techniques provide software immutability, but others such as TCFS do not. Hardware and media-enforced immutability often do not provide the metadata protection necessary for ta   </text>
<query_num> 15713 </query_num>
<text>   ne important factor is the choice of cipher. 3DES can result in throughput as low as 10MB/sec, while Blowfish can approach 53 MB/sec [32]. Performance comparisons of different ciphers can be found in =-=[8, 27, 18]-=-. The new AES cipher, Rijndael, can be very high performing if properly implemented. On the other hand, public key cryptographic operations, critical in many digital signature algorithms, are among th   </text>
<query_num> 15714 </query_num>
<text>   new keys distributed, while NCryptFS relies on a cumbersome timeout threshold, forcing users to periodically re-authenticate themselves during operations. Similar to the file system solutions, PLUTUS =-=[15]-=- and SiRiUS [11] implement a cryptographic storage system over a standard remote file system. The work of encryption, decryption, and signing is done on the hosts, allowing a scalability that is not p  write perforTable 1. Comparison of performance of different systems with different cryptographic methods. Note that the performance results were derived from unrelated studies of individual systems =-=[3, 15]-=- and cannot be used for a direct comparison. Storage System Sequential Read Sequential Write OpenAFS 1.28 s 1.57 s PLUTUS w/o crypto 1.39 s 1.59 s PLUTUS DES 4.58 s 4.27 s PLUTUS 3-DES 7.84 s 7.92 s N a protection necessary for tamperproofing. 103 Figure 2. Relationship between tamperresistant and immutable techniques Cryptographic systems may provide tamper-proofing in addition to confidentiality =-=[34, 3, 15, 11]-=-. Contentaddressable storage systems [9, 10] inherently include tamper-proofing. Two general techniques exist: cryptographic hashes and cryptographic signatures. In a hashing system, a summary of the   </text>
<query_num> 15715 </query_num>
<text>   pendent server failures. 2.2. Tamper-proofing and immutability Tamper-proofing and immutability are two techniques to improve the integrity of data. Tamper-proofing provides only integrity guarantees =-=[14]-=-. The purpose of tamperproofing is to detect if modifications have occurred. Generally cryptographic hashing or signing is used to effect tamper-proofing. Immutability, on the other hand, provides bot amper-proofing fill different roles. Figure 2 illustrates the relationship between tamperresistant, immutable, and cryptographic systems. Tamperresistant systems use cryptography to provide integrity =-=[14]-=-, while some cryptographic systems provide only confidentiality [33]. Some tamper-resistant techniques provide software immutability, but others such as TCFS do not. Hardware and media-enforced immuta   </text>
<query_num> 15716 </query_num>
<text>   ta or a successful compromise of the storage server can prevent an authorized user from accessing data,k but the attacker will not be able to recover the data itself as it is encrypted on disk. PASIS =-=[34]-=-, described in more detail later, provides a method to ensure availability, making it an exception to this rule. Performance within a cryptographic storage system is largely dependent on the type of c ally, PLUTUS employs a key rotation scheme that automates the key management process and limits the amount of individual user management. Figure 1. PLUTUS key system As an alternative solution, PASIS =-=[34]-=- uses a threshold secret sharing mechanism to separate data among several servers based on the assumption of independent failures. If fewer servers than configurable thresholds fail, no data loss or l a protection necessary for tamperproofing. 103 Figure 2. Relationship between tamperresistant and immutable techniques Cryptographic systems may provide tamper-proofing in addition to confidentiality =-=[34, 3, 15, 11]-=-. Contentaddressable storage systems [9, 10] inherently include tamper-proofing. Two general techniques exist: cryptographic hashes and cryptographic signatures. In a hashing system, a summary of the   </text>
<query_num> 15717 </query_num>
<text>   the management overhead of more complicated backup techniques. 3.3. Resilience to attack It is difficult to measure the strength of security techniques. Although frameworks attempting to do so exist =-=[22]-=-, they are mostly qualitative, and attempting to assign quantitative meaning to them seems arbitrary. Therefore, we attempt a qualitative comparison of the security properties of the different technol   </text>
<query_num> 15718 </query_num>
<text>   uted, while NCryptFS relies on a cumbersome timeout threshold, forcing users to periodically re-authenticate themselves during operations. Similar to the file system solutions, PLUTUS [15] and SiRiUS =-=[11]-=- implement a cryptographic storage system over a standard remote file system. The work of encryption, decryption, and signing is done on the hosts, allowing a scalability that is not possible with a c a protection necessary for tamperproofing. 103 Figure 2. Relationship between tamperresistant and immutable techniques Cryptographic systems may provide tamper-proofing in addition to confidentiality =-=[34, 3, 15, 11]-=-. Contentaddressable storage systems [9, 10] inherently include tamper-proofing. Two general techniques exist: cryptographic hashes and cryptographic signatures. In a hashing system, a summary of the   </text>
</top>
<top>
<paper_num> 158 </paper_num>
<paper_title>   Learning explicit and implicit visual manifolds by information projection.  </paper_title>
<paper_abstract>   Natural images have a vast amount of visual patterns distributed in a wide spectrum of subspaces of varying complexities and dimensions. Understanding the characteristics of these subspaces and their compositional structures is of fundamental importance for pattern modeling, learning and recognition. In this paper, we start with small image patches and define two types of atomic subspaces: explicit manifolds of low dimensions for structural primitives and implicit manifolds of high dimensions for stochastic textures. Then we present an information theoretical learning framework that derives common models for these manifolds through information projection, and study a manifold pursuit algorithm that clusters image patches into those atomic subspaces and ranks them according to their information gains. We further show how those atomic subspaces change over an image scaling process and how they are composed to form larger and more complex image patterns. Finally, we integrate the implicit and explicit manifolds to form a primal sketch model as a generic representation in early vision and to generate a hybrid image template representation for object category recognition in high level vision. The study of the mathematical structures in the image space sheds lights on some basic questions in human vision, such as atomic elements in visual perception, the perceptual metrics in various manifolds, and the perceptual transitions over image scales.  </paper_abstract>
<query_num> 15801 </query_num>
<text>   Fourthly, we present two case studies in Section 5 that integrate the implicit and explicit manifolds for image representation. One is the primal sketch model for generic images at the middle level (=-=Guo et al., 2007-=-), and the other is the mixed templates for object categories (=-=Si et al., 2009-=-). The two cases demonstrate that the two atomic manifold can be combined to represent general images. Finally we conclude ero-crossings unsuccessfully and his effort was mostly limited by the lack of proper models of texture. In Section 5.1, we will present a mathematical model for primal sketch based on our early work (=-=Guo et al., 2007-=-), which integrates the implicit and explicit manifolds seamlessly. We refer to two early papers on texton (=-=Zhu et al., 2005-=-) and primal sketch (=-=Guo et al., 2007-=-) for detailed discussions. In summary,   </text>
<query_num> 15802 </query_num>
<text>   es of a vehicle taken by motion camera, lie in low dimensional appearance manifolds. This is the pillow of many well-known dimension reduction techniques, such as Isomap, local linear embedding (LLE)(=-=Roweis and Saul, 2000-=-). People who applied LLE to image patches cropped from daily photos will be disappointed. The reason is intuitively discussed in the previous subsection, the image space is not a low dimensional mani   </text>
<query_num> 15803 </query_num>
<text>   ils than Gaussian, and often remains invariant when the images are down-scaled. This has inspired much research in the 1990s and early 2000s studying the statistics of natural images (=-=Ruderman, 1994; Zhu and Mumford, 1997; Huang and Mumford, 1999; Lee et al., 2003; Mumford and Gidas, 2001-=-). Here, by natural images, people usually mean photos taken in natural scenes which have a rich set of objects of various sizes in   </text>
<query_num> 15804 </query_num>
<text>   in previous sections. 2, As a discriminative representation, many image features are extracted for objects, the most popular one in recognition is the HoG template (Histogram of oriented Gradients) (=-=Dalal and Triggs, 2005-=-), and recently part based HoG models are also studied (=-=Felzenszwalb et al., 2009-=-). The HoG representation divides the image domain into regular m × n grid with each cell being a small image patch, fo   </text>
<query_num> 15805 </query_num>
<text>   lent. In today’s terminology, a texture is a set of images that share the same feature statistics H(I) = h. This is the implicit manifold that we defined in Equation (2) and named the Julesz ensemble(=-=Zhu et al., 2000-=-). Julesz’s texture quest was not very fruitful, since there was very limited knowledge about the neural functions (such as Gabor filters) in selecting the features and statistics h. Given some statis RAME model for texture modeling In the first case study, we illustrate the pursuit of implicit manifolds for texture modeling following the work of FRAME model (=-=Zhu et al., 1997-=-) and Julesz ensemble (=-=Zhu et al., 2000-=-). The feature dictionary ∆im = {Fk} consists of Gabor sine and cosine filters, Laplacian of Gaussian (LoG) and gradient filters of varies sizes. The features extracted are filter responses 〈Fk(x, y),   </text>
<query_num> 15806 </query_num>
<text>   oundary between the upper part of nose and the rest of the face. For a scene where the elements have a narrow range of sizes, such as the maple scenes, at a critical scale, a catastrophic transition (=-=Wang and Zhu, 2008-=-) occurs when we discard all the shape variables in W (describing the explicit manifolds for leaves) and switch to a statistical description W− = h for implicit manifolds. In the latter case, h is an   </text>
<query_num> 15807 </query_num>
<text>   res are extracted for objects, the most popular one in recognition is the HoG template (Histogram of oriented Gradients) (=-=Dalal and Triggs, 2005-=-), and recently part based HoG models are also studied (=-=Felzenszwalb et al., 2009-=-). The HoG representation divides the image domain into regular m × n grid with each cell being a small image patch, for instance, 8 pixels. At each pixel, a gradient is calculated, and a histogram is   </text>
<query_num> 15808 </query_num>
<text>   seemingly humble request of finding distinct features for weak classifiers turns out to be very hard to meet for many object categories. For example, for detecting vehicles in streets using Adaboost (=-=Freund and Schapire, 1997-=-), we may run out of weak classifiers rather quickly. This approach sounds very similar to the Chinese herb clinics, which have been practiced for more than a thousand years. A herb clinic typically h   </text>
<query_num> 15809 </query_num>
<text>   vocabulary. They can then be recursively composed into more complicated image categories or visual words, which in turn serve as non-terminal nodes 46in an hierarchical AND-OR graph representation (=-=Zhu and Mumford, 2006-=-). In this representation, the AND node denotes the composition of its components, while the OR node denotes multiple ways of compositions. We refer to Zhu and Mumford (2006) for a lengthy discussion   </text>
</top>
<top>
<paper_num> 159 </paper_num>
<paper_title>   Enabling large-scale wireless broadband: the case for TAPs.  </paper_title>
<paper_abstract>   The vision is tantalizing: a high-performance, scalable, and widely deployed wireless Internet that facilitates services ranging from radically new and unforeseen applications to true wireless &amp;quot;broadband&amp;quot; to residences and public spaces at rates of 10s of Mb/sec. However, while high-speed wireless access is easy to achieve in an enterprise network via low-cost IEEE 802.11 (WiFi) access points, wireless technology in public spaces is in its infancy. &amp;quot;Hot spots&amp;quot; provide high-speed wireless access, but do so in very few isolated &amp;quot;islands&amp;quot; at immense costs. Likewise, while fixed wireless (e.g. LMDS) and 3G can provide ubiquitous coverage and 3G can support mobility, throughputs can often be two orders of magnitude slower than WiFi.  </paper_abstract>
<query_num> 15901 </query_num>
<text>   -like network where packet forwarding of wireless APs and user access compete for the scarce spectrum, pushing the system capacity dramatically to the well-known scalability limits of ad hoc networks =-=[9]-=-. We envision therefore an architecture as depicted in Figure 1, where TAPs equipped with sector antennas allow for geographically focused as well as omni-directional transmission of data. Being equip e key to a high-performance scalable system is to ensure that packets consume minimal system resources to reach their destination. In particular, the scalability limitations of purely ad hoc networks =-=[9]-=- arise because each forwarding hop consumes additional resources. Moreover, equally crucial scaling impediments can be observed in measurement studies [8], [16] which show that actual implementations   </text>
<query_num> 15902 </query_num>
<text>   Any medium access and scheduling decision requires distributed resource management rather than a purely local decision. Unlike schedulers designed for cellular and wired-AP networks, (e.g., [3], [5], =-=[11]-=-, [17]), scheduling in TAP networks is inherently a distributed operation. Nodes in the network are not aware of the channel conditions or queue backlog of other nodes. It is evident that naive exchan   </text>
<query_num> 15903 </query_num>
<text>   congestion control. However, relying on TCP alone is not enough. First, TCP&amp;apos;s congestion control has welldocumented performance limitations over both multi-hop and single-hop wireless networks (e.g., =-=[2]-=-, [6], [10]). Second, TCP&amp;apos;s congestion control necessarily operates at end-to-end timescales of 100s of milliseconds -- too coarse to address the fast timescale dynamics of contention and realistic ch   </text>
<query_num> 15904 </query_num>
<text>   control. However, relying on TCP alone is not enough. First, TCP&amp;apos;s congestion control has welldocumented performance limitations over both multi-hop and single-hop wireless networks (e.g., [2], [6], =-=[10]-=-). Second, TCP&amp;apos;s congestion control necessarily operates at end-to-end timescales of 100s of milliseconds -- too coarse to address the fast timescale dynamics of contention and realistic channels. Fin   </text>
<query_num> 15905 </query_num>
<text>   e 4: Any medium access and scheduling decision requires distributed resource management rather than a purely local decision. Unlike schedulers designed for cellular and wired-AP networks, (e.g., [3], =-=[5]-=-, [11], [17]), scheduling in TAP networks is inherently a distributed operation. Nodes in the network are not aware of the channel conditions or queue backlog of other nodes. It is evident that naive   </text>
<query_num> 15906 </query_num>
<text>   e at that timescale. A fundamental bound on the capacity of beamforming for a system with M transmit antenna elements and a single receive antenna using B bits of channel information was presented in =-=[21]-=-. First of its kind, this bound uses no asymptotic approximations and is thus valid for all practical cases of interest. These results can form the basis to study the relationship between channel cohe   </text>
<query_num> 15907 </query_num>
<text>   e model for achieving fairness and spatial reuse in TAP networks. This TAP-aggregated fairness model differs fundamentally from both proportional fairness as approximately achieved by TCP [12], [19], =-=[20]-=- and max-min fairness as targeted by some ATM congestion control algorithms [14]. While the solution to achieve this desired reference model for the scenario of Figure 2 is immediate, the general case   </text>
<query_num> 15908 </query_num>
<text>   e-art beamforming techniques [4] assume that only either sender or receiver are equipped with multiple antenna elements, but the TAP architecture assumes both. Second, MIMO space-time encoding (e.g., =-=[26]-=-) assumes that antenna elements are spaced sufficiently far apart to create independent fading at each element so that the antenna beam patterns are not focused, whereas TAPs require focusing. Thus, u   </text>
<query_num> 15909 </query_num>
<text>   equally crucial scaling impediments can be observed in measurement studies [8], [16] which show that actual implementations perform significantly worse than the predicted information theoretic bounds =-=[7]-=-, [9] because they assume perfect &amp;quot;zero-overhead&amp;quot; protocols. Thus, while representing an important step in understanding the behavior of large-scale wireless ad hoc networks, existing theoretical capa   </text>
<query_num> 15910 </query_num>
<text>   ference model for achieving fairness and spatial reuse in TAP networks. This TAP-aggregated fairness model differs fundamentally from both proportional fairness as approximately achieved by TCP [12], =-=[19]-=-, [20] and max-min fairness as targeted by some ATM congestion control algorithms [14]. While the solution to achieve this desired reference model for the scenario of Figure 2 is immediate, the genera   </text>
<query_num> 15911 </query_num>
<text>   ion. Likewise, significant progress has been made in distributed media access and scheduling algorithms designed to balance fairness and spatial reuse objectives in ad hoc networks (e.g., [18], [22], =-=[27]-=-). There are two critical aspects of the TAP network that require a fundamentally new look at distributed resource allocation. First, the network has a distinct structure as compared to general ad hoc   </text>
<query_num> 15912 </query_num>
<text>   location. Likewise, significant progress has been made in distributed media access and scheduling algorithms designed to balance fairness and spatial reuse objectives in ad hoc networks (e.g., [18], =-=[22]-=-, [27]). There are two critical aspects of the TAP network that require a fundamentally new look at distributed resource allocation. First, the network has a distinct structure as compared to general   </text>
<query_num> 15913 </query_num>
<text>   oordinated schedulers has been developed that allows packets that are &amp;quot;late&amp;quot; or under-serviced upstream to catch up at downstream nodes by coordinating a packet&amp;apos;s priority index across multiple nodes =-=[15]-=-. In TAP networks, multihop coordination to best achieve system-wide performance objectives must take variable channel conditions into account and must interact with the random access MAC protocol. IV   </text>
<query_num> 15914 </query_num>
<text>   patial location. Likewise, significant progress has been made in distributed media access and scheduling algorithms designed to balance fairness and spatial reuse objectives in ad hoc networks (e.g., =-=[18]-=-, [22], [27]). There are two critical aspects of the TAP network that require a fundamentally new look at distributed resource allocation. First, the network has a distinct structure as compared to ge   </text>
<query_num> 15915 </query_num>
<text>   stion control. However, relying on TCP alone is not enough. First, TCP&amp;apos;s congestion control has welldocumented performance limitations over both multi-hop and single-hop wireless networks (e.g., [2], =-=[6]-=-, [10]). Second, TCP&amp;apos;s congestion control necessarily operates at end-to-end timescales of 100s of milliseconds -- too coarse to address the fast timescale dynamics of contention and realistic channel   </text>
<query_num> 15916 </query_num>
<text>   ty limitations of purely ad hoc networks [9] arise because each forwarding hop consumes additional resources. Moreover, equally crucial scaling impediments can be observed in measurement studies [8], =-=[16]-=- which show that actual implementations perform significantly worse than the predicted information theoretic bounds [7], [9] because they assume perfect &amp;quot;zero-overhead&amp;quot; protocols. Thus, while represen he nearest TAP(s) and can be restricted to traverse the MU&amp;apos;s local neighborhood. This restriction bounds the average path length traversed by route requests, resulting in improved traffic scalability =-=[16]-=-, and providing a foundation for scalable routing. Moreover, by exploiting the network is the channel framework, the overhead in discovering new and better routes can be balanced with the quality of t   </text>
</top>
<top>
<paper_num> 160 </paper_num>
<paper_title>   Resilience of mutual exclusion algorithms to transient memory faults.  </paper_title>
<paper_abstract>   We study the behavior of mutual exclusion algorithms in the presence of unreliable shared memory subject to transient memory faults. It is well-known that classical 2-process mutual exclusion algorithms, such as Dekker and Peterson’s algorithms, are not faulttolerant; in this paper we ask what degree of fault tolerance can be achieved using the same restricted resources as Dekker and Peterson’s algorithms, namely, three binary read/write registers. We show that if one memory fault can occur, it is not possible to guarantee both mutual exclusion and deadlock-freedom using three binary registers; this holds in general when fewer than2f +1 binary registers are used and f may be faulty. Hence we focus on algorithms that guarantee (a) mutual exclusion and starvationfreedom in fault-free executions, and (b) only mutual exclusion in faulty executions. We show that using only three binary registers it is possible to design an 2-process mutual exclusion algorithm which tolerates a single memory fault in this manner. Further, by replacing one read/write register with a test&amp;set register, we can guarantee mutual exclusion in executions where one variable experiences unboundedly many faults. In the more general setting where up tof registers may be faulty, we show that it is not possible to guarantee mutual exclusion using 2f +1 binary read/write registers if each faulty register can exhibit unboundedly many faults. On the positive side, we show that an n-variable single-fault tolerant algorithm satisfying certain conditions can be transformed into an ((n − 1)f + 1)-variable f-fault tolerant algorithm with the same progress guarantee as the original. In combination with our three-variable algorithm, this implies that there is a(2f+1)-variable mutual exclusion algorithm tolerating a single fault in up tof variables without violating mutual exclusion.  </paper_abstract>
<query_num> 16001 </query_num>
<text>   ability and refinement was already explored in [12], where the notion of linearizability was first introduced, and it has also been extensively studied in the formal methods community recently, e.g., =-=[9, 15]-=-). We saw above that we can choose any register and linearize all operations when they last access it. In particular, we can choose the last register accessed, xf. When we embed linearization points u   </text>
<query_num> 16002 </query_num>
<text>   able memories has also attracted interest. Problems such as fault-resilient selection, sorting, and matrix computations in various failure models have attracted a lot of interest in recent years (see =-=[10]-=- for a survey). Faulty memory has also been studied in multiprocessors. There is significant research in the parallel computing literature devoted to deliver general simulation mechanisms of fully ope   </text>
<query_num> 16003 </query_num>
<text>   e-only solutions such as error-correcting codes, watchdog co-processors or redundant hardware threads (e.g. [16, 17]) as well as software-only techniques that use both single and multiple cores (e.g. =-=[20, 18]-=-). These solutions are typically “heavy-weight” and quite costly in terms of memory and performance. Resilient algorithms. In the area of algorithms, designing resilient algorithms for unreliable memo   </text>
<query_num> 16004 </query_num>
<text>   ence a memory fault. In the sequel we use Read and Write to denote high-level operations on x (as opposed to low-level operations onx1,...,xf). To be useful, our implementation should be linearizable =-=[12]-=-: the operations invoked on x should appear to take place instantaneously, as though the algorithm were accessing an atomic faulty read/write register. However, unlike many fault-tolerant simulations  formally, the linearization induces a trace simulation between the lowlevel implementation to the high-level algorithm. The relationship between linearizability and refinement was already explored in =-=[12]-=-, where the notion of linearizability was first introduced, and it has also been extensively studied in the formal methods community recently, e.g., [9, 15]). We saw above that we can choose any regis   </text>
<query_num> 16005 </query_num>
<text>   f which at most one is faulty. Fault-tolerant mutual exclusion. The issue of fault-tolerance in mutual exclusion algorithms was one of the principal themes of Lamport’s paper on non-atomic algorithms =-=[14]-=-. Several failure models are considered. Among many other malfunctions, one failure type studied are transient faults, which allows arbitrary changes to the shared memory (and local) variables of the   </text>
<query_num> 16006 </query_num>
<text>   ns. In the shared memory distributed computing literature, the problem of implementing fault-tolerant registers (and other objects) from faulty objects under various fault models was studied, e.g. in =-=[1, 2, 13]-=-. With regard to our simulation in Section 8, the most relevant results are the ones given in [1] and [13] on implementing various read/write registers from faulty registers in the arbitrary, responsi   </text>
<query_num> 16007 </query_num>
<text>   ns. In the shared memory distributed computing literature, the problem of implementing fault-tolerant registers (and other objects) from faulty objects under various fault models was studied, e.g. in =-=[1, 2, 13]-=-. With regard to our simulation in Section 8, the most relevant results are the ones given in [1] and [13] on implementing various read/write registers from faulty registers in the arbitrary, responsi ntext, however, the simulation in Section 8 serves a different purpose; we do not seek to mask faults completely, as the high-level 1 Much better results are known for more benign failure modes, e.g. =-=[13, 11]-=-mutual exclusion algorithm that uses the objects can tolerate some degree of faulty behavior. Instead, we seek to reduce f faults to a single fault, which can then be handled by the algorithm. Togeth   </text>
<query_num> 16008 </query_num>
<text>   ocessors. There is significant research in the parallel computing literature devoted to deliver general simulation mechanisms of fully operational parallel machines on their faulty counterparts, e.g. =-=[7, 8]-=-. Fault-tolerant simulations. In the shared memory distributed computing literature, the problem of implementing fault-tolerant registers (and other objects) from faulty objects under various fault mo   </text>
<query_num> 16009 </query_num>
<text>   rithm can accomplish this. THEOREM 4.1. No deadlock-free mutual exclusion algorithm that uses two binary variables can satisfy the HD2 property. This result is similar in spirit to the lower bound of =-=[6]-=-, which shows that n shared variables are necessary for n-process mutual exclusion; in other words, each process must have a variable that it “owns” in some sense. Technically, however, the proof of T   </text>
<query_num> 16010 </query_num>
<text>   t that fault rates will further increase in the future. Faster clock rates, increasing transistor density, decreasing voltages and smaller feature sizes all contribute to increasing fault rates, e.g. =-=[3, 22]-=-. In fact, fault rates in modern processors have been increasing at a rate of approximately 8% per generation [5]. To counter soft errors, computer architects and compiler researchers have proposed va   </text>
<query_num> 16011 </query_num>
<text>   y to computations in one way or another. For instance, there are proposals involving hardware-only solutions such as error-correcting codes, watchdog co-processors or redundant hardware threads (e.g. =-=[16, 17]-=-) as well as software-only techniques that use both single and multiple cores (e.g. [20, 18]). These solutions are typically “heavy-weight” and quite costly in terms of memory and performance. Resilie   </text>
</top>
<top>
<paper_num> 161 </paper_num>
<paper_title>   Resource recycling: putting idle resources to work on a composable accelerator.  </paper_title>
<paper_abstract>   Mobile computing platforms in the form of smart phones, netbooks, and personal digital assistants have become an integral part of our everyday lives. Moving ahead to the future, mobile multimedia support will become a key differentiating factor for customers. Features such as high-definition audio and video, video conferencing, 3D graphics, and image projection will lead to the adoption of one phone over another. However, in contrast to wireless signal processing which is dominated by vectorizable computation, mobile multimedia applications often contain complex control flow and variable computational requirements. Moreover, data access is more complex where media applications typically operate on multi-dimensional vectors of data rather than single-dimensional vectors with simple strides. To handle these complexities, composable accelerators such as the Polymorphic Pipeline Array, or  </paper_abstract>
<query_num> 16101 </query_num>
<text>   age. Our work also considers composable architecture specific features such as resource conflict and reconfiguration overhead whereas these works targeted fixed multi-core solutions(RAW architectures =-=[14]-=- and Cell processors [9]). Resource borrowing on dynamic partition is a similar concept to Work stealing [3] but our approach is performed in more fine-grained level, not thread level. 6. CONCLUSION T   </text>
<query_num> 16102 </query_num>
<text>   ead performance in mobile environments where power consumption and hardware cost is a first-class constraint. The building blocks of PPA are simple in-order cores similar to clustered VLIW processors =-=[25]-=-. Also, the statically controlled point-to-point interconnect provides a fast inter-core communication, allowing PPA to exploit fine grain pipeline parallelism efficiently for multimedia applications.   </text>
<query_num> 16103 </query_num>
<text>   es enter the mainstream. Even this type of parallelism has many advantages compared to other types of parallelism, adapting in real situation is difficult because of program-inherent data dependences =-=[22]-=-. To overcome this difficulty, [22] has proposed a dynamic analysis tool to extract a stream graph from legacy C code in order to give a programmer hints for manual parallelization. [22] also tries lo   </text>
<query_num> 16104 </query_num>
<text>   fter the hardware is constructed. For wireless signal processing, programmable designs that exploit high degrees of single-instruction multiple-data (SIMD) parallelism have emerged to challenge ASICs =-=[2, 1, 5, 15, 24]-=-. While these solutions suffice for wireless signal processing, multimedia applications contain more complex data dependence patterns and frequent control flow for which wide-SIMD is inefficient. Thus   </text>
<query_num> 16105 </query_num>
<text>   ilize the VLIW part of the array. So, it cannot pipeline the application in a coarser granularity as PPA. With identical resources, PPA outperforms our best approximation of ADRES by 1.43x. PipeRench =-=[6]-=- is a 1-D architecture in which processing elements are arranged in stripes to facilitate pipelining, but it has a fixed configuration of resource partitioning for pipelining while PPA can partition t   </text>
<query_num> 16106 </query_num>
<text>   ons because the hardware is more flexible and can accelerate the code in multiple ways [19]. Coarse-grain pipeline parallelism is exploited by concurrently executing filters in streaming applications =-=[7, 8, 12]-=-, as well as finegrain instruction level parallelism is also found by modulo scheduling innermost loops [21]. A PPA is a generalization of a coarsegrain reconfigurable architecture (CGRA) shown in Fig   </text>
<query_num> 16107 </query_num>
<text>   ons because the hardware is more flexible and can accelerate the code in multiple ways [19]. Coarse-grain pipeline parallelism is exploited by concurrently executing filters in streaming applications =-=[7, 8, 12]-=-, as well as finegrain instruction level parallelism is also found by modulo scheduling innermost loops [21]. A PPA is a generalization of a coarsegrain reconfigurable architecture (CGRA) shown in Fig hitectures, and propose a new compilation process to solve the difficulties. In this framework, we adapt the key concept from the stream graph modulo scheduling algorithm for coarse-grain parallelism =-=[12]-=-. The main difference is that parallel composition of the each filter is not performed with split-joins, but by modulo scheduling across larger core-groups. With this change, the PPA compiler can be u on different number of resources. 2.2 Stream Graph Modulo Scheduling This paper presents a compiler technique specifically for composable accelerators based on stream graph modulo scheduling, or SGMS =-=[12]-=-. SGMS is a modulo scheduling algorithm for mapping streaming applications onto multicore systems. Modulo scheduling is traditionally a form of software pipelining applied at the instruction level to  rder to give a programmer hints for manual parallelization. [22] also tries load balancing by changing a program but this paper’s focus is more on compile time optimization for given program. [8] and =-=[12]-=- are similar to this paper to exploit coarse-grained pipeline parallelism but the parallelization mechanism is limited only to stateless components as using StreamIt language. Our work also considers   </text>
<query_num> 16108 </query_num>
<text>   ons because the hardware is more flexible and can accelerate the code in multiple ways [19]. Coarse-grain pipeline parallelism is exploited by concurrently executing filters in streaming applications =-=[7, 8, 12]-=-, as well as finegrain instruction level parallelism is also found by modulo scheduling innermost loops [21]. A PPA is a generalization of a coarsegrain reconfigurable architecture (CGRA) shown in Fig t of autonomous actors (also called filters) that operate on data and communicate through firstin first-out data channels [23]. During program execution, actors fire repeatedly in a periodic schedule =-=[8]-=-. Each actor has a separate instruction stream and an independent address space, thus all dependences between actors are made explicit through the communication channels. Compilers can leverage these  ode in order to give a programmer hints for manual parallelization. [22] also tries load balancing by changing a program but this paper’s focus is more on compile time optimization for given program. =-=[8]-=- and [12] are similar to this paper to exploit coarse-grained pipeline parallelism but the parallelization mechanism is limited only to stateless components as using StreamIt language. Our work also c   </text>
<query_num> 16109 </query_num>
<text>   parallelism is exploited by concurrently executing filters in streaming applications [7, 8, 12], as well as finegrain instruction level parallelism is also found by modulo scheduling innermost loops =-=[21]-=-. A PPA is a generalization of a coarsegrain reconfigurable architecture (CGRA) shown in Figure 1 [17]. It consists of an array of simple processing elements (PEs) that are tightly interconnected by a nally a form of software pipelining applied at the instruction level to find a valid schedule for a loop such that the interval between successive iterations (initiation interval, or II) is minimized =-=[21]-=-. SGMS is the same technique on a coarse-grain stream graph to pipeline the actors across multiple cores. The objective is to maximize concurrent execution of actors while hiding communication overhea   </text>
<query_num> 16110 </query_num>
<text>   s work, we focus on stream-style C code where a program is represented as a set of autonomous actors (also called filters) that operate on data and communicate through firstin first-out data channels =-=[23]-=-. During program execution, actors fire repeatedly in a periodic schedule [8]. Each actor has a separate instruction stream and an independent address space, thus all dependences between actors are ma eduling across larger core-groups. With this change, the PPA compiler can be used for more generic code by removing the restrictions of static data rates on stream programming languages like StreamIt =-=[23]-=-. Edge-centric modulo scheduling (EMS) [18], which focuses on routing of values between functional units, is used as the modulo scheduling technique for exploiting fine-grain parallelism. The compilat   </text>
<query_num> 16111 </query_num>
<text>   tween the granularity of parallelism in workloads and the granularity of processing cores inspired a flexible execution model that allows the aggregation of small cores to create larger logical cores =-=[11]-=-,[10]. Composable accelerators are multi-core accelerator designs that incorporate this flexible execution model in embedded systems. Multiple small cores enable the parallel execution of individual t n the fastest 5. RELATED WORK Architectures: Combining cores to create a bigger logical core is relatively a new technique, recently proposed by Core fusion [10] and Composable Lightweight Processors =-=[11]-=-. Core Fusion is a CMP architecture that can dynamically allocate independent cores together for a single thread execution maintaining ISA compatibility. CLPs also allows dynamic allocation of cores t   </text>
<query_num> 16112 </query_num>
<text>   uration overhead whereas these works targeted fixed multi-core solutions(RAW architectures [14] and Cell processors [9]). Resource borrowing on dynamic partition is a similar concept to Work stealing =-=[3]-=- but our approach is performed in more fine-grained level, not thread level. 6. CONCLUSION The popularity of mobile computing platforms has led to the development of feature packed devices that suppor   </text>
</top>
<top>
<paper_num> 162 </paper_num>
<paper_title>   Multiscale symmetric part detection and grouping.  </paper_title>
<paper_abstract>   Skeletonization algorithms typically decompose an object’s silhouette into a set of symmetric parts, offering a powerful representation for shape categorization. However, having access to an object’s silhouette assumes correct figure-ground segmentation, leading to a disconnect with the mainstream categorization community, which attempts to recognize objects from cluttered images. In this paper, we present a novel approach to recovering and grouping the symmetric parts of an object from a cluttered scene. We begin by using a multiresolution superpixel segmentation to generate medial point hypotheses, and use a learned affinity function to perceptually group nearby medial points likely to belong to the same medial branch. In the next stage, we learn higher granularity affinity functions to group the resulting medial branches likely to belong to the same object. The resulting framework yields a skeletal approximation that’s free of many of the instabilities plaguing traditional skeletons. More importantly, it doesn’t require a closed contour, enabling the application of skeleton-based categorization systems to more realistic imagery. 1.  </paper_abstract>
<query_num> 16201 </query_num>
<text>   (h) in contrast, classical skeletonization algorithms require a closed contour which, for real images, must be approximated by a region boundary. In this case, the parameters of the N-cuts algorithm =-=[17]-=- were tuned to give the best region (maximal size without region undersegmentation) for the swimmer. A standard medial axis extraction algorithm applied to the smoothed silhouette produces a skeleton  occur; as can be seen in Fig. 1(b), we segment an image into 25, 50, 100 and 200 superpixels. To generate superpixels at each scale, we employ a modified version [12] of the normalized cuts algorithm =-=[17]-=- since it yields compact superpixels. Each superpixel segmentation yields a superpixel graph, where nodes represent superpixels and edges represent superpixel adjacencies. If a superpixel represents a   </text>
<query_num> 16202 </query_num>
<text>   ature in the perceptual grouping community. Many computational models exist for symmetry-based grouping, including Brady and Asada [5], Cham and Cipolla [6], SaintMarc et al. [15], Ylä-Jääski and Ade =-=[21]-=- and, more re-cently, Stahl and Wang [20], among others. Such systems face one or more important limitations: 1) the complexity of pairwise contour grouping to detect symmetry-related contour pairs;   </text>
<query_num> 16203 </query_num>
<text>   ial axis transform [3] decomposes a closed 2-D shape into a set of skeletal parts and their connections, providing a powerful parts-based decomposition of the shape that’s suitable for shape matching =-=[19, 16]-=-. While the medial axis-based research community is both active and diverse, it has not kept pace with the mainstream object recognition (categorization) community that seeks to recognize objects from   </text>
<query_num> 16204 </query_num>
<text>   ies) as local maxima in a Laplacian pyramid, linked together by spatial overlap to form a tree structure. Object matching was then formulated as comparing paths through two trees. Shokoufandeh et al. =-=[18]-=- proposed a more elaborate matching framework based on Lindeberg’s multiscale blob model [10]. This family of approaches can be characterized as imposing a strong part-based symmetry prior, detecting   </text>
<query_num> 16205 </query_num>
<text>   nt scales at which we expect parts to occur; as can be seen in Fig. 1(b), we segment an image into 25, 50, 100 and 200 superpixels. To generate superpixels at each scale, we employ a modified version =-=[12]-=- of the normalized cuts algorithm [17] since it yields compact superpixels. Each superpixel segmentation yields a superpixel graph, where nodes represent superpixels and edges represent superpixel adj   </text>
<query_num> 16206 </query_num>
<text>   obtain the final pairwise region affin1 ity, converted to edge weights as W (i, j) = As(i,j) . The resulting graph is used in conjunction with an efficient agglomerative clustering algorithm based on =-=[8]-=- (complexity: O(|S|), where S is a set of all superpixels) to obtain medial parts (medial point clusters). The clustering algorithm initializes all medial point hypotheses as singletons, and maintains , j) (3) + [Jij = 2] · A2(i, j) Having defined all the components of the affinity function Ap(i, j) (Equation 1), we use these affinities to cluster parts that are attached. We use the same algorithm =-=[8]-=- used to cluster medial points into parts. 4.2. Medial Part Selection Our affinity-based grouping yields a set of part clusters, each presumed to correspond to a set of attached parts belonging to a s   </text>
<query_num> 16207 </query_num>
<text>   ph-based segmentation algorithm, with abstracted symmetry axes overlaid onto the abstracted parts; (g) in contrast, a Laplacian-based multiscale blob and ridge decomposition, such as that computed by =-=[10]-=-, shown, yields many false positive and false negative parts; (h) in contrast, classical skeletonization algorithms require a closed contour which, for real images, must be approximated by a region bo s spurious branches, branch instability, and poor part delineation. Our approach offers clear advantages over competing approaches. For example, classical multiscale blob and ridge detectors, such as =-=[10]-=- (Fig. 1(g)), yield many spurious parts, a challenging form of noise for any graph-based indexing or matching strategy. And even if an opportunistic setting of a region segmenter’s parameters yields a e structure. Object matching was then formulated as comparing paths through two trees. Shokoufandeh et al. [18] proposed a more elaborate matching framework based on Lindeberg’s multiscale blob model =-=[10]-=-. This family of approaches can be characterized as imposing a strong part-based symmetry prior, detecting parts at multiple scales, and grouping them based on a simple model of spatial proximity. How r system assumes parts with straight symmetry axes. Finally, to provide a quantitative evaluation of our part detection strategy, we compare its precision and recall to the method of Lindeberg et al. =-=[10]-=-, used to generate the symmetric parts shown in Fig. 1(g). Both methods are evaluated on 61 test images from the Weizmann Horse Dataset [4]. A ground truth part is considered to be recovered if its no k centered at the centroid), providing a more standard skeletal description. a threshold (0.4). Our part detection offers a significant improvement in both precision and recall (Fig. 8). Moreover, in =-=[10]-=-, no effort is made to distinguish part occlusion from part attachment; parts are simply grouped if they overlap. Note that both methods achieve low precision. This is partially due to the fact that t   </text>
<query_num> 16208 </query_num>
<text>   solving a relaxed quadratic programming problem, in which real values are rounded to 0 or 1 [14]. 5. Results To evaluate the method, we train the various components using the Weizmann Horse Database =-=[4]-=-, consisting of images of horses together with figure-ground segmentations; in addition, we manually mark the elongated parts of the horses, together with their attachment relations. Fig. 6 illustrate  its precision and recall to the method of Lindeberg et al. [10], used to generate the symmetric parts shown in Fig. 1(g). Both methods are evaluated on 61 test images from the Weizmann Horse Dataset =-=[4]-=-. A ground truth part is considered to be recovered if its normalized overlap (in area) with one of the detected parts is above 1 Note that although we choose to display only the abstract representati   </text>
</top>
<top>
<paper_num> 163 </paper_num>
<paper_title>   Exploring the limits of leakage power reduction in caches.  </paper_title>
<paper_abstract>   If current technology scaling trends hold, leakage power dissipation will soon become the dominant source of power consumption. Caches, due to the fact that they account for the largest fraction of on-chip transistors in most modern processors, are a primary candidate for attacking the leakage problem. While there has been a flurry of research in this area over the Last several years, a major question remains unanswered. What is the total potential of existing architectural and circuit techniques to address this important design concern? In this paper, we explore the limits in which existing circuit and architecture technologies may address this growing problem. We first formally propose a parameterized model that can determine the optimal leakage savings based on the perfect knowledge of the address trace. By carefully applying the sleep and drowsy modes, we find that the total leakage power from the L1 instruction cache, data cache, and a unified L2 cache may be reduced to mere 3.6%, 0.9%, and 2.3%, respectively, of the unoptimized case. We further study how such a model can be extended to obtain the optimal leakage power savings for different cache configurations.  </paper_abstract>
<query_num> 16301 </query_num>
<text>   DRI-cache =-=[Powell et al. 2001]-=- uses the Gated-Vdd technique to dynamically adjust the size of the active portion of the cache by turning off a bank of cache lines based on the miss rates. DRG-cache [=-=Agarwal et al. 2002-=-] employs Gated-Vdd to reduce leakage power by turning off the gated-Ground transistor, while data is restored when the gated-Ground transistor is turned on. DTSRAM =-=[H.Kim and Roy 2002]-=- uses body bias   </text>
<query_num> 16302 </query_num>
<text>   body biasing to separately control the Vt of each cache line. To minimize the energy and delay overhead, a cache line is switched to high Vt when it is not likely to be used anymore. Kaxiras et al. [=-=Hu et al. 2002; Kaxiras et al. 2001-=-] proposed the cache line decay scheme to turn off the cache lines in the dead periods of their cache generations using the Gated-Vdd technique. Instead of placing both the tag an   </text>
<query_num> 16303 </query_num>
<text>   cache banks are not accessed. Hanson [=-=Hanson et al. 2001-=-] found that for L1 caches, MTCMOS, which is a state-preserving technique that operates multiple threshold voltages, outperforms Gated-Vdd. In [=-=Li et al. 2003-=-], the authors presented several architectural techniques that exploit the data duplication across the different levels of cache hierarchy. They found that the best strategy in terms of energy and ene   </text>
<query_num> 16304 </query_num>
<text>   d: either dynamically (due to the switching activity of repeated capacitance charge and discharge on the output of the millions of gates), or statically (mainly due to sub-threshold and gate leakage [=-=Kim et al. 2003; Rabaey et al. 2002-=-]). Dynamic power consumption is proportional to the square of the supply voltage, which reduces as process technology scales. While the scaling down of transistor geometries enabl   </text>
<query_num> 16305 </query_num>
<text>   edbackcontrol theory to adaptively adjust the cache decay interval and cache lines are turned off accordingly. Another approach to reducing leakage power is called drowsy cache =-=[Flautner et al. 2002; Kim et al. 2004; 2002]-=-, which decreases the supply voltage of idle cache lines. Specifically, all cache lines are periodically placed into drowsy mode. =-=[Kim and Mudge 2004]-=- studied techniques for data retention with  The stall will lead to significant energy consumption as the big circle indicates. Similar things happen to the drowsy mode. But the drowsy mode preserves the data and only takes a couple of cycles [=-=Kim et al. 2004-=-] to wake up the cache line. So, without just-in-time refetch (Figure 4(d)), the amount of energy the drowsy mode consumes is less than that of the sleep mode during the system stalling, which is indi ell size is usually smaller than a L1 SRAM cell. The area overhead of implementing either drowsy or sleep technique in L2 caches is about 5-8%, and sleep mode may cause instability of L2 memory cells[=-=Kim et al. 2004-=-]. However, as machines and working-sets grow, the L2 caches are becoming increasingly performance critical, and integrated even closer to the processor (especially when a third level of hierarchy is   </text>
<query_num> 16306 </query_num>
<text>   gy in terms of energy and energydelay product is to place the L2 subblock into a state-preserving mode as soon as its contents are moved to L1 and to reactive it only when it is accessed. Bai et al. [=-=Bai et al. 2005-=-] investigated the impact of Tox and Vth on power performance tradeoffs for on-chip caches. In contrast, =-=[Sankaranarayanan and Skadron 2004; Zhang et al. ACM Transactions on Architecture and Code Opti -=-  </text>
<query_num> 16307 </query_num>
<text>   s with HotLeakage and showed Gated-Vdd is superior for a set of faster L2 latencies. Heo =-=[Heo et al. 2002]-=- reduced bitline leakage by leaving bitlines open whose cache banks are not accessed. Hanson [=-=Hanson et al. 2001-=-] found that for L1 caches, MTCMOS, which is a state-preserving technique that operates multiple threshold voltages, outperforms Gated-Vdd. In =-=[Li et al. 2003]-=-, the authors presented several architect   </text>
<query_num> 16308 </query_num>
<text>   separately control the Vt of each cache line. To minimize the energy and delay overhead, a cache line is switched to high Vt when it is not likely to be used anymore. Kaxiras et al. [=-=Hu et al. 2002; Kaxiras et al. 2001-=-] proposed the cache line decay scheme to turn off the cache lines in the dead periods of their cache generations using the Gated-Vdd technique. Instead of placing both the tag and the data into the s s to the add instructions depends on the range of the inner loop |high(i) − low(i)|. the cache, it is regarded as dead. Besides turning off cache lines in dead periods as the cache decay scheme does [=-=Kaxiras et al. 2001-=-], our method also explores the live period of a cache generation, which demonstrates great potential for leakage reduction. In fact we found that dead periods did not contribute a large amount of lea the cache access interval using the sleep technique is s = s1 + s2 + s3 + s4, and that of using the drowsy mode is d = d1 + d2 + d3. For the sleep mode, the data has been lost due to an induced miss [=-=Kaxiras et al. 2001-=-] and must be re-fetched from the memory hierarchy. as such, there is a significant amount of power consumed by the dynamic activity required to fetch the data from the L2 cache, marked with “*” in Fi discussing the implementation of each technique, we define an access interval of a cache line as Ti. OPT-Drowsy puts the line into the drowsy 2 The sleep(10K) is similar to the cache-decay scheme in [=-=Kaxiras et al. 2001-=-], in which the decay interval was set to be 10K cycles, and the extra leakage power consumed by the counter per cache line was taken into account. ACM Transactions on Architecture and Code Optimizati her than OPTSleep(1M), and 31% higher than OPT-Drowsy, which indicates that there is still far more potential left in the existing techniques. 3 The sleep(1M) is similar to the cache-decay scheme in [=-=Kaxiras et al. 2001-=-], in which the decay interval was set to be 1M cycles, and the extra leakage power consumed by the counter per cache line was taken into account. ACM Transactions on Architecture and Code Optimizatio   </text>
<query_num> 16309 </query_num>
<text>   tents are moved to L1 and to reactive it only when it is accessed. Bai et al. [=-=Bai et al. 2005-=-] investigated the impact of Tox and Vth on power performance tradeoffs for on-chip caches. In contrast, [=-=Sankaranarayanan and Skadron 2004; Zhang et al. ACM Transactions on Architecture and Code Optimization, Vol. x, No. x, xx 20xx.s6 · Y. Meng, T. Sherwood and R. Kastner 2002-=-] studied software approaches. =-=[Sankaranarayanan and Skadron -=-  </text>
<query_num> 16310 </query_num>
<text>   the consumed leakage power is to “turn off” those transistors that are not needed. While this is the easiest to think about, it is by no means the easiest to implement. One such approach, Gated-Vdd [=-=Powell et al. 2001-=-], attempts to solve this problem by reducing leakage through the use of a high threshold sleep transistor (between pull-down NMOS and virtual Vss) to break the connection and thus increases the L1 ca ACCs exploit the fact that in ordinary programs most of the bits in caches are zeros for both the data and instruction streams, and provide significant leakage reduction in the zero state. DRI-cache [=-=Powell et al. 2001-=-] uses the Gated-Vdd technique to dynamically adjust the size of the active portion of the cache by turning off a bank of cache lines based on the miss rates. DRG-cache [=-=Agarwal et al. 2002-=-] employs G   </text>
<query_num> 16311 </query_num>
<text>   the supply voltage of idle cache lines. Specifically, all cache lines are periodically placed into drowsy mode. =-=[Kim and Mudge 2004]-=- studied techniques for data retention with lower supply voltage. [=-=Hu et al. 2003-=-] employed drowsy cache to exploit program hot-spots and code sequentiality for instruction cache leakage management. Parikh et al. =-=[Li et al. 2004]-=- compared Gated-Vdd and drowsy cache at different L2   </text>
<query_num> 16312 </query_num>
<text>   w long is long enough for each mode? While our paper attempts to address a previously unanswered question, there is a great deal of prior work aimed at reducing leakage power in caches. Azizi et al. [=-=Azizi et al. 2003-=-] introduced asymmetric dual-Vt SRAM cell caches(ACCs). ACCs exploit the fact that in ordinary programs most of the bits in caches are zeros for both the data and instruction streams, and provide sign   </text>
</top>
<top>
<paper_num> 164 </paper_num>
<paper_title>   SHARP: a hybrid adaptive routing protocol for mobile ad hoc networks.  </paper_title>
<paper_abstract>   A central challenge in ad hoc networks is the design of routing protocols that can adapt their behavior to frequent and rapid changes in the network. The performance of proactive and reactive routing protocols varies with network characteristics, and one protocol may outperform the other in different network conditions. The optimal routing strategy depends on the underlying network topology, rate of change, and traffic pattern, and varies dynamically. This paper introduces the Sharp Hybrid Adaptive Routing Protocol (SHARP), which automatically finds the balance point between proactive and reactive routing by adjusting the degree to which route information is propagated proactively versus the degree to which it needs to be discovered reactively. SHARP enables each node to use a different application-specific performance metric to control the adaptation of the routing layer. This paper describes application-specific protocols built on top of SHARP for minimizing packet overhead, bounding loss rate, and controlling jitter. Simulation studies show that the resulting protocols outperform the purely proactive and purely reactive protocols across a wide range of network characteristics.  </paper_abstract>
<query_num> 16401 </query_num>
<text>   . Higher layer protocols such as TCP are quite sensitive to packet loss in the underlying layers. A routing protocol that results in a high loss rate will experience greatly diminished TCP throughput =-=[8, 16]-=-. Other applications such as compressed video also exhibit relatively low tolerance to loss rate. There is a fundamental trade-off in terms of overhead and reliability between the proactive and reacti   </text>
<query_num> 16402 </query_num>
<text>   We also provide the intuition justifying why they work well by examining a non-adaptive version of SHARP with statically selected, fixed values for the zone radius. SHARP was implemented in GloMoSim =-=[27]-=-, a scalable packet-level simulator with an accurate radio model. The many parameters necessary for a realistic simulation were derived from real-world applications. The MAC protocol and radio charact   </text>
<query_num> 16403 </query_num>
<text>   ], and OLSR [9], exchange routing information periodically between hosts, and constantly maintain a set of available routes for all nodes in the network. In contrast, reactive protocols, such as AODV =-=[25]-=-, DSR [11], and TORA [20], delay route discovery until a particular route is required, and propagate routing information only on demand. There are also a few hybrid protocols, such as ZRP [7], HARP [1 t protocol, incurring low loss rate and predictable overhead.4.2 Reactive Routing Component SHARP’s reactive routing protocol is based on the Ad-hoc On-demand Distance Vector (AODV) routing protocol =-=[25]-=-. SHARP simply uses standard AODV [23] for reactive routing, and includes several optimizations such as route caching and expanding ring search. SHARP integrates the proactive and the reactive compone   </text>
<query_num> 16404 </query_num>
<text>   as AODV [25], DSR [11], and TORA [20], delay route discovery until a particular route is required, and propagate routing information only on demand. There are also a few hybrid protocols, such as ZRP =-=[7]-=-, HARP [18], and ZHLS [10], that combine proactive and reactive routing strategies. There is a fundamental trade-off between proactive dissemination and reactive discovery of routing information. Whil  pursue disparate application-specific goals. In this Section, we provide a brief overview of hybrid and adaptive routing protocols and summarize how they differ from SHARP. The Zone Routing Protocol =-=[7]-=- (ZRP) was the first hybrid routing protocol with both a proactive and a reactive routing component. ZRP defines a zone around each node consisting of its k-neighborhood. Routing within a zone is perf   </text>
<query_num> 16405 </query_num>
<text>   ce. FSR [6], Fisheye State Routing, is a link-state protocol that exchanges periodic link-state information. The period of link state propagation is determined by the distance to the destination. ADV =-=[1]-=- is Adaptive Distance Vector algorithm that exhibits on-demand characteristics by varying the frequency and size of routing updates. Some researchers have examined supplanting reactive protocols with   </text>
<query_num> 16406 </query_num>
<text>   en independently with the lifetime distributed exponentially with mean λ, the average frequency at which all the downstream links fail at a node can be es1 2βnλ r SPR D h n 1 i=1 i timated to be (see =-=[24]-=-), where βn equals , and n is the number of downstream links at that node. In addition, an event triggered by link-failures at a node alters the status of several of its upstream links. This results i o within a few hops of h by using the expanding ring search. AODV routes fail at an average rate of h per second under the assumption of indepenλ dent and exponentially distributed link-failures (see =-=[24]-=-). Thus, the route maintenance overhead of AODV, for active routes, is approximately proportional to N S h h . Other facλ tors such as congestion could also cause route breaks. In this analysis, we as   </text>
<query_num> 16407 </query_num>
<text>   or service providers are contacted by several nodes in the network. In many general-purpose applications, including messaging, and wireless audio, the popularity of nodes follows a Zipf-distribution =-=[4]-=-, making some nodes more popular than others. Figure 1 shows the SHARP topology in a typical deployment. SHARP maintains proactive routing zones around popular destinations A, B, and C. It achieves th   </text>
<query_num> 16408 </query_num>
<text>   routing information periodically between hosts, and constantly maintain a set of available routes for all nodes in the network. In contrast, reactive protocols, such as AODV [25], DSR [11], and TORA =-=[20]-=-, delay route discovery until a particular route is required, and propagate routing information only on demand. There are also a few hybrid protocols, such as ZRP [7], HARP [18], and ZHLS [10], that c s an efficient protocol engineered with techniques borrowed from different routing algorithms such as Destination Sequenced Distance Vector (DSDV) [22] and Temporally Ordered Routing Algorithm (TORA) =-=[20]-=-. SPR maintains routes to a single destination in each zone. This enables SPR to employ low overhead mechanisms to build and maintain routes and to offer several good properties, including predictable   </text>
<query_num> 16409 </query_num>
<text>   s all nodes, which is determined offline from a past history of link-failure statistics. Several researchers have examined the behavior of reactive and proactive routing protocols through simulations =-=[2, 5]-=-. Our primary goal in this paper is not to perform protocol comparisons, but to identify and evaluate the trade-off between the two routing regimes. Our measurements show that the choice of the optima   </text>
<query_num> 16410 </query_num>
<text>   s all nodes, which is determined offline from a past history of link-failure statistics. Several researchers have examined the behavior of reactive and proactive routing protocols through simulations =-=[2, 5]-=-. Our primary goal in this paper is not to perform protocol comparisons, but to identify and evaluate the trade-off between the two routing regimes. Our measurements show that the choice of the optima that performing local repairs distorts the optimality of the DAG, since the correct distance to the destination is no longer known. This has been shown to significantly affect the performance of TORA =-=[2]-=-. SHARP compensates for this problem with the construction protocol, which periodically rebuilds the DAG from scratch. Hence, SPR’s DAGs contain relatively short paths to the destination most of the t eam links have failed, link reversal is performed to repair the DAG and the data packet is forwarded to the best newly formed downstream node. Temporary routing loops could be created (as observed in =-=[2]-=-) if the data packets reach the new downstream node ahead of the update packet. To prevent this, the new height of the node is piggybacked with the data packet, so that the downstream node can update   </text>
</top>
<top>
<paper_num> 165 </paper_num>
<paper_title>   A plan-based agent architecture for interpreting natural language dialogue.  </paper_title>
<paper_abstract>   This paper describes a plan-based agent architecture for modeling NL cooperative dialogue; in particular, the paper focuses on the interpretation dialogue and explanation of its coherence by means of the recognition of the speakers&amp;apos; underlying intentions. The approach we propose makes it possible to analyze and explain in a uniform way several apparently unrelated linguistic phenomena, which have been often studied separately and treated via ad-hoc methods in the models of dialogue presented in the literature. Our model of linguistic interaction is based on the idea that dialogue can be seen as any other interaction among agents: therefore, domain-level and linguistic actions are treated in a similar way. Our agent architecture is based on a two-level representation of the knowledge about acting: at the metalevel, the Agent Modeling plans describe the recipes for plan formation and execution (they are a declarative representation of a reactive planner); at the object level, the domain ...  </paper_abstract>
<query_num> 16501 </query_num>
<text>   ed approaches to dialogue modeling have discussed and analyzed the importance of considering also the agents&amp;apos; metalevel activity for modeling their interaction with other agents (=-=Litman &amp; Allen 1987,Ramshaw 1991, Carberry et al. 1992-=-): this metalevel activity is important to understand behaviors like plan exploration. 3 An architecture based on plans In our architecture, the actions which an agent can execut   </text>
<query_num> 16502 </query_num>
<text>   ing. In contrast to the multi-level approach to dialogue processing, Grosz, Sidner and Lochbaum only use domain-level plans to model collaborative dialogue (=-=Grosz &amp; Sidner 1986, Lochbaum et al. 1990, Lochbaum 1991, Lochbaum 1995-=-). We agree with Lochbaum that it is better not to introduce multiple plan types, but: - Although we have three plan libraries, only the object-level and the metalevel plan libraries ha   </text>
<query_num> 16503 </query_num>
<text>   requires some means to choose the best hypothesis. Various methods have been proposed in the literature to cope with this problem. For example, the model described in Carberry (1990a) and the one in (=-=Bauer 1995-=-) are based on the Dempster-Shafer theory of evidence, whose main advantage with respect to classical probability theory is its ability to keep apart uncertainty and ignorance. Although both approache   </text>
<query_num> 16504 </query_num>
<text>   t to the multi-level approach to dialogue processing, Grosz, Sidner and Lochbaum only use domain-level plans to model collaborative dialogue (=-=Grosz &amp; Sidner 1986, Lochbaum et al. 1990, Lochbaum 1991, Lochbaum 1995-=-). We agree with Lochbaum that it is better not to introduce multiple plan types, but: - Although we have three plan libraries, only the object-level and the metalevel plan libraries have functionally   </text>
<query_num> 16505 </query_num>
<text>   t&amp;apos;s beliefs, and the &amp;quot;Interpret&amp;quot; action, that represents the interpretation of the turns of a dialogue. All the libraries are organized on the basis of a Decomposition and a Generalization Hierarchy (=-=Kautz 1991, Weida &amp; Litman 1992-=-). The first hierarchy specifies how a complex action can be performed by executing simpler actions. The Generalization Hierarchy supports feature inheritance among actions: this   </text>
</top>
<top>
<paper_num> 166 </paper_num>
<paper_title>   Locally Weighted Learning for Control.  </paper_title>
<paper_abstract>   Lazy learning methods provide useful representations and training algorithms for learning about complex phenomena during autonomous adaptive control of complex systems. This paper surveys ways in which locally weighted learning, a type of lazy learning, has been applied by ustocontrol tasks. We explain various forms that control tasks can take, and how this a ects the choice of learning paradigm. The discussion section explores the interesting impact that explicitly remembering all previous experiences has on the problem of learning to control.  </paper_abstract>
<query_num> 16601 </query_num>
<text>   , 1995). We will see how the explicit representation of speci c memories can speed up convergence and improve the robustness and autonomy of optimization and control algorithms (=-=Atkeson et al., 1995; Moore and Schneider, 1995-=-). It is frustrating to watch a robot repeat its mistakes, with only a slight improvement on each attempt. The goal of the learning algorithms described here is to improve performance as rapidly as po arning is to optimize a particular criterion, rather than achieve a particular outcome. Lazy learning can be used to represent the cost function directly and to speed the search for maxima or minima (=-=Moore and Schneider, 1995-=-). A linear local model can be used to estimate the rst derivatives (gradient) and a quadratic local model can be used to estimate the second derivatives (Hessian) of the cost function at the current   </text>
<query_num> 16602 </query_num>
<text>   1995). Cheap cross validation makes search for model parameters routine, and we have explored procedures that take advantage of this (=-=Atkeson et al., 1995; Maron and Moore, 1994; Moore et al., 1992; Moore and Lee, 1994-=-). We have extended the locally weighted learning approach to give information about the reliability of the predictions and local linearizations generated, based on the local density and distribution   </text>
<query_num> 16603 </query_num>
<text>   d outcomes is misinterpreted by the inverse model. Even if the inverse model had interpreted the data correctly, any locally weighted averaging on u would have led to incorrect actions (=-=Moore, 1991a; Jordan and Rumelhart, 1992-=-). In subsequent sections on temporally dependent tasks, we will discuss how sometimes the action selected by the inverse function is too aggressive. 2.2 Control Using Forward Models The forward model in a local optimum. In contrast, training nonlinear parametric models can get stuck in local minima. However, some of the control law design algorithms we have surveyed can become stuck (=-=Moore, 1992; Jordan and Rumelhart, 1992-=-). The inverse-model method can become stuck with non-monotonic or highly noisy systems. The shifting setpoint algorithm can become stuck in principle, although this has not yet occurred in practice.   </text>
<query_num> 16604 </query_num>
<text>   iterature on such learning control problems has sprung up in recent years, with the general name of reinforcement learning. Overviews may be found in (=-=Sutton, 1988; Barto et al., 1990; Watkins, 1989; Barto et al., 1995; Moore and Atkeson, 1993-=-). In this paper we will restrict discussion to the applications of lazy learning to these problems. Again, we proceed by learning an empirical forward modelbxk+1 = b f(xk;buk   </text>
<query_num> 16605 </query_num>
<text>   of estimating uncertainty with locally weighted methods is small. Nonlinear parametric representations such asmulti-layer sigmoidal neural networks can also be adapted to return con dence intervals (=-=MacKay, 1992; Pomerleau, 1994-=-), but approximations are required, and the computational cost is larger. Worse, parametric models (e.g., global polynomial regression) that predict con dence statistically are typica   </text>
<query_num> 16606 </query_num>
<text>   point in the database. There are a surprisingly large number of algorithms available for doing this, mostly based on kd-trees (=-=Preparata and Shamos, 1985; Omohundro, 1987; Moore, 1990; Grosse, 1989;Quinlan, 1993; Omohundro, 1991; Deng and Moore, 1995-=-). 26sIs the curse of dimensionality a problem for lazy learning for control? The curse of dimensionality is the exponential dependence of needed resources on di   </text>
<query_num> 16607 </query_num>
<text>   se data points which, given the uncertainty inherent in the prediction, are considered most likely to achieve the desired outcome. This can considerably reduce the exploration required (=-=Moore, 1991a;Cohn et al., 1995-=-). 2.5 ATemporally Independent Task: Billiards In order to explore the e cacy of lazy learning methods for the control of temporally independent tasks, the previously described approaches were impleme ploration heuristics can be used, all based on the idea that it is worth exploring only where there is little con dence in the empirical model (=-=Sutton, 1990; Kaelbling, 1993; Moore and Atkeson, 1993; Cohn et al., 1995-=-). 24s4 Lazy Learning of Models: Pros and Cons Lazy learning of models leads to new forms of autonomous control. The control algorithms explicitly perform empirical nonlinear modeling as well as simul rediction. This can be done heuristically with the local density of the data providing an uncertainty estimate =-=(Moore, 1991a)-=- or by making sensible statistical assumptions (=-=Schaal and Atkeson, 1994b; Cohn et al., 1995-=-). In either case, this has been shown empirically to dramatically reduce the amount of exploration needed when the uncertainty estimates guide the experiment design. The cost of estimating uncertaint   </text>
<query_num> 16608 </query_num>
<text>   se. Less expensive forms of dynamic programming would normally perform value iteration only at the end of each trial (as we do in the example in Section 3.6.1), or as an incremental parallel process (=-=Sutton, 1990; Moore and Atkeson, 1993; Peng and Williams, 1993-=-). 3.6.1 A Simulation Example: The Puck We illustrate this form of learning by means of a simple simulated example. Figure 12 depicts a frictionless p ven that we were working with a xed discretization. All transitions between cells experienced by the system were remembered in a discrete state transition model. A learning algorithm similar to Dyna (=-=Sutton, 1990-=-) was used with full value iteration carried out on the discrete model every time-step. Exploration was achieved by assuming any unvisited state had a future cost of zero. The action, which is onedime n action. To resolve this dilemma, a number of useful exploration heuristics can be used, all based on the idea that it is worth exploring only where there is little con dence in the empirical model (=-=Sutton, 1990; Kaelbling, 1993; Moore and Atkeson, 1993; Cohn et al., 1995-=-). 24s4 Lazy Learning of Models: Pros and Cons Lazy learning of models leads to new forms of autonomous control. The control algorithms exp   </text>
<query_num> 16609 </query_num>
<text>   the reliability of the predictions and local linearizations generated, based on the local density and distribution of the data and an estimate of the local variance (=-=Atkeson et al., 1995; Schaal and Atkeson, 1994 a,b-=-). This allows a robot to monitor its own skill level, protect itself from its ignorance by designing robust policies, and guide its exploratory behavior. Another attractive feature of locally weig ance. 3.1.1 An Implementation of Deadbeat Control: Devil Sticking I Deadbeat control using lazy learning models was explored by implementing it for a juggling task known as devil sticking (=-=Schaal and Atkeson, 1994 a,b-=-). A center stick is batted back and forth between two handsticks. Figure 8 shows a sketch of our devil sticking robot. The juggling robot uses its top two joints to perform planar devil sticking.  ieve some higher level task description. However, the setpoint of the task can be manipulated during learning to improve exploration. This is done by the shifting setpoint algorithm (SSA) (=-=Schaal and Atkeson, 1994a-=-). 16sSSA attempts to decompose the control problem into two separate control tasks on different time scales. At the fast time scale, it acts as a dynamic regulator by trying to keep the controlled s  algorithms, for planning, or for further exploration. 3.5 Dynamic Regulation With An Unspeci ed Setpoint: Devil Sticking III The SSA method was tested on the devil sticking juggling task (=-=Schaal and Atkeson, 1994a,b-=-). In this case it had the following steps. 1. Regardless of the poor juggling quality of the robot (i.e., at most two or three hits per trial), the SSA made the robot repeat these initial actions  retization for computing the value function. Several researchers are investigating methods for reducing the cost of value iteration when a model has been learned (e.g. (=-=Moore, 1991b; Mahadevan, 1992; Atkeson, 1994-=-)). 3.6.2 Exploration The approach wehave described does not explicitly explore. If the learned model contains serious errors, a part of state space that wrongly looks unrewarding will never be visite   </text>
</top>
<top>
<paper_num> 167 </paper_num>
<paper_title>   Acknowledgement.  </paper_title>
<paper_abstract>   I would like to acknowledge the support and valuable advice received from my research advisor Professor Steven Low. Also I would like to acknowledge Lachlan Andrew, Ryan Witt, and Bartek Wydrowski for their help with de-velopment of the many underlying theories, implementation of the protocol in the Linux kernel, and last but not least, the advice they provided. iii iv We present the first implementation of TCP MaxNet, a conges-tion control algorithm which uses multi-bit explicit congestion sig-nal. While the theoretical properties of the MaxNet framework have been studied in depth, we have been lacking an implementation of the protocol that would confirm or refute what theory predicts; namely, stability, quick speed of convergence, max-min fairness, short router queues and low latency. The protocol consists of two components: an AQM algorithm and a source algorithm. The AQM algorithm resides in routers and calculates the congestion price. The source algorithm reacts to the calculated price and adjusts the sending rate accord-ingly. We provide pseudocode of both algorithms, and implement the protocol in the Linux kernel by modifying the TCP/IP stack. Imple-mentation related issues, such as fractional calculations and evaluation of exponentials in the kernel are described. We show that our imple-mentation, which uses a 23-bit TCP Option to encode and transmit the price, scales from 32 bits/sec to 1 Peta-bit/sec rates with sufficient accuracy. In the initial phase of data transmissions, we use a variant of the QuickStart algorithm, which allows the sender to start send-ing almost immediately at the available capacity, and prevents queue buildup. Using state of the art WAN-in-Lab equipment, we evaluate the performance of TCP MaxNet and confirm favorable properties of the protocol. We observe that within a few RTTs the sending rates converge to their max-min fair shares as dictated by the price. We show that the latency is much lower than that of the other TCP pro-tocols, and the protocol operates with nearly empty router queues. Finally, since the protocol does not treat loss as congestion signal, it has superior performance under loss.  </paper_abstract>
<query_num> 16701 </query_num>
<text>   (t) is the amount of traffic which has been enqueued since the last price update and is destined to leave through link l, 1 is a constant controlling C the rate of convergence of the MaxNet algorithm =-=[27]-=-, and µ is a constant. In steady state the price stabilizes and yl(t) = µC. Choosing µ &amp;lt; 1 results 18s1. On SYN arrival: QSrate ← C(µ + ɛ) − y dt if QSrate &amp;lt; QSminrate then QSrate ← QSminrate end if �   </text>
<query_num> 16702 </query_num>
<text>   . ECN [16], an extension of TCP Reno, marks a single bit in the header to indicate congestion. The higher the rate of packet marking, the higher the congestion level is. Even more interesting are XCP =-=[17]-=- and RCP [18], new protocols that use multi-bit explicit congestion signaling. XCP uses the routers to communicate the sender’s window in the packet headers, while RCP communicates directly the sendin   </text>
<query_num> 16703 </query_num>
<text>   P BIC achieves higher average throughput in high delay bandwidth product networks by increasing the sender’s window more aggressively 2 after loss, but this results in longer router queues. TCP Vegas =-=[12]-=-, and TCP FAST [13] [14] use a new approach and treat delay as an indicator of congestion. The biggest advantage 2 TCP BIC halves its window after loss and then in each subsequent RTT increases its wi   </text>
<query_num> 16704 </query_num>
<text>   aggressively push data through the network, and disobey the price signal. We suggest separating the MaxNet queue in the routers, and using a variant of selective packet dropping method, such as CHOKE =-=[32]-=-, to enforce approximate fairness. Separating the MaxNet queue from the standard queue seems to be advantageous for the following two reasons. First, the separate queue allows MaxNet to enjoy low late nts has not received much attention in connection with other recent TCP implementations, we believe that this challenge can be overcome by using the approximate fair sharing schemes presented e.g. in =-=[32]-=-, [33] and [34]. Although our implementation could theoretically scale up to 1 Petabit/sec, increasing CPU utilization will become a performance bottleneck at higher speeds. In order to save CPU cycle   </text>
<query_num> 16705 </query_num>
<text>   arkably well considering the explosive growth of the Internet in the last two decades. However, it is well known that the performance of the protocol degrades as the bandwidth delay product increases =-=[4]-=- [5] [6]. Equation (2) implies that in ”equilibrium” the window of the sender oscillates between the maximal value it attains and one half of that value. In order to achieve full utilization, Reno req be expressed as ˙v = 1 T − 2 3 qw2 , (4) T and in equilibrium we have ˙v = 0. We have the following relation between the equilibrium loss probability q ∗ and the equilibrium window w ∗ , described in =-=[4]-=- [5] [6]: q ∗ = 3 1 . (5) 2 w∗2 Consider, for example, 1 Gbps link with 50 ms RTT, a link we use in our experiments in chapter 5, and assuming the packet size is 1500 bytes, we calculate that the aver   </text>
<query_num> 16706 </query_num>
<text>   bly well considering the explosive growth of the Internet in the last two decades. However, it is well known that the performance of the protocol degrades as the bandwidth delay product increases [4] =-=[5]-=- [6]. Equation (2) implies that in ”equilibrium” the window of the sender oscillates between the maximal value it attains and one half of that value. In order to achieve full utilization, Reno require xpressed as ˙v = 1 T − 2 3 qw2 , (4) T and in equilibrium we have ˙v = 0. We have the following relation between the equilibrium loss probability q ∗ and the equilibrium window w ∗ , described in [4] =-=[5]-=- [6]: q ∗ = 3 1 . (5) 2 w∗2 Consider, for example, 1 Gbps link with 50 ms RTT, a link we use in our experiments in chapter 5, and assuming the packet size is 1500 bytes, we calculate that the average   </text>
<query_num> 16707 </query_num>
<text>   cing 14 ms of one-way propagation delay. Our experimental setup consisting of two Linux routers and four Linux end hosts is depicted in Fig. 9. Figure 9: The topology of the Wan-in-Lab. 34sA dummynet =-=[31]-=- machine resides on the link between Goldenrod and Yellow. This allows to introduce propagation delays between Goldenrod and Yellow in both directions, simulating a real link. TCP MaxNet router code i   </text>
<query_num> 16708 </query_num>
<text>   eived much attention in connection with other recent TCP implementations, we believe that this challenge can be overcome by using the approximate fair sharing schemes presented e.g. in [32], [33] and =-=[34]-=-. Although our implementation could theoretically scale up to 1 Petabit/sec, increasing CPU utilization will become a performance bottleneck at higher speeds. In order to save CPU cycles, we suggest r   </text>
<query_num> 16709 </query_num>
<text>   er average throughput in high delay bandwidth product networks by increasing the sender’s window more aggressively 2 after loss, but this results in longer router queues. TCP Vegas [12], and TCP FAST =-=[13]-=- [14] use a new approach and treat delay as an indicator of congestion. The biggest advantage 2 TCP BIC halves its window after loss and then in each subsequent RTT increases its window by half of the   </text>
<query_num> 16710 </query_num>
<text>   er, recent research showed that even these implementations are not flawless. XCP was shown to only achieve constrained max-min fairness, and one of its flows may be assigned an arbitrarily small rate =-=[19]-=-. Moreover, the protocol’s stability has only been shown for a single link scenario shared by flows with identical RTTs. RCP faces another problem. It calculates d0 [18], an estimate of the weighted a   </text>
<query_num> 16711 </query_num>
<text>   erage throughput in high delay bandwidth product networks by increasing the sender’s window more aggressively 2 after loss, but this results in longer router queues. TCP Vegas [12], and TCP FAST [13] =-=[14]-=- use a new approach and treat delay as an indicator of congestion. The biggest advantage 2 TCP BIC halves its window after loss and then in each subsequent RTT increases its window by half of the diff   </text>
<query_num> 16712 </query_num>
<text>   increase of a transmission rate of a sender does not result in the decrease of the transmission rate of a competing user that has equal or lower rate. Moreover, MaxNet has fast convergence properties =-=[23]-=-. Since the MaxNet sender decreases its sending rate as the price calculated by routers increases, proper selection of the rate control algorithms allows operation of the protocol with empty router qu   </text>
<query_num> 16713 </query_num>
<text>   making it nearly impossible to sustain full utilization of the network. A number of extensions of the original packet loss based scheme, such as TCP BIC [7], HS [8], H [9], Scalable [10] and Westwood =-=[11]-=-, have been suggested. These protocols often dramatically improve the performance of TCP Reno. However, some fundamental problems remain. Oscillations are a natural consequence of only one-bit congest   </text>
<query_num> 16714 </query_num>
<text>   oss and then in each subsequent RTT increases its window by half of the difference of the current window and the window before loss. 11sof these delay-based proposals is that they are provably stable =-=[15]-=- [6] [14] and achieve excellent throughput and responsiveness. However, due to their delay based design, both protocols still require nonzero router queues, and thus introduce additional delay that gr   </text>
<query_num> 16715 </query_num>
<text>   s not received much attention in connection with other recent TCP implementations, we believe that this challenge can be overcome by using the approximate fair sharing schemes presented e.g. in [32], =-=[33]-=- and [34]. Although our implementation could theoretically scale up to 1 Petabit/sec, increasing CPU utilization will become a performance bottleneck at higher speeds. In order to save CPU cycles, we   </text>
<query_num> 16716 </query_num>
<text>   stion based loss, making it nearly impossible to sustain full utilization of the network. A number of extensions of the original packet loss based scheme, such as TCP BIC [7], HS [8], H [9], Scalable =-=[10]-=- and Westwood [11], have been suggested. These protocols often dramatically improve the performance of TCP Reno. However, some fundamental problems remain. Oscillations are a natural consequence of on   </text>
<query_num> 16717 </query_num>
<text>   tocols using passive, active, binary and multi-bit types of congestion level signaling has been suggested, analyzed and implemented. The most widely used congestion control protocol today is TCP Reno =-=[2]-=- and its improved variant New Reno [3]. The algorithm is an improved version of TCP Tahoe, which dates to 1988 when it was suggested as a response to a series of congestion collapses in the Internet.  er example of a TCP option is SACK [25], an important and widely adopted extension of TCP Reno. SACK, an extension of the TCP algorithm allowing selective acknowledgement of received segments of data =-=[2]-=-, has been introduced because of poor performance of the standard algorithm in the presence of multiple packet losses. Fig. 2 shows a typical receiver queue when multiple packets were lost. The origin   </text>
<query_num> 16718 </query_num>
<text>   well considering the explosive growth of the Internet in the last two decades. However, it is well known that the performance of the protocol degrades as the bandwidth delay product increases [4] [5] =-=[6]-=-. Equation (2) implies that in ”equilibrium” the window of the sender oscillates between the maximal value it attains and one half of that value. In order to achieve full utilization, Reno requires bu ssed as ˙v = 1 T − 2 3 qw2 , (4) T and in equilibrium we have ˙v = 0. We have the following relation between the equilibrium loss probability q ∗ and the equilibrium window w ∗ , described in [4] [5] =-=[6]-=-: q ∗ = 3 1 . (5) 2 w∗2 Consider, for example, 1 Gbps link with 50 ms RTT, a link we use in our experiments in chapter 5, and assuming the packet size is 1500 bytes, we calculate that the average send nd then in each subsequent RTT increases its window by half of the difference of the current window and the window before loss. 11sof these delay-based proposals is that they are provably stable [15] =-=[6]-=- [14] and achieve excellent throughput and responsiveness. However, due to their delay based design, both protocols still require nonzero router queues, and thus introduce additional delay that grows   </text>
</top>
<top>
<paper_num> 168 </paper_num>
<paper_title>   Bilateral and Multilateral Exchanges for Peer-Assisted Content Distribution.  </paper_title>
<paper_abstract>   Abstract—Users of the BitTorrent file-sharing protocol and its variants are incentivized to contribute their upload capacity in a bilateral manner: Downloading is possible in return for uploading to the same user. An alternative is to use multilateral exchange to match user demand for content to available supply at other users in the system. We provide a formal comparison of peer-to-peer system designs based on bilateral exchange with those that enable multilateral exchange via a price-based market mechanism to match supply and demand. First, we compare the two types of exchange in terms of the equilibria that arise. A multilateral equilibrium allocation is Pareto-efficient, while we demonstrate that bilateral equilibrium allocations are not Pareto-efficient in general. We show that Pareto efficiency represents the “gap” between bilateral and multilateral equilibria: A bilateral equilibrium allocation corresponds to a multilateral equilibrium allocation if and only if it is Pareto-efficient. Our proof exploits the fact that Pareto efficiency implies reversibility of an appropriately constructed Markov chain. Second, we compare the two types of exchange through the expected percentage of users that can trade in a large system, assuming a fixed file popularity distribution. Our theoretical results as well as analysis of a BitTorrent dataset provide quantitative insight into regimes where bilateral exchange may perform quite well even though it does not always give rise to Pareto-efficient equilibrium allocations. Index Terms—Asymptotic analysis, market equillibria, Markov processes, peer-to-peer systems, random graphs.  </paper_abstract>
<query_num> 16801 </query_num>
<text>   13]. The transactions role of money is surveyed in [40]. Finally, as discussed in the Introduction, we note that a number of studies consider incentives in peer-to-peer systems (e.g., [3], [9], [11], =-=[14]-=-, [15], [24], [33], [38], [40]). Our paper contributes to this broad line of literature. VII. CONCLUSION This paper provides a formal comparison of two peer-to-peer system designs: bilateral barter sy   </text>
<query_num> 16802 </query_num>
<text>   14 This result is analogous to the same result for undirected Erdös–Rényi random graphs and can be proven using similar counting arguments for threshold behavior of those graphs; see, e.g., [19] and =-=[20]-=-.) Using Markov’s inequality as . In other words, if we sample items with replacement from a bin of items as described above, then we obtain distinct items with probability approaching 1 as . It follo   </text>
<query_num> 16803 </query_num>
<text>   CTION E ARLY peer-to-peer systems did not provide any incentives for participation, leading to extensive free riding: Many peers were using the resources of other peers without contributing their own =-=[2]-=-, [18]. The peer-to-peer community responded with mechanisms to prevent free riding by incentivizing sharing on a bilateral exchange basis, as used by BitTorrent [11] and its variants [32], [39], [35]   </text>
<query_num> 16804 </query_num>
<text>   E ARLY peer-to-peer systems did not provide any incentives for participation, leading to extensive free riding: Many peers were using the resources of other peers without contributing their own [2], =-=[18]-=-. The peer-to-peer community responded with mechanisms to prevent free riding by incentivizing sharing on a bilateral exchange basis, as used by BitTorrent [11] and its variants [32], [39], [35]. Acco   </text>
<query_num> 16805 </query_num>
<text>   [2], [18]. The peer-to-peer community responded with mechanisms to prevent free riding by incentivizing sharing on a bilateral exchange basis, as used by BitTorrent [11] and its variants [32], [39], =-=[35]-=-. According to the BitTorrent protocol, each user splits his available upload rate among users from which he gets the highest download rates. As a result, an increase in the upload rate to one user ma   </text>
<query_num> 16806 </query_num>
<text>   eases when downloading and increases when uploading. Monetary incentives with a virtual currency have been previously proposed to encourage contribution in peer-to-peer systems [17], [38], [36], [6], =-=[5]-=-. However, such designs are usually more complex than bilateral protocols and are not widespread. The two system designs present a significant tradeoff: Bilateral exchange without money is simple, whi urthermore, for the purposes of this paper, we also assume that there are no constraints in the middle of the network, though our prior work suggests a natural approach for including such constraints =-=[5]-=-. Let be the set of feasible rate vectors. In particular, this ensures that: 1) all rates are nonnegative; 2) users only upload files they possess; and 3) each user does not violate his upload capacit rice according to the mismatch between requests received and available capacity. Simulations suggest that such price dynamics converge for a variety of topologies in the case of multilateral exchange =-=[5]-=-. The same approach can be applied to BE by considering exchange ratios instead. IV. EFFICIENCY OF EQUILIBRIA This section rigorously analyzes the efficiency properties of bilateral and multilateral e nge in a single swarm to chunks of the same file. 12 We conclude by noting that our paper has considered a static “snapshot” view of a file-sharing system. This is complemented by our earlier work in =-=[5]-=-, where we considered a system design for multilateral exchange in a dynamic setting. In a dynamic system, the number of simultaneous matches possible may be quite small, even if all content possessed xible version of mixed bundling, adapted to allow trade across files in a bundle.APERJIS et al.: BILATERAL AND MULTILATERAL EXCHANGES FOR PEER-ASSISTED CONTENT DISTRIBUTION 1299 later. Our design in =-=[5]-=- leverages this advantage by designing a system where pricing mechanisms serve only as algorithmic devices to ensure efficient exchange; to keep the system simple, we never expose prices directly to t   </text>
<query_num> 16807 </query_num>
<text>   file, i.e., for all users . There are 7323 files in the system, and the number of users is shown on the horizontal axis. exponent is greater than one, bilateral exchange performs reasonably well. In =-=[7]-=-, we show that if , then the conclusions of Proposition 6 hold even if users possess and desire multiple files. In fact, the proof of Proposition 7 in [7] shows that when each user possesses files and n . The remainder of the proof shows that there exists an invariant distribution such that is optimal for the Multilateral Peer Optimization problem under . Due to space limitations it is provided in =-=[7]-=-. Proof of Proposition 5: We first show the result when for all users. Let Since is finite and , it suffices to show that as . We observe that for any fixed The first term approaches zero as ; the sec   </text>
<query_num> 16808 </query_num>
<text>   he transactions role of money is surveyed in [40]. Finally, as discussed in the Introduction, we note that a number of studies consider incentives in peer-to-peer systems (e.g., [3], [9], [11], [14], =-=[15]-=-, [24], [33], [38], [40]). Our paper contributes to this broad line of literature. VII. CONCLUSION This paper provides a formal comparison of two peer-to-peer system designs: bilateral barter systems   </text>
<query_num> 16809 </query_num>
<text>   ir own [2], [18]. The peer-to-peer community responded with mechanisms to prevent free riding by incentivizing sharing on a bilateral exchange basis, as used by BitTorrent [11] and its variants [32], =-=[39]-=-, [35]. According to the BitTorrent protocol, each user splits his available upload rate among users from which he gets the highest download rates. As a result, an increase in the upload rate to one u   </text>
<query_num> 16810 </query_num>
<text>   ne price per user, and users optimize with 1 Note that we allow users to bilaterally exchange content over multiple files. This is partly supported by swarming systems like BitTorrent through bundles =-=[28]-=-. BitTorrent also has a mixed bundling mechanism, which could be adapted to allow trade across files in a bundle (but currently is not flexible).s1294 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 19, NO. lative to BitTorrent [33]. Liu et al. study a similar mechanism assuming that peers belong to an underlying social network [25]. Finally, the performance implications of bundling have been considered =-=[28]-=-. Our work is also related to the study of equilibria in economies where not all trades are allowed. Kakade et al. introduce a graph-theoretic generalization of classical Arrow–Debreu economics, in wh   </text>
<query_num> 16811 </query_num>
<text>   nsactions role of money is surveyed in [40]. Finally, as discussed in the Introduction, we note that a number of studies consider incentives in peer-to-peer systems (e.g., [3], [9], [11], [14], [15], =-=[24]-=-, [33], [38], [40]). Our paper contributes to this broad line of literature. VII. CONCLUSION This paper provides a formal comparison of two peer-to-peer system designs: bilateral barter systems such a   </text>
<query_num> 16812 </query_num>
<text>   nsidered [29], [13]. The transactions role of money is surveyed in [40]. Finally, as discussed in the Introduction, we note that a number of studies consider incentives in peer-to-peer systems (e.g., =-=[3]-=-, [9], [11], [14], [15], [24], [33], [38], [40]). Our paper contributes to this broad line of literature. VII. CONCLUSION This paper provides a formal comparison of two peer-to-peer system designs: bi   </text>
<query_num> 16813 </query_num>
<text>   ptimal solution of the Multilateral Peer Optimization problem of each user (Step 3). We show each of these steps by demonstrating that if the desired conclusion of the step does 13 See recent work in =-=[34]-=- for a similar approach in peer-to-peer distributed backup. not hold, then there exists a rate vector that Pareto-improves —violating the assumption that is Pareto-efficient. Before beginning the proo   </text>
<query_num> 16814 </query_num>
<text>   r that decreases when downloading and increases when uploading. Monetary incentives with a virtual currency have been previously proposed to encourage contribution in peer-to-peer systems [17], [38], =-=[36]-=-, [6], [5]. However, such designs are usually more complex than bilateral protocols and are not widespread. The two system designs present a significant tradeoff: Bilateral exchange without money is s   </text>
<query_num> 16815 </query_num>
<text>   rades to pass through one intermediary improves performance and incentives relative to BitTorrent [33]. Liu et al. study a similar mechanism assuming that peers belong to an underlying social network =-=[25]-=-. Finally, the performance implications of bundling have been considered [28]. Our work is also related to the study of equilibria in economies where not all trades are allowed. Kakade et al. introduc   </text>
<query_num> 16816 </query_num>
<text>   red [29], [13]. The transactions role of money is surveyed in [40]. Finally, as discussed in the Introduction, we note that a number of studies consider incentives in peer-to-peer systems (e.g., [3], =-=[9]-=-, [11], [14], [15], [24], [33], [38], [40]). Our paper contributes to this broad line of literature. VII. CONCLUSION This paper provides a formal comparison of two peer-to-peer system designs: bilater   </text>
<query_num> 16817 </query_num>
<text>   that cannot trade multilaterally, we have . The following propositions imply that in a large system, 5 Gummadi et al. find that peer-to-peer file popularity follows a flattened Zipflink distribution =-=[16]-=-. However, the Zipf distribution is still the closest approximation for which analytical work is possible.1296 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 19, NO. 5, OCTOBER 2011 may be significant whe   </text>
<query_num> 16818 </query_num>
<text>   this literature. The superiority of monetary exchange has been studied [37], and dynamics of bilateral trading processes have been considered [29], [13]. The transactions role of money is surveyed in =-=[40]-=-. Finally, as discussed in the Introduction, we note that a number of studies consider incentives in peer-to-peer systems (e.g., [3], [9], [11], [14], [15], [24], [33], [38], [40]). Our paper contribu   </text>
<query_num> 16819 </query_num>
<text>   to each user that decreases when downloading and increases when uploading. Monetary incentives with a virtual currency have been previously proposed to encourage contribution in peer-to-peer systems =-=[17]-=-, [38], [36], [6], [5]. However, such designs are usually more complex than bilateral protocols and are not widespread. The two system designs present a significant tradeoff: Bilateral exchange withou   </text>
<query_num> 16820 </query_num>
<text>   tric with 14 This result is analogous to the same result for undirected Erdös–Rényi random graphs and can be proven using similar counting arguments for threshold behavior of those graphs; see, e.g., =-=[19]-=- and [20].) Using Markov’s inequality as . In other words, if we sample items with replacement from a bin of items as described above, then we obtain distinct items with probability approaching 1 as .   </text>
</top>
<top>
<paper_num> 169 </paper_num>
<paper_title>   Separation Logic Semantics for Communicating Processes.  </paper_title>
<paper_abstract>   This paper explores a unification of the ideas of Concurrent Separation Logic with those of Communicating Sequential Processes. It extends separation logic by an operator for separation in time as well as separation in space. It extends CSP in the direction of the pi-calculus: dynamic change of alphabet is achieved by communication of channel names. Separation is exploited to ensure that each channel still has only two ends. For purposes of exploration, the model is the simplest possible, confined to traces without refusals. The treatment is sufficiently general to facilitate extensions by standard techniques for sharing multiplexed channels and heap state. 1  </paper_abstract>
<query_num> 16901 </query_num>
<text>   at most one sender and receiver [12,25]. The type systems have been used to characterize special classes of behaviour, such as sequential behaviour [1], and are now being used in work on web services =-=[7]-=-. Our use of partial alphabet composition is similar to how substructural typing limits the number of processes that can access a channel. Beyond that basic similarity, the techniques developed here a   </text>
<query_num> 16902 </query_num>
<text>   bitrary partial commutative monoids in place of alphabets, ⊑ might have infinite descending sequences, in which case Proposition 6.4 would fail. Fractional permissions [2] provide one such model. See =-=[22]-=- for further information on the theory of footprints (where the term “footprint” is used in a related, but inequivalent, way). 18Hoare and O’Hearn 7 Related Work In the Introduction we acknowledged t   </text>
<query_num> 16903 </query_num>
<text>   channels and heap state. 1 Introduction This paper reports on work bringing together semantic ideas lying behind Concurrent Separation Logic (CSL, [18,4]) and Communicating Sequential Processes (CSP, =-=[11]-=-). CSL provides a modular way of reasoning about shared-memory programs. It is based on the principle that that, at any time, it is possible to partition the state into that “owned” by separate proces   </text>
<query_num> 16904 </query_num>
<text>   hese kinds of features. In this paper we do not consider structures such as failures, divergences, or infinite traces, that have been used in the semantics of CSP to account for deadlock and liveness =-=[23,3]-=-. The extension to the these sorts of properties is a problem for future work. The presentation that follows is designed to suit readers with some prior acquaintance either of process algebra or of se   </text>
<query_num> 16905 </query_num>
<text>   ide denotational models of session typing systems, where the changing alphabet is an explicit part of the semantics. A number of authors have used substructural logics to reason about process calculi =-=[9,8,20]-=-. The approach has been to first set down an operational semantics of a process calculus, and then use the parallel composition to define a separating conjunction connective as described in Section 2. els, and then use the induced separation connectives in the description of the denotational semantics of processes. So, we use separation conjunctions to provide the semantics of process terms, where =-=[9,8,20]-=- use process terms to provide the semantics of separating conjunctions (in the generalized sense of Section 2). Although our approach is inverse, we share a long-term aim with these works: We would li   </text>
<query_num> 16906 </query_num>
<text>   larger portions of resource: as a consequence, we do not have to think about the entire global state of a system to understand a computation. This is related to the intuition behind the frame rule of =-=[17]-=-. Technical Note. When you “extend with a separate alphabet” you must be sure to satisfy the Synchronization Property. For example, {c!}[c!3]{c!} ∗ {c?}[ ]{c?} is undefined, because the trace {c!, c?} ootprints Above, we remarked how the A∗skip part of condition Unity is related to the frame rule of Separation Logic. We can take this locality idea one step further, by revisiting the footprint idea =-=[17]-=-. Say that the footprint of a program execution is the set of resources touched by it, including those that are allocated. Extrapolating from this, we can talk about a footprint execution or trace, wh   </text>
<query_num> 16907 </query_num>
<text>   of closely related work on substructural typing and logic. There has been a significant body of work on type disciplines that capture constraints on channel usage in pi-calculus. Linear type systems =-=[13]-=- ensure point-to-point channels and more: that each channel is used at most once. Session types are closer to the approach here, in that they ensure point-to-point channels but allow a channel to be r   </text>
<query_num> 16908 </query_num>
<text>   posed by Routley and Meyer in their ternary-relation semantics of substructural logics [24]. Special cases, where the relation can be replaced by a (partial) function, have been used in Bunched Logic =-=[19,21]-=- and Duration Calculus [10]. On the other hand, such conjunctions can be seen as binary modal operators [5]. Like other modal operators, they are definable in predicate calculus by restricted forms of   </text>
<query_num> 16909 </query_num>
<text>   restriction that we never have {P (s), V (s)} ⊆ E, two operations on a single semaphore in a single event set. The event composition ⊎ is union of disjoint sets. This model is similar to the one from =-=[6]-=-, but allows operations on different semaphores to happen at the same time. A variation on this model, which allows concurrent read access to a location, is obtained using permission models [2]. We ca   </text>
<query_num> 16910 </query_num>
<text>   sion types are closer to the approach here, in that they ensure point-to-point channels but allow a channel to be reused multiple times while maintaining that there is at most one sender and receiver =-=[12,25]-=-. The type systems have been used to characterize special classes of behaviour, such as sequential behaviour [1], and are now being used in work on web services [7]. Our use of partial alphabet compos   </text>
<query_num> 16911 </query_num>
<text>   sions by standard techniques for sharing multiplexed channels and heap state. 1 Introduction This paper reports on work bringing together semantic ideas lying behind Concurrent Separation Logic (CSL, =-=[18,4]-=-) and Communicating Sequential Processes (CSP, [11]). CSL provides a modular way of reasoning about shared-memory programs. It is based on the principle that that, at any time, it is possible to parti   </text>
<query_num> 16912 </query_num>
<text>   used multiple times while maintaining that there is at most one sender and receiver [12,25]. The type systems have been used to characterize special classes of behaviour, such as sequential behaviour =-=[1]-=-, and are now being used in work on web services [7]. Our use of partial alphabet composition is similar to how substructural typing limits the number of processes that can access a channel. Beyond th   </text>
</top>
<top>
<paper_num> 170 </paper_num>
<paper_title>   Building Secure High-Performance Web Services with OKWS.  </paper_title>
<paper_abstract>   OKWS is a toolkit for building fast and secure Web services. It provides Web developers with a small set of tools that has proved powerful enough to build complex systems with limited effort. Despite its emphasis on security, OKWS shows performance improvements compared to popular systems: when servicing fully dynamic, non-disk-bound database workloads, OKWS’s throughput and responsiveness exceed that of Apache 2 [3], Flash [23] and Haboob [44]. Experience with OKWS in a commercial deployment suggests it can reduce hardware and system management costs, while providing security guarantees absent in current systems. 1  </paper_abstract>
<query_num> 17001 </query_num>
<text>   . The Exokernel operating system [16] allows its Cheetah Web server to directly access the TCP/IP stack, in order to reduce buffer copies allow for more effective caching. The Denali isolation kernel =-=[45]-=- can isolate Web services by running them on separate virtual machines. 9 Summary and Future Work OKWS is a toolkit for serving dynamic Web content, and its architecture fits naturally into a compelli   </text>
<query_num> 17002 </query_num>
<text>   WS shows performance improvements compared to popular systems: when servicing fully dynamic, non-disk-bound database workloads, OKWS’s throughput and responsiveness exceed that of Apache 2 [3], Flash =-=[23]-=- and Haboob [44]. Experience with OKWS in a commercial deployment suggests it can reduce hardware and system management costs, while providing security guarantees absent in current systems. 1 Introduc roduce OKWS, the OK Web Server. Unlike typical Web servers, OKWS is specialized for dynamic content and is not well-suited to serving files from disk. It relies on existing Web servers, such as Flash =-=[23]-=- or Apache [3], to serve images and other static content. We argue (in Section 5.4) that this separation of static and dynamic content is natural and, moreover, contributes to security. What OKWS does  site-specific services shown in Figure 1 communicate among themselves with SFS’s implementation of Sun RPC [32]; they communicate with external Web clients via HTTP. Unlike other event-based servers =-=[23, 44, 47]-=-, OKWS exposes the event architecture to Web developers. To use OKWS, an administrator installs the helper binaries (okld, okd, pubd and oklogd) to a standard directory such as /usr/local/sbin, and in us Database Proxies OKWS provides Web developers with a generic library for translating between asynchronous RPC and any given blocking client library, in a manner similar to Flash’s helper processes =-=[23]-=-, and “manual calling automatic” in [1]. OKWS users can thus simply implement database proxies: asynchronous RPC front-ends to standard databases, such as MySQL [21] or Berkeley DB [29]. Our libraries ce 490 895 1,590 2,401 10 Services 225 760 1,232 2,089 Change −54.0% −15.1% −22.5% −13.0% Table 2: Average throughputs in connections per second Highly-optimized event-based Web servers such as Flash =-=[23]-=- and Zeus [47] have eclipsed Apache in terms of performance. While Flash in particular has a history of outstanding performance serving static content, our performance studies here indicate that its a   </text>
<query_num> 17003 </query_num>
<text>   ance improvements compared to popular systems: when servicing fully dynamic, non-disk-bound database workloads, OKWS’s throughput and responsiveness exceed that of Apache 2 [3], Flash [23] and Haboob =-=[44]-=-. Experience with OKWS in a commercial deployment suggests it can reduce hardware and system management costs, while providing security guarantees absent in current systems. 1 Introduction Most dynami pendency. The size of the process pool P is determined by the number of concurrent active HTTP sessions; each process pi serves only one of these connections. Javabased systems like the Haboob Server =-=[44]-=- employ only one process; thus P = {p1}, and dependencies (p1, sj) exist for all j. Figures 2(a)-(c) depict graphs of Apache, Flash and Haboob hosting two services for two remote users. Assuming that   site-specific services shown in Figure 1 communicate among themselves with SFS’s implementation of Sun RPC [32]; they communicate with external Web clients via HTTP. Unlike other event-based servers =-=[23, 44, 47]-=-, OKWS exposes the event architecture to Web developers. To use OKWS, an administrator installs the helper binaries (okld, okd, pubd and oklogd) to a standard directory such as /usr/local/sbin, and in   </text>
<query_num> 17004 </query_num>
<text>   eover, an OS-level jail ought to hide all setuid executables from the Web server, since many privilege escalation attacks require such files (examples include the ptrace and bind attacks mentioned in =-=[17]-=-). Privilege escalation is possible without setuid executables but requires OS-level bugs or race conditions that are typically rarer. An adversary can still do damage without control of the Web serve   </text>
<query_num> 17005 </query_num>
<text>   es events but the same results are possible with an appropriate threads library. An expansive body of literature argues the merits of one scheme over the other, and most recently, Capriccio’s authors =-=[34]-=- argue that threads can achieve the same performance as events in the context of Web servers, while providing programmers with a more intuitive interface. Other recent work suggests that threads and e   </text>
<query_num> 17006 </query_num>
<text>   performance advantages of compiled C++ code over Java systems. Other work has proposed changes to underlying operating systems to make Web servers fast and more secure. The Exokernel operating system =-=[16]-=- allows its Cheetah Web server to directly access the TCP/IP stack, in order to reduce buffer copies allow for more effective caching. The Denali isolation kernel [45] can isolate Web services by runn   </text>
<query_num> 17007 </query_num>
<text>   s as Flash but with a process pool whose size is independent of the number of concurrent HTTP connections. 4 Implementation OKWS is a portable, event-based system, written in C++ with the SFS toolkit =-=[18]-=-. It has been successfully tested on Linux and FreeBSD. In OKWS, the different helper processes and site-specific services shown in Figure 1 communicate among themselves with SFS’s implementation of S   </text>
<query_num> 17008 </query_num>
<text>   s to security. What OKWS does provide is a simple, powerful, and secure toolkit for building dynamic content pages (also known as Web services). OKWS enforces the natural principle of least privilege =-=[27]-=- so that those aspects of the system most vulnerable to attack are the least useful to attackers. Further, OKWS separates privileges so that the different components of the system distrust each other.   </text>
<query_num> 17009 </query_num>
<text>   stigated here, many other Web development environments are in widespread use. Zope [48], a Python-based platform, has gained popularity due to its modularity and support for remote collaboration. CSE =-=[13]-=- allows developers to write Web services in C++ and uses some of the same sandboxing schemes we use here to achieve fault isolation. In more commercial settings, Java-based systems often favor thin We   </text>
<query_num> 17010 </query_num>
<text>   such as Ninja [33] build on SEDA’s infrastructure to create clusters of Web servers with the same appealing properties. Other work has used the SFS toolkit to build static Web Servers and Web proxies =-=[46]-=-. Though the current OKWS architecture is well-suited for SMP machines, the adoption of libasync-mp would allow for finer-grained sharing of a Web workload across many CPUs. OKWS uses events but the s   </text>
<query_num> 17011 </query_num>
<text>   velopers with a generic library for translating between asynchronous RPC and any given blocking client library, in a manner similar to Flash’s helper processes [23], and “manual calling automatic” in =-=[1]-=-. OKWS users can thus simply implement database proxies: asynchronous RPC front-ends to standard databases, such as MySQL [21] or Berkeley DB [29]. Our libraries provide the illusion of a standard asy l_select_db(&amp;quot;testdb&amp;quot;, $db); $id = $HTTP_GET_VARS[&amp;quot;id&amp;quot;]; $qry = &amp;quot;SELECT x,y FROM tab WHERE x=$id&amp;quot;; $result = mysql_query(&amp;quot;$qry&amp;quot;, $db); $myrow = mysql_fetch_row($result); print(&amp;quot;QRY $id $myrow[0] $myrow=-=[1]-=-\n&amp;quot;); ?&amp;gt; &amp;lt;/body&amp;gt; &amp;lt;/html&amp;gt; Figure 5: PHP version of the null service should use the provided “safe” strings classes when generating HTML output, and they should use only autogenerated RPC stubs for netw s can achieve the same performance as events in the context of Web servers, while providing programmers with a more intuitive interface. Other recent work suggests that threads and events can coexist =-=[1]-=-. Such techniques, if applied to OKWS, would simplify stack management for Web developers. In addition to the PHP [25] scripting language investigated here, many other Web development environments are   </text>
</top>
<top>
<paper_num> 171 </paper_num>
<paper_title>   Building a reusable test collection for question answering.  </paper_title>
<paper_abstract>   In contrast to traditional information retrieval systems, which return ranked lists of documents that users must manually browse through, a question answering system attempts to directly answer natural language questions posed by the user. Although such systems possess language processing capabilities, they still rely on traditional document retrieval techniques to generate an initial candidate set of documents. In this paper, we argue that document retrieval for question answering represents a different task than retrieving documents in response to more general retrospective information needs. Thus, to guide future system development, specialized question answering test collections must be constructed. We have shown that the current evaluation resources have major shortcomings, and to remedy the situation, we have manually created a small, reusable question answering test collection for research purposes. This article describes our methodology for building this test collection and discusses issues we encountered along the way regarding the notion of “answer correctness”. 1.  </paper_abstract>
<query_num> 17101 </query_num>
<text>   =-=and Lin, 2003-=-) or attempting to “justify” the answer using an abductive proof (=-=Harabagiu et al., 2000-=-). In general, knowledge-based approaches (=-=e.g., Prager et al., 2000-=-), Web-based techniques (=-=e.g., Brill et al., 2001-=-), and statistical methods (=-=e.g., Echihabi and Marcu, 2003-=-) are well represented in question answering systems. In a typical question answering system, a document retriever is employed to produce a ca iversity in the types of documents that are retrieved, since most question answering teams rely on linguistically-uninformed keyword-based techniques (=-=e.g., Srihari and Li, 1999; Clarke et al., 2000;Brill et al., 2001-=-). Furthermore, according to Monz’s (2003) calculation, 28% of TREC 2002 participants and 21% of TREC 2003 participants simply used the results of the PRISE system provided by NIST. Although some syst   </text>
<query_num> 17102 </query_num>
<text>   agraph-sized fragments for subsequent analysis. Most often, passage retrieval algorithms perform a density-based weighting of query terms, i.e., they favor query terms that appear close together (see =-=Tellex et al., 2003-=- for a survey). The insight is that answers to a question are likely to occur in extents containing many closely-clustered query terms while documents containing all query terms spaced far apart are l   </text>
<query_num> 17103 </query_num>
<text>   articipants simply used the results of the PRISE system provided by NIST. Although some systems do employ advanced techniques, e.g., abductive inferencing (=-=Harabagiu et al., 2000-=-) and feedback loops (=-=Moldovan et al., 2002-=-), such systems are in the minority. Besides, it is unclear what effects these advanced techniques have on the document retrieval aspect of question answering because they primarily operate on previou   </text>
<query_num> 17104 </query_num>
<text>   as been previously reported in Bilotti (2004) and Bilotti et al. (2004). Our test collection was created using a much simplified one-shot variant of the searchguided relevance assessment methodology (=-=Cormack et al., 1998; Cieri et al., 2002-=-). Working from known answers to the questions (provided by NIST assessors), we manually crafted boolean queries with terms selected from each question and its answer⎯terms which w   </text>
<query_num> 17105 </query_num>
<text>   e the use of heuristic rules (=-=Hovy et al. 2001-=-) and machine learning techniques (=-=Li and Roth, 2002-=-), both of which may refer to a custom question type hierarchy or existing resources such as WordNet (=-=Harabagiu et al., 2000-=-). As an example, the expected answer type of the question “Where was Kennedy assassinated?” is location. The question analysis module is often also responsible for formulating one or more queries to  nguistic processing technology, such as matching syntactic relations from the questions with those from the corpus (=-=Katz and Lin, 2003-=-) or attempting to “justify” the answer using an abductive proof (=-=Harabagiu et al., 2000-=-). In general, knowledge-based approaches (=-=e.g., Prager et al., 2000-=-), Web-based techniques (=-=e.g., Brill et al., 2001-=-), and statistical methods (=-=e.g., Echihabi and Marcu, 2003-=-) are well represented in REC 2002 participants and 21% of TREC 2003 participants simply used the results of the PRISE system provided by NIST. Although some systems do employ advanced techniques, e.g., abductive inferencing (=-=Harabagiu et al., 2000-=-) and feedback loops (=-=Moldovan et al., 2002-=-), such systems are in the minority. Besides, it is unclear what effects these advanced techniques have on the document retrieval aspect of question answerin   </text>
<query_num> 17106 </query_num>
<text>   er. Another issue that Monz has studied is the effect of stemming: whereas previous studies in ad hoc retrieval have reported mixed results regarding its impact on precision and recall (=-=Harman, 1991;Krovetz, 1993; Hull, 1996-=-), he demonstrates clear precision and recall improvements that can be directly attributed to stemming document terms. Retrieval emphasis marks another divergence between ad hoc retrieval   </text>
<query_num> 17107 </query_num>
<text>   ered query terms while documents containing all query terms spaced far apart are less likely to contain the answer. In some systems, document and passage retrieval are performed simultaneously (=-=e.g., Clarke et al., 2000-=-). Finally, the answer extraction module searches the passages for the actual answers. The basic strategy is to find named entities that match the expected answer type (=-=Srihari and Li, 1999-=-), although s, there is limited diversity in the types of documents that are retrieved, since most question answering teams rely on linguistically-uninformed keyword-based techniques (=-=e.g., Srihari and Li, 1999; Clarke et al., 2000; Brill et al., 2001-=-). Furthermore, according to Monz’s (2003) calculation, 28% of TREC 2002 participants and 21% of TREC 2003 participants simply used the results of the PRISE system provided by NIST   </text>
<query_num> 17108 </query_num>
<text>   f ad hoc test collections, including the effect of topic size (=-=Voorhees and Buckley 2002-=-), the effect of incomplete judgments (=-=Buckley and Voorhees, 2004-=-), the effect of different evaluation metrics (=-=Buckley and Voorhees, 2000-=-), and different notions of relevance (=-=Voorhees 2000; Voorhees 2001a; Sormunen 2002-=-); in general, they have confirmed the reliability of existing test collections as a laboratory tool for experimentat   </text>
<query_num> 17109 </query_num>
<text>   ffect of incomplete judgments (=-=Buckley and Voorhees, 2004-=-), the effect of different evaluation metrics (=-=Buckley and Voorhees, 2000-=-), and different notions of relevance (=-=Voorhees 2000; Voorhees 2001a; Sormunen 2002-=-); in general, they have confirmed the reliability of existing test collections as a laboratory tool for experimentation and validated the general TREC methodology as an effective means for creating s   </text>
<query_num> 17110 </query_num>
<text>   iverse set of retrieval techniques. Both of these assumptions are false in the case of question answering. The average performance of current systems is still poor, despite a few outliers (see, e.g., =-=Voorhees, 2003-=-). For the TREC 2004 evaluation, Voorhees reported that the median score of 92.2% of all questions is zero. As a result, the list of known relevant documents is also quite small, averaging 3.95 releva g the performance of a system that did not participate in the original TREC evaluations. The reliability and stability of results for participating teams have been well established (see, for example, =-=Voorhees, 2003-=- and other TREC QA track overview papers). 4. Building a Reusable QA Test Collection In the previous section, we have discussed why presently available resources for evaluating question answering syst   </text>
<query_num> 17111 </query_num>
<text>   matic, the relative differences were still quite substantial. These results illustrate the unreliability of evaluating new question answering systems using the current set of relevance judgments (see =-=Lin,2005-=- for a more in-depth analysis). This is especially true for better performing systems, which are generally capable of answering more “difficult” questions. The more difficult a question is, the fewer   </text>
<query_num> 17112 </query_num>
<text>   measured by the pooled judgments. Researchers have probed other aspects of ad hoc test collections, including the effect of topic size (=-=Voorhees and Buckley 2002-=-), the effect of incomplete judgments (=-=Buckley and Voorhees, 2004-=-), the effect of different evaluation metrics (=-=Buckley and Voorhees, 2000-=-), and different notions of relevance (=-=Voorhees 2000; Voorhees 2001a; Sormunen 2002-=-); in general, they have confirmed the relia   </text>
<query_num> 17113 </query_num>
<text>   n analysis component classifies user questions into the expected semantic type of the answer. Typical approaches include the use of heuristic rules (=-=Hovy et al. 2001-=-) and machine learning techniques (=-=Li and Roth, 2002-=-), both of which may refer to a custom question type hierarchy or existing resources such as WordNet (=-=Harabagiu et al., 2000-=-). As an example, the expected answer type of the question “Where was Kenned   </text>
<query_num> 17114 </query_num>
<text>   nce of an answer. Another issue that Monz has studied is the effect of stemming: whereas previous studies in ad hoc retrieval have reported mixed results regarding its impact on precision and recall (=-=Harman, 1991; Krovetz, 1993; Hull, 1996-=-), he demonstrates clear precision and recall improvements that can be directly attributed to stemming document terms. Retrieval emphasis marks another divergence between ad   </text>
<query_num> 17115 </query_num>
<text>   swering systems today can be decomposed into four major components (see Figure 1): question analysis, document retrieval, passage retrieval, and answer extraction (=-=cf. Hirschman and Gaizauskas, 2001;Voorhees, 2001-=-). The question analysis component classifies user questions into the expected semantic type of the answer. Typical approaches include the use of heuristic rules (=-=Hovy et al. 2001-=-) and machine learnin other processing modules. 3. The TREC Question Answering Tracks Over the past few years, the question answering tracks at the Text Retrieval Conferences (TRECs) (=-=Voorhees and Tice, 1999, 2000, 2000a; Voorhees, 2001, 2002, 2003-=-), sponsored by the National Institute of Standards and Technology (NIST), have brought formal and rigorous evaluation methodologies to bear on the question answering task: features includ ley 2002), the effect of incomplete judgments (=-=Buckley and Voorhees, 2004-=-), the effect of different evaluation metrics (=-=Buckley and Voorhees, 2000-=-), and different notions of relevance (=-=Voorhees 2000; Voorhees 2001a; Sormunen 2002-=-); in general, they have confirmed the reliability of existing test collections as a laboratory tool for experimentation and validated the general TREC methodology as an effective mean   </text>
<query_num> 17116 </query_num>
<text>   using an abductive proof (=-=Harabagiu et al., 2000-=-). In general, knowledge-based approaches (=-=e.g., Prager et al., 2000-=-), Web-based techniques (=-=e.g., Brill et al., 2001-=-), and statistical methods (=-=e.g.,Echihabi and Marcu, 2003-=-) are well represented in question answering systems. In a typical question answering system, a document retriever is employed to produce a candidate set of documents for further linguistic processing   </text>
</top>
<top>
<paper_num> 172 </paper_num>
<paper_title>   Firewall Policy Queries.  </paper_title>
<paper_abstract>   Abstract—Firewalls are crucial elements in network security, and have been widely deployed in most businesses and institutions for securing private networks. The function of a firewall is to examine each incoming and outgoing packet and decide whether to accept or to discard the packet based on its policy. Due to the lack of tools for analyzing firewall policies, most firewalls on the Internet have been plagued with policy errors. A firewall policy error either creates security holes that will allow malicious traffic to sneak into a private network or blocks legitimate traffic and disrupts normal business processes, which in turn could lead to irreparable, if not tragic, consequences. Because a firewall may have a large number of rules and the rules often conflict, understanding and analyzing the function of a firewall has been known to be notoriously difficult. An effective way to assist firewall administrators to understand and analyze the function of their firewalls is by issuing queries. An example of a firewall query is “Which computers in the private network can receive packets from a known malicious host in the outside Internet? ” Two problems need to be solved in order to make firewall queries practically useful: how to describe a firewall query and how to process a firewall query. In this paper, we first introduce a simple and effective SQL-like query language, called the Structured Firewall Query Language (SFQL), for describing firewall queries. Second, we give a theorem, called the Firewall Query Theorem, as the foundation for developing firewall query processing algorithms. Third, we present an efficient firewall query processing algorithm, which uses decision diagrams as its core data structure. Fourth, we propose methods for optimizing firewall query results. Finally, we present methods for performing the union, intersect, and minus operations on firewall query results. Our experimental results show that our firewall query processing algorithm is very efficient: it takes less than 10 milliseconds to process a query over a firewall that has up to 10,000 rules. Index Terms—Network security, firewall queries, firewall testing, firewall correctness. Ç 1  </paper_abstract>
<query_num> 17201 </query_num>
<text>   .LIU AND GOUDA: FIREWALL POLICY QUERIES 773 Fig. 5. A full-length ordered FDD before reduction. nodes level by level from the terminal nodes to the root node using signatures to speed up comparisons =-=[33]-=-. Starting from the bottom level, at each level, we compute a signature for each node at that level. For a terminal node v, set vs signature to be its label. For a nonterminal node v, suppose v has k   </text>
<query_num> 17202 </query_num>
<text>   all this algorithm the FDD-based firewall query processing algorithm. Here, we give a brief introduction to firewall decision diagrams [20]. A similar data structure was used by Rovniagin and Wool in =-=[40]-=- and by Dobkin and Lipton in [12]. Definition 1. (Firewall Decision Diagram). A Firewall Decision Diagram (FDD) with a decision set DS and over fields F1; ...;Fd is an acyclic and directed graph that   </text>
<query_num> 17203 </query_num>
<text>   ang only supports queries over accept traffic. Third, we formulate firewall queries using an SQL-like language. Some firewall analysis methods have been proposed in [4], [14], [15], [19], [25], [26], =-=[29]-=-, [30], [31], [36]. In [29], Liu presented algorithms for performing the change impact analysis of firewall policies. In [30], Liu presented an algorithm for verifying firewall policies. The verificat   </text>
<query_num> 17204 </query_num>
<text>   egitimate access to the private network, they cannot find the errors that disable legitimate communication between the private network and the outside Internet. Firewall policy testing was studied in =-=[27]-=-. 3 FORMAL DEFINITIONS We now formally define the concepts of fields, packets, firewalls, and the Firewall Compression Problem. A field Fi is a variable whose domain, denoted DðFiÞ, is a finite interv   </text>
<query_num> 17205 </query_num>
<text>   eries over discard traffic, while Fang only supports queries over accept traffic. Third, we formulate firewall queries using an SQL-like language. Some firewall analysis methods have been proposed in =-=[4]-=-, [14], [15], [19], [25], [26], [29], [30], [31], [36]. In [29], Liu presented algorithms for performing the change impact analysis of firewall policies. In [30], Liu presented an algorithm for verify lyze firewall rules. Clearly, building an expert system just for analyzing a firewall is overwrought and impractical. Detecting potential firewall policy errors by conflict detection was discussed in =-=[4]-=-, [14], [25], [36]. Similar to conflict detection, some anomalies are defined and techniques for detecting anomalies are presented in [2], [47]. Examining each conflict or anomaly is helpful in reduci   </text>
<query_num> 17206 </query_num>
<text>   es in the original inconsistent firewalls. In the absence of publicly available firewalls, we create synthetic firewalls according to the characteristics of real-life packet classifiers discovered in =-=[3]-=-, [23]. Note that a firewall is also a packet classifier. Each rule has the following five fields: interface, source IP address, destination IP address, destination port number, and protocol type. The   </text>
<query_num> 17207 </query_num>
<text>   fic, while Fang only supports queries over accept traffic. Third, we formulate firewall queries using an SQL-like language. Some firewall analysis methods have been proposed in [4], [14], [15], [19], =-=[25]-=-, [26], [29], [30], [31], [36]. In [29], Liu presented algorithms for performing the change impact analysis of firewall policies. In [30], Liu presented an algorithm for verifying firewall policies. T ll rules. Clearly, building an expert system just for analyzing a firewall is overwrought and impractical. Detecting potential firewall policy errors by conflict detection was discussed in [4], [14], =-=[25]-=-, [36]. Similar to conflict detection, some anomalies are defined and techniques for detecting anomalies are presented in [2], [47]. Examining each conflict or anomaly is helpful in reducing potential   </text>
<query_num> 17208 </query_num>
<text>   gh-level packet filtering policies in [24]. Bartal et al. proposed a UML-like language for specifying global filtering policies in [5]. Design of high-performance ATM firewalls was discussed in [45], =-=[46]-=- with focus on firewall architectures. Firewall vulnerabilities were discussed and classified in [45], [46]. However, the focus of [17], [28] are the vulnerabilities of the packet filtering software a   </text>
<query_num> 17209 </query_num>
<text>   wall policy errors by conflict detection was discussed in [4], [14], [25], [36]. Similar to conflict detection, some anomalies are defined and techniques for detecting anomalies are presented in [2], =-=[47]-=-. Examining each conflict or anomaly is helpful in reducing potential firewall policy errors; however, the number of conflicts or anomalies in a firewall is typically large, and manual checking of eac   </text>
<query_num> 17210 </query_num>
<text>   y is unreliable because the meaning of each rule depends on the current order of the rules in the firewall, which may be incorrect. Some firewall design methods have been proposed in [5], [20], [21], =-=[22]-=-, [24], [32]. These works aim at creating firewall rules, while we aim at analyzing firewall rules. Gouda and Liu proposed to use decision diagrams for designing firewalls in [20], [22]. In [32], Liu  hey both point to the same node. Fig. 5 shows an FDD before reduction and Fig. 6 shows the corresponding FDD after reduction. A brute force deep comparison algorithm for FDD reduction was proposed in =-=[22]-=-. In this paper, we use a more efficient FDD reduction algorithm that processes the Authorized licensed use limited to: Michigan State University. Downloaded on October 21, 2009 at 17:49 from IEEE Xpl   </text>
</top>
<top>
<paper_num> 173 </paper_num>
<paper_title>   Variation-Aware Application Scheduling and Power Management for Chip Multiprocessors.  </paper_title>
<paper_abstract>   Within-die process variation causes individual cores in a Chip Multiprocessor (CMP) to differ substantially in both static power consumed and maximum frequency supported. In this environment, ignoring variation effects when scheduling applications or when managing power with Dynamic Voltage and Frequency Scaling (DVFS) is suboptimal. This paper proposes variation-aware algorithms for application scheduling and power management. One such power management algorithm, called LinOpt, uses linear programming to find the best voltage and frequency levels for each of the cores in the CMP — maximizing throughput at a given power budget. In a 20core CMP, the combination of variation-aware application scheduling and LinOpt increases the average throughput by 12–17 % and reduces the average ED 2 by 30–38 %  — all relative to using variation-aware scheduling together with a simple extension to Intel’s Foxton power management algorithm.  </paper_abstract>
<query_num> 17301 </query_num>
<text>   ., the number of cores in the same voltage-frequency domain). They find that having a small number of cores per domain produces the most complexity-effective design. Other researchers like Heo et al. =-=[9]-=- and Stavrou and Trancoso [34] minimize power density or temperature hot spots by judiciously scheduling jobs or migrating them from core to core. In embedded systems, linear programming has been used algorithms with the additional goal of keeping the temperature of the CMP as uniform as possible. This can be achieved through aggressive migration of applications from active to inactive cores as in =-=[9]-=-, and through temperature-aware mapping of applications to cores and assignment of (V,f) pairs. The result is likely to be fewer hot spots and lower power consumption, but it comes at the cost of incr   </text>
<query_num> 17302 </query_num>
<text>   1] study how heterogeneous CMPs impact parallel workloads. They suggest fine-granularity threading as a solution for alleviating the performance instability caused by heterogeneous CMPs. Kumar et al. =-=[18]-=- propose a CMP with different-complexity cores and the same ISA. The goal is to reduce power consumption by using the simpler, more power-efficient cores to run memory-bound applications. They schedul his pool to construct multi-programmed workloads that contain from 1 to 20 applications — where each application runs on a different core. This approach to construct workloads has been used elsewhere =-=[12, 18]-=-. Each experiment is repeated 20 times; each time with a different set of applications. We report the average outcome of the 20 trials. We use the simulation points present in SESC to run the most rep   </text>
<query_num> 17303 </query_num>
<text>   and leakage power for the scaled technology and frequency relative to the reference values. We use HotSpot [31] to estimate on-chip temperatures. To do so, we use the iterative approach of Su et al. =-=[35]-=-: the temperature is estimated based on the current total power; the leakage power is estimated based on the current temperature; and the leakage power is included in the total power. This is repeated   </text>
<query_num> 17304 </query_num>
<text>   and the voltage and frequency levels at which the CMP runs. They apply DVFS chip-wide rather than independently per core, which reduces the flexibility and impact of the optimization. Kadayif et al. =-=[15]-=- enable embedded systems to exploit the heterogeneity of workloads. Specifically, they use the compiler to assign different voltages and frequencies to different processors depending on the characteri   </text>
<query_num> 17305 </query_num>
<text>   annot be ignored at the architecture and system levels. Indeed, variation has major implications, such as increased leakage power consumption in the chips and limited processor frequency improvements =-=[2]-=-. In the context of Chip Multiprocessors (CMP), within-die process variation in current and near-future technologies causes individual cores in the chip to differ substantially in the amount of power   Background: Process Variation Process variation refers to changes in transistor parameters beyond their nominal values, and results from manufacturing difficulties in very small feature technologies =-=[2]-=-. For the purposes of modeling and analysis, variation is generally broken down into dieto-die (D2D) and within-die (WID). WID variation can be further divided into systematic and random effects. In t   </text>
<query_num> 17306 </query_num>
<text>   be set independently, although all cores have the same voltage. Currently, multiple on-chip voltages are provided by off-chip voltage regulators, which are bulky and costly. Recent work by Kim et al. =-=[16]-=- describes designs of on-chip regulators. They are able to perform voltage changes in nanoseconds rather than in microseconds, and consume little energy. Designs similar to these will enable wide use  regulators that generate the different voltages. Currently, such regulators are on the board, but new technologies will soon make it possible to place them on the processor package or even on the die =-=[16]-=-. In this case, voltage transition speeds will be orders of magnitude faster. In this paper, however, we conservatively assume that the voltage and frequency transition speeds are those of current sys   </text>
<query_num> 17307 </query_num>
<text>   both application scheduling and power management have to be variationaware. Moreover, given the large design space, we need to use an intelligent way to prune the design space. Herbert and Marculescu =-=[10]-=- examine the tradeoffs of using different DVFS granularities (i.e., the number of cores in the same voltage-frequency domain). They find that having a small number of cores per domain produces the mos   </text>
<query_num> 17308 </query_num>
<text>   cache), average frequency of the active cores, throughput (measured in millions of instructions per second or MIPS), and the energy delay-square product (ED 2 ). We also give the weighted throughput =-=[32]-=-, which uses the weighted IPCs of the applications. The weighted IPC of an application is computed as the application’s IPC normalized to the application’s IPC at reference conditions. This metric giv   </text>
<query_num> 17309 </query_num>
<text>   der a range of values for V th ’s σ/µ, namely 0.03– 0.12, and use as default 0.12. Moreover, we assume that the random and systematic components have equal variances. For φ, we use Friedberg’s et al. =-=[8]-=- measurement that the gate length had a correlation range close to 0.5 of the chip’s width. Since the systematic component of V th ’s variation directly depends on the gate length’s variation, we set   </text>
<query_num> 17310 </query_num>
<text>   e Model To estimate power, we scale the results given by popular tools using technology projections from ITRS [14]. Specifically, we use SESC, which is augmented with dynamic power models from Wattch =-=[3]-=- to estimate dynamic power at a reference technology and frequency. In addition, we use HotLeakage [39] to estimate leakage power at the same reference technology. Then, we obtain ITRS’s scaling proje   </text>
<query_num> 17311 </query_num>
<text>   ne solving the optimization problem of Section 4.3.1 using a non-linear algorithm. We choose simulated annealing (SAnn) — a well-known probabilistic algorithm for solving global optimization problems =-=[17]-=-. The goal of SAnn is the same as the one used with LinOpt: maximize throughput under power constraints. For a given mapping of threads to cores, the search space of the SAnn algorithm consists of all   </text>
<query_num> 17312 </query_num>
<text>   ores leak different amounts. Chip manufacturers hardly release any measurements on process variation for current or future technologies. As a result, we rely on statistical models of variation (e.g., =-=[11, 20, 21, 28, 30, 37]-=-) driven by values projected by the ITRS [14]. In this paper, we use the VARIUS model [30, 37], which we briefly summarize here. To model systematic variation, the chip is divided into a grid. Each gr le 4 summarizes the architecture configuration. In the following, we discuss the different parts of our infrastructure. 6.1. Variation Model Parameters To model WID variation, we use the VARIUS model =-=[30, 37]-=- to generate V th and L eff variation maps. We then superimpose these maps on our floorplan as shown in Figure 3. This allows us to model how variation affects core power and frequency. Table 4 shows   </text>
<query_num> 17313 </query_num>
<text>   with the goal ofsmaximizing the chip-wide performance/power ratio. We combine application scheduling and global DVFS power management. Scheduling for Heterogeneous Architectures. Balakrishnan et al. =-=[1]-=- study how heterogeneous CMPs impact parallel workloads. They suggest fine-granularity threading as a solution for alleviating the performance instability caused by heterogeneous CMPs. Kumar et al. [1   </text>
</top>
<top>
<paper_num> 174 </paper_num>
<paper_title>   Content-based retrieval of technical drawings.  </paper_title>
<paper_abstract>   This paper presents a new approach to classify, index and retrieve technical draw-ings by content. Our work uses spatial relationships, shape geometry and high-dimensional indexing mechanisms to retrieve complex drawings from CAD databa-ses. This contrasts with conventional approaches which use mostly textual metadata for the same purpose. Creative designers and draftspeople often re-use data from previous projects, publications and libraries of ready-to-use components. Usually, retrieving these drawings is a slow, complex and error-prone endeavor, requiring either exhaustive visual examination, a solid memory, or both. Unfortunately, the widespread use of CAD systems, while making it easier to create and edit drawings, exacerbates this problem, insofar as the number of projects and drawings grows enormously, without providing adequate searching mechanisms to support retrieving these documents. We describe an approach that supports automatic indexation of technical drawing databases through drawing simplification, feature extraction and efficient algorithms to index large amounts of data. We describe in detail all the steps of our classification 1 process to content-based retrieval of technical drawings (CAD) and present results from usability tests on our prototype.  </paper_abstract>
<query_num> 17401 </query_num>
<text>   22s5.2 Indexing Structure In this section we shortly describe experimental comparison of our indexing structure (NB-Tree) to the most popular approaches available, such as the SR-Tree [43], the ATree =-=[44]-=- and the Pyramid Technique [45]. All experiments were performed on a Intel Pentium II @ 233 MHz running Linux 2.4.8 and with 384 MB of RAM. More detailed reports can be found in [29, 30]. Figure 13.a   </text>
<query_num> 17402 </query_num>
<text>   and search simple image datasets using hand-sketches as queries. Size functions are a relatively new class of shape descriptors, based on geometric-topological theory of critical points. Shock trees =-=[22]-=- are another method to describe and compare shapes. Pelillo presented a solution to matching two shock trees by constructing the association graph [23]. Authors illustrate the power of this approach b   </text>
<query_num> 17403 </query_num>
<text>   chnical drawings through textual databases. However, these fail to use the rich visual association mechanisms and designer’s use of sketches to recover information. Gross’s Electronic Cocktail Napkin =-=[12, 2, 13]-=- addressed a visual retrieval scheme based on diagrams, to indexing databases of architectural drawings. Users draw sketches of buildings, which are compared with annotations (diagrams), stored in a d   </text>
<query_num> 17404 </query_num>
<text>   escriptors with low dimensions. To acquire geometric information about drawings we use a general shape recognition library able to identify a set of geometric shapes and gestural commands called CALI =-=[27]-=-. This enables us to use either drawing data or sketches as input, which is a desirable feature of our system, as we shall see later on. In our approach instead of using CALI to recognize a shape or a   </text>
<query_num> 17405 </query_num>
<text>   iscarded. The algorithm described herein uses the method recently proposed by Haperin and Pecker in [31], which is based on the method presented by Goodrich, Guibas, Hershberger and Paul Tanenbaum in =-=[32]-=-. Either approach preserve the topological properties of the original line segments, which is important for our content-based retrieval approach. After snap rounding and intersection removal our algor   </text>
<query_num> 17406 </query_num>
<text>   ly yield approximate matches to the query. However, nearest neighbor search in high-dimensional data spaces is a difficult problem. We developed a new multidimensional indexing structure, the NB-Tree =-=[29, 30]-=-, that satisfies the requirements enumerated before, providing us with an efficient indexing mechanism for high–dimensional data points of variable dimension. The NB-Tree is a simple, yet efficient in e [43], the ATree [44] and the Pyramid Technique [45]. All experiments were performed on a Intel Pentium II @ 233 MHz running Linux 2.4.8 and with 384 MB of RAM. More detailed reports can be found in =-=[29, 30]-=-. Figure 13.a depicts the performance of nearest neighbor searches for synthetic data sets of uniformly distributed data points, when data dimension increases. We can see that the NB-Tree outperforms   </text>
<query_num> 17407 </query_num>
<text>   namely Fourier descriptors (FD), grid-based (GB), Delaunay triangulation (DT) and Touchpoint-vertex-angle-sequence (TPVAS). To that end we used results of an experiment previously performed by Safar =-=[46]-=-, where he compared his method (TPVAS) with the FD, GB and DT methods. In that experiment, authors used a database containing 100 contours 24sPrecision 100 90 80 70 60 50 40 30 20 10 0 0 10 20 30 40 5 y to the query. For each of the five queries, we determined the positions for the 10 similar shapes in the ordered response set. Using results from our method and the values presented in Table 2 from =-=[46]-=- we derived the precision-recall plot shown in Figure 14. Looking at Figure 14 we can see that our technique outperforms all the other methods, yielding good precision figures for recall values up to   </text>
<query_num> 17408 </query_num>
<text>   nical drawings. 22s5.2 Indexing Structure In this section we shortly describe experimental comparison of our indexing structure (NB-Tree) to the most popular approaches available, such as the SR-Tree =-=[43]-=-, the ATree [44] and the Pyramid Technique [45]. All experiments were performed on a Intel Pentium II @ 233 MHz running Linux 2.4.8 and with 384 MB of RAM. More detailed reports can be found in [29, 3   </text>
<query_num> 17409 </query_num>
<text>   olic representation of the query gets compared to all the sym5sbolic representations stored in the database, making this work difficult to scale up for large collections of images. Funkhouser et. al. =-=[20]-=- describe a method for retrieving 3D shapes using sketched contours. However, their approach relies on silhouettes and their fitting to projections of 3D images, unlike our method which is based on st   </text>
<query_num> 17410 </query_num>
<text>   pological theory of critical points. Shock trees [22] are another method to describe and compare shapes. Pelillo presented a solution to matching two shock trees by constructing the association graph =-=[23]-=-. Authors illustrate the power of this approach by matching articulated and deformed shapes described by shock trees. Shokoufandeh et al developed another approach to perform shock tree matching based   </text>
<query_num> 17411 </query_num>
<text>   pproach by matching articulated and deformed shapes described by shock trees. Shokoufandeh et al developed another approach to perform shock tree matching based on graph spectrum and Voronoi diagrams =-=[24]-=-. While these approaches use trees (graphs) to describe the contour of simple shapes, we use graphs to represent the spatial structure of complex drawings. More recently Shokoufandeh et al presented a rt graphs into vector descriptors that can be manipulated using a multidimensional indexing structure. The spectrum of a graph is calculated from the eigenvalues of its adjacency matrix. According to =-=[26, 24]-=- the use of eigenvalues (spectrum) of a graph as an indexing method is valid since (1) it captures local topology, (2) is invariant to subgraph re-order and (3) is stable, since small changes in the g e little changes in its spectrum. However, resulting descriptors are not unique. More than one graph can have the same spectrum, which gives rise to collisions similar to these in hashing schemes. In =-=[24]-=- authors argue that these collisions occur rather infrequently, a claim seemingly verified by our experiments. B A C D E A B C D E A 0 1 1 1 0 B 1 0 1 0 0 C 1 1 0 0 0 D 1 0 0 0 1 E 0 0 0 1 0 0.25 0.68 e absolute values to obtain the topology descriptor. Adjacency matrices are symmetric, assuring that eigenvalues are always real. 20sExperimental Comparison While previous work by Shokoufandeh et al. =-=[24]-=- is also based on eigenvalues, they sums these to reduce the dimension of data rather than using eigenvalues by themselves. This is because efficient indexing structures for high dimensional data poin   </text>
<query_num> 17412 </query_num>
<text>   ropriated for large collections of drawings, since they perform a sequential scan through the database comparing the query with all indexed drawings. Leung and Chen proposed a sketch retrieval method =-=[17]-=- for general unstructured free-form hand-drawings stored in the form of multiple strokes. They use shape information from each stroke exploiting the geometric relationship between multiple strokes for   </text>
<query_num> 17413 </query_num>
<text>   rse specification of queries. This method is also applicable to query-by-example, without modifications. We have also used our approach to develop a Sketch-Based Retrieval system for ClipArt drawings =-=[47]-=-. Although this is another domain of application, where the geometric information is more relevant than topology, experimental evaluation yielded good results with a larger database (1,000 drawings an   </text>
<query_num> 17414 </query_num>
<text>   shapes, we use graphs to represent the spatial structure of complex drawings. More recently Shokoufandeh et al presented a framework for shape matching through scale-space decomposition of 3D models =-=[25]-=-. Their algorithm is based on efficient hierarchical decomposition of metric data using its spectral properties. 3D objects are mapped into rooted trees, thus recasting the problem of finding a match   </text>
</top>
<top>
<paper_num> 175 </paper_num>
<paper_title>   RI-MAC: a receiver-initiated asynchronous duty cycle mac protocol for dynamic traffic loads in wireless sensor networks.  </paper_title>
<paper_abstract>   The problem of idle listening is one of the most significant sources of energy consumption in wireless sensor nodes, and many techniques have been proposed based on duty cycling to reduce this cost. In this paper, we present a new asynchronous duty cycle MAC protocol, called Receiver-Initiated MAC (RI-MAC), that uses receiver-initiated data transmission in order to efficiently and effectively operate over a wide range of traffic loads. RI-MAC attempts to minimize the time a sender and its intended receiver occupy the wireless medium to find a rendezvous time for exchanging data, while still decoupling the sender and receiver’s duty cycle schedules. We show the performance of RI-MAC through detailed ns-2 simulation and through measurements of an implementation in TinyOS in a testbed of MICAz motes. Compared to the prior asynchronous duty cycling approach of X-MAC, RI-MAC achieves higher throughput, packet delivery ratio, and power efficiency under a wide range of traffic loads. Especially when there are contending flows, such as bursty traffic or transmissions from hidden nodes, RI-MAC significantly improves throughput and packet delivery ratio. Even under light traffic load for which X-MAC is optimized, RI-MAC achieves the same high performance in terms of packet delivery ratio and latency while maintaining comparable power efficiency.  </paper_abstract>
<query_num> 17501 </query_num>
<text>   5 Special Frame Short Preamble — Beacon Special Frame Size 6 B — 6 – 9 B Dwell Time 10.5 ms 100 ms Variable sensing range to the transmission range are also observed in some state-of-art sensor nodes =-=[2]-=-. Table 2 summarizes the MAC protocol parameters we used in our simulations. Backoff strategy and retransmission have not been explicitly discussed in prior work [3, 15], as X-MAC is optimized for lig   </text>
<query_num> 17502 </query_num>
<text>   detecting a collision, a receiver calculates the new BW value that will be used in the next beacon, by employing some backoff strategy such as binary exponential backoff (BEB) in IEEE 802.11 or Sift =-=[14,21]-=-, depending on the density of a network. BEB is used in our implementation in TinyOS, as we found it adapts to networks of different densities and resolves collisions efficiently in RI-MAC in our eval   </text>
<query_num> 17503 </query_num>
<text>   e to convergecast [25] and correlated-event workload traffic [13], where multiple sensors that have detected the same event send their reports to the sink node or to a node that does data aggregation =-=[11]-=-. As traffic in a WSN can be quite dynamic, depending on the events being sensed and the sensing application and protocols being used, an ideal WSN MAC protocol should perform well under a wide range   </text>
<query_num> 17504 </query_num>
<text>   edium is not occupied, some of them could experience significant delay. This is often the case under bursty or high traffic load such as due to convergecast [25] and correlated-event workload traffic =-=[13]-=-, where multiple sensors that have detected the same event send their reports to the sink node or to a node that does data aggregation [11]. As traffic in a WSN can be quite dynamic, depending on the   is 200 meters from its neighbors, and the sink node is at the center. In our simulations, we used a Random Correlated-Event (RCE) traffic model [20]. This model, based on a correlated-event workload =-=[13]-=-, simulates the impulse traffic triggered by spatiallycorrelated events commonly observed in detection and tracking applications. RCE picks a random (x,y) location for each event. IfTable 3: Average   </text>
<query_num> 17505 </query_num>
<text>   er and over again. Synchronized duty cycle MAC protocols, such as S-MAC [23], T-MAC [7], RMAC [8], and DW-MAC [20], also achieve great energy efficiency in WSNs; so too, hybrid approaches such as SCP =-=[24]-=-. The major difference between RI-MAC and these MAC protocols is that RI-MAC does not require any synchronization, thus saving the overhead and complexity of clock synchronization. Even though no node ensor node. Most of these parameters are from the data sheet of CC2420 radio [4], which is used in popular motes such as MICAz and TelosB. The RSSI sampling delay for CC2420 was reported by Ye et al. =-=[24]-=-; we use this delay as the time for a single CCA (clear channel assessment) check, i.e., the delay before actual transmission starts after a STXONCCA command is strobed [4]. The transmission range and hortest path between any two nodes. We also ensure that no network used in our simulations is partitioned. As energy consumption of different radios varies significantly, even in the same radio state =-=[24]-=-, we report effective duty cycle in evaluating power efficiency, as done in prior work [3, 15]. The sleep interval for all three MAC protocols is 1 second, and the initial wakeup time of each node was   </text>
<query_num> 17506 </query_num>
<text>   io turned on, listening for a possible packet to be received even though none has been sent. Many solutions to the problem of idle listening have been proposed utilizing the technique of duty cycling =-=[19, 23]-=-. In this technique, each sensor node turns its radio on only periodically, alternating between active and sleeping states. For example, with a 10% duty cycle, a node has its radio on only 10% of the  red synchronization introduces extra overhead and complexity, and a node may need to wake up multiple times if its neighbors are on different schedules. Existing asynchronous approaches such as B-MAC =-=[19]-=-, X-MAC [3], and WiseMAC [9], on the other hand, allow nodes to operate independently, with each node on its own duty cycle schedule. Such protocols typically employing low power listening (LPL), in w As such, features of RI-MAC such as back-to-back data transmission and collision detection and recovery were not discussed in LPP. In sensor network MAC protocols not using receiver-initiation, B-MAC =-=[19]-=- and X-MAC [3] are representative asynchronous duty cycle-based protocols. In B-MAC, each node periodically wakes up to check if there is any activity currently on the wireless channel. If so, the nod   </text>
<query_num> 17507 </query_num>
<text>   io turned on, listening for a possible packet to be received even though none has been sent. Many solutions to the problem of idle listening have been proposed utilizing the technique of duty cycling =-=[19, 23]-=-. In this technique, each sensor node turns its radio on only periodically, alternating between active and sleeping states. For example, with a 10% duty cycle, a node has its radio on only 10% of the  ts radio to save energy. Contention-based duty cycle MAC protocols in the literature can be roughly categorized into two categories: synchronous and asynchronous. Synchronous approaches such as S-MAC =-=[23]-=-, T-MAC [7], RMAC [8], and DW-MAC [20] synchronize neighboring nodes in order to align their active or sleeping periods. Neighbor nodes start exchanging packets only within the common active time, ena r time, allowing problems such as starvation due to repeated collisions between competing nodes that wake up at the same time over and over again. Synchronized duty cycle MAC protocols, such as S-MAC =-=[23]-=-, T-MAC [7], RMAC [8], and DW-MAC [20], also achieve great energy efficiency in WSNs; so too, hybrid approaches such as SCP [24]. The major difference between RI-MAC and these MAC protocols is that RI   </text>
<query_num> 17508 </query_num>
<text>   ization introduces extra overhead and complexity, and a node may need to wake up multiple times if its neighbors are on different schedules. Existing asynchronous approaches such as B-MAC [19], X-MAC =-=[3]-=-, and WiseMAC [9], on the other hand, allow nodes to operate independently, with each node on its own duty cycle schedule. Such protocols typically employing low power listening (LPL), in which, prior es of RI-MAC such as back-to-back data transmission and collision detection and recovery were not discussed in LPP. In sensor network MAC protocols not using receiver-initiation, B-MAC [19] and X-MAC =-=[3]-=- are representative asynchronous duty cycle-based protocols. In B-MAC, each node periodically wakes up to check if there is any activity currently on the wireless channel. If so, the node remains acti  in some state-of-art sensor nodes [2]. Table 2 summarizes the MAC protocol parameters we used in our simulations. Backoff strategy and retransmission have not been explicitly discussed in prior work =-=[3, 15]-=-, as X-MAC is optimized for light traffic load. We use 32 as the initial backoff window and 8 as the congestion backoff window, which are the default values used in the UPMA package distributed with T on. The backoff window size for beacon transmission is fixed at 32 slots in RI-MAC. Although retransmission was not included in X-MAC’s original design (none was specified in X-MAC’s published design =-=[3, 15, 22]-=-), for fair comparison with RI-MAC in which retransmission is included, we evaluated X-MAC and X-MAC-UPMA both with and without retransmission in our simulations. When retransmission was enabled, we u arting a timer that does CCA checks every 20 ms, and each CCA check lasts longer than the gap between short preambles. The time 20 ms was used because that is the wake time used in X-MAC’s evaluation =-=[3]-=-. In X-MAC-UPMA, a node that has detected busy medium turns off its radio if no packet is received within 100 ms, according to the code in the UPMA distribution. In our simulated X-MAC-UPMA, similar t   </text>
<query_num> 17509 </query_num>
<text>   save energy. Contention-based duty cycle MAC protocols in the literature can be roughly categorized into two categories: synchronous and asynchronous. Synchronous approaches such as S-MAC [23], T-MAC =-=[7]-=-, RMAC [8], and DW-MAC [20] synchronize neighboring nodes in order to align their active or sleeping periods. Neighbor nodes start exchanging packets only within the common active time, enabling a nod wing problems such as starvation due to repeated collisions between competing nodes that wake up at the same time over and over again. Synchronized duty cycle MAC protocols, such as S-MAC [23], T-MAC =-=[7]-=-, RMAC [8], and DW-MAC [20], also achieve great energy efficiency in WSNs; so too, hybrid approaches such as SCP [24]. The major difference between RI-MAC and these MAC protocols is that RI-MAC does n   </text>
<query_num> 17510 </query_num>
<text>   sfully transmit a packet to the receiver even if the transmission overlaps with others, especially when senders have different distances to the receiver (and thus different received signal strengths) =-=[16, 17]-=-. 3.5. Collision Detection and Retransmissions By coordinating DATA frame transmissions at receivers, RI-MAC greatly reduces the cost for detecting collisions and recovering lost DATA frames compared   </text>
<query_num> 17511 </query_num>
<text>   ta. As these nodes have to wait until the medium is not occupied, some of them could experience significant delay. This is often the case under bursty or high traffic load such as due to convergecast =-=[25]-=- and correlated-event workload traffic [13], where multiple sensors that have detected the same event send their reports to the sink node or to a node that does data aggregation [11]. As traffic in a   </text>
<query_num> 17512 </query_num>
<text>   ued packets to be transmitted immediately. We refer to this duration as the dwell time in the rest of this paper. The UPMA (Unified Power Management Architecture for Wireless Sensor Networks) package =-=[15]-=- implemented a variation of X-MAC in TinyOS, in which the DATA frame itself is used as the short preamble, as illustrated in Figure 2. This strategy simplifies implementation and helps a sender to det limitations, we leave further investigation and experimental work on broadcast traffic as future work. 3.8. RI-MAC Implementation in TinyOS We implemented our RI-MAC protocol under the UPMA framework =-=[15]-=- in TinyOS on a network of MICAz sensor motes. The composition of RI-MAC under the UPMA framework in our implementation is shown in Figure 8. We implemented RI-MAC for the CC2420 radio, which is a pac network of MICAz motes running TinyOS; our experimental results match our results obtained in simulation and further verify RI-MAC’s performance advantages over existing protocols. Since Klues et al. =-=[15]-=- have implemented X-MAC-UPMA on real motes and shown that X-MAC-UPMA outperforms B-MAC and SCP, in this paper, we compared RI-MAC only against X-MAC and X-MAC-UPMA. 4.1. Simulation Evaluation In our s  in some state-of-art sensor nodes [2]. Table 2 summarizes the MAC protocol parameters we used in our simulations. Backoff strategy and retransmission have not been explicitly discussed in prior work =-=[3, 15]-=-, as X-MAC is optimized for light traffic load. We use 32 as the initial backoff window and 8 as the congestion backoff window, which are the default values used in the UPMA package distributed with T on. The backoff window size for beacon transmission is fixed at 32 slots in RI-MAC. Although retransmission was not included in X-MAC’s original design (none was specified in X-MAC’s published design =-=[3, 15, 22]-=-), for fair comparison with RI-MAC in which retransmission is included, we evaluated X-MAC and X-MAC-UPMA both with and without retransmission in our simulations. When retransmission was enabled, we u   </text>
<query_num> 17513 </query_num>
<text>   y. Contention-based duty cycle MAC protocols in the literature can be roughly categorized into two categories: synchronous and asynchronous. Synchronous approaches such as S-MAC [23], T-MAC [7], RMAC =-=[8]-=-, and DW-MAC [20] synchronize neighboring nodes in order to align their active or sleeping periods. Neighbor nodes start exchanging packets only within the common active time, enabling a node to sleep ems such as starvation due to repeated collisions between competing nodes that wake up at the same time over and over again. Synchronized duty cycle MAC protocols, such as S-MAC [23], T-MAC [7], RMAC =-=[8]-=-, and DW-MAC [20], also achieve great energy efficiency in WSNs; so too, hybrid approaches such as SCP [24]. The major difference between RI-MAC and these MAC protocols is that RI-MAC does not require   </text>
</top>
<top>
<paper_num> 176 </paper_num>
<paper_title>   Memory System Characterization of Commercial Workloads.  </paper_title>
<paper_abstract>   Commercial applications such as databases and Web servers constitute the largest and fastest-growing segment of the market for multiprocessor servers. Ongoing innovations in disk subsystems, along with the ever increasing gap between processor and memory speeds, have elevated memory system design as the critical performance factor for such workloads. However, most current server designs have been optimized to perform well on scientific and engineering workloads, potentially leading to design decisions that are non-ideal for commercial applications. The above problem is exacerbated by the lack of information on the performance requirements of commercial workloads, the lack of available applications for widespread study, and the fact that most representative applications are too large and complex to serve as suitable benchmarks for evaluating trade-offs in the design of processors and servers. This paper presents a detailed performance study of three important classes of commercial workl...  </paper_abstract>
<query_num> 17601 </query_num>
<text>   ark suite. Similarly, design of multiprocessor architectures, along with academic research in this area, have been heavily influenced by popular scientific and engineering benchmarks such as SPLASH-2 =-=[26]-=- and STREAMS [13], with only a handful of published architectural studies that have in some way tried to address issues specific to commercial workloads [3, 7, 9, 12, 14, 16, 20, 21, 24]. The lack of   </text>
<query_num> 17602 </query_num>
<text>   ation detail, enabling the user to choose the most appropriate trade-off between simulation detail and slowdown. Its fastest simulator uses an on-the-fly binary translation technique similar to Embra =-=[25]-=- to position the workload into a steady state. The ability to simulate both user and system code under SimOS is essential given the rich level of system interactions exhibited by commercial workloads.   </text>
<query_num> 17603 </query_num>
<text>   e 7.3.2 DBMS. The Oracle 7.3.2 DBMS runs on both uniprocessors and multiprocessor shared memory machines, and recent benchmark results demonstrate that the software scales well on current SMP systems =-=[4]-=-. Figure 1 illustrates the different components of Oracle. The server executes as a collection of Unix processes that share a common large shared memory segment, called the System Global Area, or SGA. as far as processor and memory system are concerned. We have verified this similarity by comparing our memory system performance results presented in Section 4 with results from audit-size TPC-C runs =-=[4]-=- on the same hardware platform (AlphaServer 4100). We have also confirmed this similarity on other Alpha multiprocessor platforms. We provide a few general statistics, gathered with the DCPI [1] profi compared to DSS or AltaVista. The cycle breakdown shows a CPI of 7.0 for our OLTP workload based on TPC-B, which is about twice the CPI of audit-size TPC-C runs on the same family of hardware servers =-=[4]-=-. This difference is primarily due to more complex and compute intensive queries in TPC-C and the better locality of that workload due to the higher number of small database tables. Nevertheless, the  es have raised the level of awareness in the architecture community to the fact that OLTP workloads have a very different behavior when compared with scientific applications. Cvetanovic and Donaldson =-=[4]-=-, in their characterization of the AlphaServer 4100, measure the performance of SPEC95 programs, the TPC-C benchmark and other industry benchmarks. Our paper differs from this study in that it focuses   </text>
<query_num> 17604 </query_num>
<text>   nd engineering benchmarks such as SPLASH-2 [26] and STREAMS [13], with only a handful of published architectural studies that have in some way tried to address issues specific to commercial workloads =-=[3, 7, 9, 12, 14, 16, 20, 21, 24]-=-. The lack of architectural research on commercial applications is partly due to the fact that I/O issues have been historically considered as the primary performance bottleneck for such workloads. Ho For example, commercial database engines have improved tremendously during the past few years in scaling to more processors in shared-memory systems. Therefore, studies that are a few years old (e.g. =-=[16]-=-) or that are based on software that is not of commercial grade (e.g., the public domain Postgres database [21]) are likely to be outdated and non-representative. The above issues are even more pronou hakkar and Sweiger [20] have measured the performance of an early OLTP benchmark (TP1) in the Sequent Symmetry multiprocessor, using both memory-resident and out-of-memory databases. Rosenblum et al. =-=[16]-=- focused on the impact of architectural trends on the operating system performance using three applications, including TPC-B. Both studies identify I/O performance as the critical factor for OLTP appl   </text>
<query_num> 17605 </query_num>
<text>   nd engineering benchmarks such as SPLASH-2 [26] and STREAMS [13], with only a handful of published architectural studies that have in some way tried to address issues specific to commercial workloads =-=[3, 7, 9, 12, 14, 16, 20, 21, 24]-=-. The lack of architectural research on commercial applications is partly due to the fact that I/O issues have been historically considered as the primary performance bottleneck for such workloads. Ho e processors in shared-memory systems. Therefore, studies that are a few years old (e.g. [16]) or that are based on software that is not of commercial grade (e.g., the public domain Postgres database =-=[21]-=-) are likely to be outdated and non-representative. The above issues are even more pronounced for Web workloads since numerous applications in this area are yet to be developed, and the existing ones  y conclude that multithreading is effective in hiding latency of OLTP. Verghese et al. [24] use a DSS workload to evaluate operating system support for page migration and replication. Trancoso et al. =-=[21]-=- use a public domain database engine to perform a detailed study of the memory performance of some TPC-D queries. However their engine could not automatically parallelize the queries, nor had the effi   </text>
<query_num> 17606 </query_num>
<text>   nd engineering benchmarks such as SPLASH-2 [26] and STREAMS [13], with only a handful of published architectural studies that have in some way tried to address issues specific to commercial workloads =-=[3, 7, 9, 12, 14, 16, 20, 21, 24]-=-. The lack of architectural research on commercial applications is partly due to the fact that I/O issues have been historically considered as the primary performance bottleneck for such workloads. Ho ith, the Alpha 21164 processor provides a rich set of event counters that can be used to construct a detailed view of processor behavior, including all activity within the three-level cache hierarchy =-=[3]-=-. We used the IPROBE monitoring tool to gain access to the processor event counters. A typical monitoring experiment involved multiple runs of the workload with IPROBE, measuring a single event in eac   </text>
<query_num> 17607 </query_num>
<text>   nd engineering benchmarks such as SPLASH-2 [26] and STREAMS [13], with only a handful of published architectural studies that have in some way tried to address issues specific to commercial workloads =-=[3, 7, 9, 12, 14, 16, 20, 21, 24]-=-. The lack of architectural research on commercial applications is partly due to the fact that I/O issues have been historically considered as the primary performance bottleneck for such workloads. Ho oblem has been addressed. We have observed that the I/O problem has indeed been addressed in modern database engines, and that memory system performance is already the main bottleneck. Perl and Sites =-=[14]-=- study Microsoft SQL Server performance on an Alpha-based Windows NT server through trace-driven simulation. They claim that OLTP performance is mostly limited by chip pin bandwidth, since the bandwid   </text>
<query_num> 17608 </query_num>
<text>   oring experiments on Alpha multiprocessor platforms, using a wide range of hardware and software monitoring tools. The second set of results uses full system simulation (using our Alpha port of SimOS =-=[15]-=-) to study the effect of architectural variations. In dealing with the large scale and complexity of these workloads, we have identified a number of simplifications that make these workloads more amen ng experiments by using simulations to primarily study the effect of architectural variations on our workloads. For this purpose, we developed an Alpha port of the SimOS simulation environment. SimOS =-=[15]-=- is a full system simulation environment originally developed at Stanford University to study MIPS-based multiprocessors. Our version of SimOS simulates the hardware components of Alpha-based multipro   </text>
<query_num> 17609 </query_num>
<text>   profiling system that is also based on the processor event counters, and is especially useful because of its ability to associate event frequencies with specific executable images or processes. ATOM =-=[17]-=- is a static binary translator that facilitates the instrumentation of an executable image. We used ATOM to instrument the Oracle binary to identify different memory regions accessed by the servers, t  user-level simulations however, other simplifications beyond scaling have to be addressed. The following are observations from experiments we have done based on tracing user-level activity with ATOM =-=[17]-=- to determine the viability of using such traces. It is important to consider the effect of not modeling operating system instructions. For OLTP, operating system activity is non-negligible but it als   </text>
<query_num> 17610 </query_num>
<text>   runs [4] on the same hardware platform (AlphaServer 4100). We have also confirmed this similarity on other Alpha multiprocessor platforms. We provide a few general statistics, gathered with the DCPI =-=[1]-=- profiling tool, on the high-level behavior of the workload. On a four processor AlphaServer, the workload spends 71% in user-mode, 18% in the kernel and 11% in the idle loop. As expected from a tuned run. 1 Event types available include counts of accesses and misses in the various caches, TLB misses, types of instructions, branch mispredicts, issue widths, memory barriers, replay traps, etc. DCPI =-=[1]-=- (Digital Continuous Profiling Infrastructure) is an extremely low overhead sampling-based profiling system that is also based on the processor event counters, and is especially useful because of its   </text>
<query_num> 17611 </query_num>
<text>   ses, allowing the memory activity of daemon processes to be ignored as a simplification. This methodology has been used to study the effectiveness of simultaneous multithreading on database workloads =-=[10]-=-. Overall, while user-level simulation studies are feasible, they require a deep understanding of the system implications of the workload to ensure representative results. 7 In our study those queries   </text>
</top>
<top>
<paper_num> 177 </paper_num>
<paper_title>   Energy Scaling Laws for Distributed Inference in Random Fusion Networks.  </paper_title>
<paper_abstract>   The energy scaling laws of multihop data fusion networks for distributed inference are considered. The fusion network consists of randomly located sensors distributed i.i.d. according to a general spatial distribution in an expanding region. Among the class of data fusion schemes that enable optimal inference at the fusion center for Markov random field (MRF) hypotheses, the scheme with minimum average energy consumption is bounded below by average energy of fusion along the minimum spanning tree, and above by a suboptimal scheme, referred to as Data Fusion for Markov Random Fields (DFMRF). Scaling laws are derived for the optimal and suboptimal fusion policies. It is shown that the average asymptotic energy of the DFMRF scheme is finite for a class of MRF models.  </paper_abstract>
<query_num> 17701 </query_num>
<text>   ays possible, and at the end of Section IV, we will discuss examples of correlation structures where scalable data fusion does not exist. C. Prior and related work The seminal work of Gupta and Kumar =-=[11]-=- on the capacity of wireless networks has stimulated extensive studies covering a broad range of networking problems with different performance metrics. See also [12]. Here, we restrict ourselves to r   </text>
<query_num> 17702 </query_num>
<text>   cement, we analyzed the minimum energy fusion scheme for optimal inference and showed that it reduces to a Steiner tree under certain constraints. We also proposed a heuristic called the DFMRF 2 . In =-=[36]-=-, we analyzed the optimal sensor density for uniform node placement which maximizes the inference error exponent under an average energy constraint, and in [37], [48], we derived the error exponent fo   </text>
<query_num> 17703 </query_num>
<text>   e energy scaling laws. Given sensor locations Vn and possibly correlated sensor measurements, finding the minimum energy fusion policy under the constraint of optimal inference is, in general, NPhard =-=[6]-=-, and hence, intractable. We will establish upper and lower bounds on the fusion energy of this optimal scheme and analyze their scaling behavior. The lower bound is obtained by a scheme conducting fu placed sensors, and energy scaling laws are not established. The results presented in this paper extend some of our earlier work in the direction of scaling-law analysis in random fusion networks. In =-=[6]-=-, [10], [35], for fixed network size and node placement, we analyzed the minimum energy fusion scheme for optimal inference and showed that it reduces to a Steiner tree under certain constraints. We a (pdf) on Q1 which is bounded away from zero and infinity. We then generate Vi by scaling Xi accordingly: Vi = √ n λXi ∈ Q n λ . A useful special case is 2 The DFMRF scheme is referred to as AggMST in =-=[6]-=-, [35]. 3 The results in this paper hold for κ defined on any convex unit area. the uniform distribution (κ ≡ 1). Let Pλ be the homogeneous Poisson distribution on R 2 with density λ. B. Graphical inf se the network size. In Fig.4c, we also find that the approximation ratio is insensitive with respect to the path loss ν. Hence, DFMRF scheme has nearly the same efficiency in the entire range of ν ∈ =-=[2,6]-=- under the k-NNG dependency. In Fig.5a, we plot the average energy consumption of DFMRF in (27) under uniform node placement and disk dependency graph with radius δ. The average energy is bounded, as   </text>
<query_num> 17704 </query_num>
<text>   ing laws for multihop wireless networks (without any data fusion) are derived under different routing strategies. The issue of node placement for desirable energy scaling has been considered in [14], =-=[15]-=-, where it is argued that uniform node placement, routinely considered in the literature, has poor energy performance. It is interesting to note that, for fusion networks, uniform sensor distribution   </text>
<query_num> 17705 </query_num>
<text>   sending data from all sensors to the fusion center certainly ensures optimal inference, it is not necessary for statistical inference. More relevant to our work is the idea of data aggregation, e.g., =-=[21]-=-–[23]. Finding aggregation policies for correlated data, however, is nontrivial; it depends on the specific applications for which the sensor network is designed. Perhaps a more precise notion of aggr   </text>
<query_num> 17706 </query_num>
<text>   ted through data fusion. One line of approach is the use of distributed compression with the aim of routing all the measurements to the fusion center. Examples of such approaches can be found in [18]–=-=[20]-=-. While sending data from all sensors to the fusion center certainly ensures optimal inference, it is not necessary for statistical inference. More relevant to our work is the idea of data aggregation   </text>
<query_num> 17707 </query_num>
<text>   th between i and j using the links in the network graph Ng(V) (set of feasible links for direct transmission). We now assume that the network graph Ng(V) is a local uenergy spanner. In the literature =-=[44]-=-, a graph Ng(V) is called a u-energy spanner, for some constant u &amp;gt; 0 called its energy stretch factor, when it satisfies E max i,j∈V SP (i,j;Ng) ESP ≤ u, (22) (i,j;Cg) where Cg(V) denotes the complet y two nodes is no worse than u-times the optimal value. Examples of energy spanners include the Gabriel graph 9 (with stretch factor u = 1 when the path-loss ν ≥ 2), the Yao graph, and its variations =-=[44]-=-. In this paper, we only require a weaker version 10 of the above property that there is at most u-energy stretch between the neighbors in the dependency graph From (23), we have E max (i,j)∈G SP (i,j   </text>
<query_num> 17708 </query_num>
<text>   tions. In this paper, we assume that optimal inference is made at the fusion center, and this places a constraint on data aggregation. It rules out sub-sampling of the sensor field, which is dealt in =-=[5]-=-. B. Summary of results and contributions In this paper, we allow data aggregation at intermediate nodes, but require that the fusion center achieves the same inference performance as if all raw obser imal sensor density for uniform node placement which maximizes the inference error exponent under an average energy constraint, and in [37], [48], we derived the error exponent for MRF hypotheses. In =-=[5]-=-, we analyzed optimal sensor selection (i.e., sub-sampling) policies for achieving tradeoff between fusion costs and inference performance. The energy scaling laws derived in this paper rely heavily o   </text>
<query_num> 17709 </query_num>
<text>   xploited through data fusion. One line of approach is the use of distributed compression with the aim of routing all the measurements to the fusion center. Examples of such approaches can be found in =-=[18]-=-–[20]. While sending data from all sensors to the fusion center certainly ensures optimal inference, it is not necessary for statistical inference. More relevant to our work is the idea of data aggreg   </text>
</top>
<top>
<paper_num> 178 </paper_num>
<paper_title>   A Multiple Hypothesis Approach to Figure Tracking.  </paper_title>
<paper_abstract>   This paper describes a probabilistic multiple-hypothesis framework for tracking highly articulated objects. In this framework, the probability density of the tracker state is represented as a set of modes with piecewise Gaussians characterizing the neighborhood around these modes. The temporal evolution of the probability density is achieved through sampling from the prior distribution, followed by local optimization of the sample positions to obtain updated modes. This method of generating hypotheses from state-space search does not require the use of discrete features unlike classical multiple-hypothesis tracking. The parametric form of the model is suited for highdimensional state-spaces which cannot be efficiently modeled using non-parametric approaches. Results are shown for tracking Fred Astaire in a movie dance sequence. 1  </paper_abstract>
<query_num> 17801 </query_num>
<text>   4]. Yamamoto and Koshikawa [5] were the first to apply modern kinematic models and gradient-based optimization techniques, but their results were limited to 2D motion. Other 3D tracking works include =-=[6, 16, 17, 18]-=-. The work of Ju and et. al. [19] is perhaps the closest to our 2D SPM. Other 2D figure tracking results can be found in [20]. Early applications of Kalman filters (KF) to rigid body tracking appear i   </text>
<query_num> 17802 </query_num>
<text>   The work of Ju and et. al. [19] is perhaps the closest to our 2D SPM. Other 2D figure tracking results can be found in [20]. Early applications of Kalman filters (KF) to rigid body tracking appear in =-=[21, 22, 23]-=-. Figure tracking schemes which use the Kalman filter are discussed in [8, 9]. All of these works employ the conventional unimodal KF. One exception is Shimada et. al. [10], in which a simple multiple   </text>
<query_num> 17803 </query_num>
<text>   inglehypothesis tracker fails to handle the self-occlusion caused by Fred Astaire’s legs crossing. in [24, 25]. An early survey of these techniques can be found in [26]. Recently, Rasmussen and Hager =-=[27]-=- used the joint probabilistic data association filter (JPDAF) [11] to track multi-part objects, such as a face and hand. In contrast to our MHT framework, the JPDAF approach uses a correspondence-base   </text>
<query_num> 17804 </query_num>
<text>   movie dance sequence. 1 Introduction Visual tracking of human motion is a key technology in a large number of areas. It has applications ranging from 3D mouse input [1] to content-based video editing =-=[2]-=-. This paper addresses the visual tracking problem for an articulated object such as the human figure, using a known kinematic model [3, 4, 5, 6]. The kinematics of an articulated object provide the m ion, which is the estimation of 2D image plane figure motion across a video sequence. Figures are described by a novel class of 2D kinematic models called Scaled Prismatic Models (SPM), introduced in =-=[2]-=-. These models enforce 2D constraints on figure motion that are consistent with an underlying 3D kinematic model. Unlike 3D kinematic models, SPM’s do not require detailed prior knowledge of figure ge ty of points along the link due to an instantaneous state change. A complete discussion of SPM models, including a derivation of the SPM Jacobian and an analysis of its singularities, can be found in =-=[2]-=-. In this report we model the figure as a branched SPM chain. Each link in the arms, v p d legs, and head is modeled as an SPM link. Each link has two degrees of freedom, leading to a total body model   </text>
<query_num> 17805 </query_num>
<text>   nd explicitly returns ‘figure features’ where each feature specifies a different skeletal configuration. One alternative is to use Monte Carlo methods such as Isard and Blake’s CONDENSATION algorithm =-=[12]-=-. While nonparametric models can represent arbitrary pdfs, their computational costs are prohibitive for the large state spaces required in figure tracking. This paper describes a novel formulation of n hypotheses between target and features, but state-space hypotheses which locally maximize the likelihood of the observed image. Alternatively Monte Carlo methods, such as the CONDENSATION algorithm =-=[12]-=-, can be used. These methods express the pdf of the tracker state non-parametrically with a fair set of samples. The number of samples required for accurately modeling the pdf increases with both the  ntation of the CONDENSATION algorithm bear out these observations. Tracking was attempted on sequences of a person walking using a 26-dimensional tracker based on templates (instead of contours as in =-=[12]-=-). When a second-order autoregressive (AR) model trained on walking dynamics was applied, tracking was successful whensFigure 4: Mode-based Multiple Hypothesis Tracking Results. Top row: the multiple   </text>
<query_num> 17806 </query_num>
<text>   of any tracking scheme is the choice of probabilistic representation for the state estimates. The Kalman filter [7] is a classical choice which has been employed in earlier figure tracking work (see =-=[8, 9, 10]-=- for examples). Unfortunately the Kalman filter is restricted to representing unimodal probability distributions. The presence of background clutter, self-occlusions, and complex dynamics during figur re tracking results can be found in [20]. Early applications of Kalman filters (KF) to rigid body tracking appear in [21, 22, 23]. Figure tracking schemes which use the Kalman filter are discussed in =-=[8, 9]-=-. All of these works employ the conventional unimodal KF. One exception is Shimada et. al. [10], in which a simple multiple hypothesis approach is used to handle reflective ambiguity under orthographi   </text>
<query_num> 17807 </query_num>
<text>   ranging from 3D mouse input [1] to content-based video editing [2]. This paper addresses the visual tracking problem for an articulated object such as the human figure, using a known kinematic model =-=[3, 4, 5, 6]-=-. The kinematics of an articulated object provide the most fundamental constraint on its motion. Kinematic models play two roles in tracking. First, they define the desired output—a state vector of jo 4]. Yamamoto and Koshikawa [5] were the first to apply modern kinematic models and gradient-based optimization techniques, but their results were limited to 2D motion. Other 3D tracking works include =-=[6, 16, 17, 18]-=-. The work of Ju and et. al. [19] is perhaps the closest to our 2D SPM. Other 2D figure tracking results can be found in [20]. Early applications of Kalman filters (KF) to rigid body tracking appear i   </text>
<query_num> 17808 </query_num>
<text>   t to apply modern kinematic models and gradient-based optimization techniques, but their results were limited to 2D motion. Other 3D tracking works include [6, 16, 17, 18]. The work of Ju and et. al. =-=[19]-=- is perhaps the closest to our 2D SPM. Other 2D figure tracking results can be found in [20]. Early applications of Kalman filters (KF) to rigid body tracking appear in [21, 22, 23]. Figure tracking s   </text>
<query_num> 17809 </query_num>
<text>   ults were limited to 2D motion. Other 3D tracking works include [6, 16, 17, 18]. The work of Ju and et. al. [19] is perhaps the closest to our 2D SPM. Other 2D figure tracking results can be found in =-=[20]-=-. Early applications of Kalman filters (KF) to rigid body tracking appear in [21, 22, 23]. Figure tracking schemes which use the Kalman filter are discussed in [8, 9]. All of these works employ the co   </text>
</top>
<top>
<paper_num> 179 </paper_num>
<paper_title>   Security Goals: Packet Trajectories and Strand Spaces.  </paper_title>
<paper_abstract>   This material was presented in a series of lectures at fosad,  a summer school on Foundations of Security Analysis and Design, at the  University of Bologna Center at Bertinoro in September 2000. It has two  main purposes. The first  </paper_abstract>
<query_num> 17901 </query_num>
<text>   We follow custom and write inv(K) as K -1 , encr(K, m) as {|m|} K , and join(a, b) as a b. If K is a set of keys, K -1 denotes the set of inverses of elements of K. We assume, like many others (e.g. =-=[29, 32, 39]-=-), that A is freely generated, which is crucial for the results in this paper. Axiom 1. A is freely generated from T and K by encr and join. Definition 7. The subterm relation @ is defined inductively   </text>
<query_num> 17902 </query_num>
<text>   ] for an extensive bibliography through the mid-nineties. Indeed, the present chapter has 50 citations, despite not citing large numbers of papers in the literature. For instance, we do not even cite =-=[4]-=-, let alone the hordes of papers derived from it. 23 distributed systems and electronic commerce. Their importance applies equally to military and civil systems. The Dolev-Yao approach of separating c   </text>
<query_num> 17903 </query_num>
<text>   cies of type C-S and E-D. We call a bundle normal if it has no redundancies of type C-S and E-D, by analogy with Prawitz&amp;apos;s normal deductions [42]. Many others have noted related properties, including =-=[7, 40, 20]-=-. We may also assume another property of the bundle C. Definition 14. C is directed if for every node n # C, there is a regular node m # C such that n # C m. Every bundle C is equivalent to some direc   </text>
<query_num> 17904 </query_num>
<text>   ds can be used to discover what packets are expected to emerge from a given interface; by sni#ng the network and sampling the headers, we can raise an alarm if an unexpected packet is found. See also =-=[22]-=-, in which specifications are used to generate test cases systematically. 6. If two given systems each meet a security goal, does each continue to meet that security goal if they are combined in a par   </text>
<query_num> 17905 </query_num>
<text>   es to draw on. Indeed, because protocol mixing has shown itself to be a significant cause of protocol failure, and makes protocol analysis more di#cult [6, 10, 23, 35, 46, 48], it has been identified =-=[36]-=- as a key problem in applying formal methods to cryptographic protocols. Moreover, in practice, di#erent protocols using cryptography are usually combined. A key distribution protocol is useful only i   </text>
<query_num> 17906 </query_num>
<text>   it were, it would be the positive node of a D-strand. And then the preceding key node would have the term K, contrary to the assumption K ## P. 5.4 Neuman-Stubblebine The Neuman-Stubblebine protocol =-=[38]-=- contains two sub-protocols. We will call the first sub-protocol the authentication protocol and the second sub-protocol the re-authentication protocol. In the authentication sub-protocol, a key disA  ly harmless because the contents are public in any 59 case. The secondary protocol would also be unable to re-use symmetric-key tickets such as those generated by the Kerberos Key Distribution Center =-=[24, 38]-=-. These are also intuitively harmless, so long as the secondary protocol does not extract private values from within them, or repackage their private contents, potentially insecurely. Hence, we allow   </text>
<query_num> 17907 </query_num>
<text>   lines contained in the configuration files. One might alternatively dispense with lists of rectangles and instead represent the sets more directly, for instance using Binary Decision Diagrams (bdds) =-=[3]-=-. In our next section, we will instead consider how bdds can be used as an auxiliary structure to discover the set of atoms that naturally describe existing configuration files. 2.12 The Abstraction P  Decision Diagrams Again we face the question how to implement the boolean algebra of sets in this context, and for our current purpose of solving the abstraction problem, the binary decision diagram =-=[3]-=- fits our needs perfectly. A binary decision diagram (bdd) is a finite directed acyclic graph with two sinks, labelled true and false. Each interior node has out-degree two, and is labeled with a prop   </text>
<query_num> 17908 </query_num>
<text>   n with the secondary protocol. One of the advantages of our approach is that the result works for all secrecy and authentication goals; in this it continues a trend visible from several recent papers =-=[31, 21, 45, 44, 19]-=-. We have an additional reason for including this 56 material here: It is a good example of the power of the machinery of paths and well-behaved bundles developed in Section 4. 6.2 Multiprotocol Stran   </text>
<query_num> 17909 </query_num>
<text>   nately, in this protocol, we cannot simply ignore the whitespace. There is another way that things can fit together, if a malicious penetrator is present, shown in Figure 5. The attack is due to Lowe =-=-=-=-[27], and was discovered a decade and a half after the protocol was published. A has initiated a session A {|Na A|}K P # P .  w {|Na A|}KB # B .  w w w w w w # {|Na N b |} KA .  w .  w {|N b |} K P #  n Watergate, is, &amp;quot;What did he know and when did he know it?&amp;quot; To illustrate an authentication goal, let us switch now to a protocol that achieves its goals, such as the Needham-Schroeder-Lowe protocol =-=[-=-27] as shown in Figure 12. The only di#erence from the Needham-Schroeder protocol is in the A {|Na A|}KB # {|Na A|}KB # B .  w # {|Na N b B|}KA # {|Na N b B|}KA .  w .  w {|N b |} KB # {|N b |} KB # . r than some other responder B # . And Figure 5 is a counterexample in which B # = P #= B. Hence we have uncovered a limitation in the authentication achieved by Needham-Schroeder, first noted by Lowe =-=[26, 27]-=-, which led Lowe to amend the protocol to contain the responder&amp;apos;s name B in the second message {|N a N b B|} KA . The Outgoing Test in Needham-Schroeder-Lowe Consider now the corrected Needham-Schroed   </text>
<query_num> 17910 </query_num>
<text>   of the crucial real-world problem of security management. 2 1.2 The Structure of these Lectures We divide the remainder of our report into five sections. Section 2 Packet Trajectories : Derived from =-=[11, 13]-=-. Coauthors: A. Herzog and J. Thayer. Contents : Introduce the packet protection problem. Define a class of security goals that filtering routers can achieve. Network model. Algorithms to determine wh nantly at security gateways, and there are large potential advantages of management in doing so. A systematic way to be sure of reaping those benefits is needed. That material is instead available in =-=[13]-=-. We start by considering the security goals that we would like to achieve. From those, we infer a way of modeling the goals (and the systems that should meet those goals) using simple mathematical no  # B, where B is some suitable collection of sets of packets. In other contexts, other security goals will be needed. For instance, the security goals we can achieve using IPsec, as have discussed in =-=[13]-=-, may take di#erent logical forms or require di#erent real world ingredients to be modeled. 2.5 Some Security Goals We have already mentioned two candidate security goals, each a possible way to preve   </text>
<query_num> 17911 </query_num>
<text>   peers. The penetrator has more unintended services to draw on. Indeed, because protocol mixing has shown itself to be a significant cause of protocol failure, and makes protocol analysis more di#cult =-=[6, 10, 23, 35, 46, 48]-=-, it has been identified [36] as a key problem in applying formal methods to cryptographic protocols. Moreover, in practice, di#erent protocols using cryptography are usually combined. A key distribut   </text>
<query_num> 17912 </query_num>
<text>   peers. The penetrator has more unintended services to draw on. Indeed, because protocol mixing has shown itself to be a significant cause of protocol failure, and makes protocol analysis more di#cult =-=[6, 10, 23, 35, 46, 48]-=-, it has been identified [36] as a key problem in applying formal methods to cryptographic protocols. Moreover, in practice, di#erent protocols using cryptography are usually combined. A key distribut ing this 56 material here: It is a good example of the power of the machinery of paths and well-behaved bundles developed in Section 4. 6.2 Multiprotocol Strand Spaces To represent multiple protocols =-=[46]-=-, we select some regular strands as being runs of the primary protocol; we call these strands primary strands. Definition 16. A multiprotocol strand space is a strand space (#, tr) together with a dis protocol strand space does not have disjoint inbound encryption. Indeed, the penetrator can use a session of the reauthentication protocol to complete a responder strand in a bundle with no initiator =-=[46-=-]. For this reason, we amend (see [46]) the re-authentication protocol to the form shown in Figure 22. To apply our independence theorem, we check that the A # B {|A K T |} KBS .  w N # a {|A K T |} K   </text>
<query_num> 17913 </query_num>
<text>   r than some other responder B # . And Figure 5 is a counterexample in which B # = P #= B. Hence we have uncovered a limitation in the authentication achieved by Needham-Schroeder, first noted by Lowe =-=[26, 27]-=-, which led Lowe to amend the protocol to contain the responder&amp;apos;s name B in the second message {|N a N b B|} KA . The Outgoing Test in Needham-Schroeder-Lowe Consider now the corrected Needham-Schroed   </text>
<query_num> 17914 </query_num>
<text>   sure that di#erent protocols never use the same key, although this may be expensive or di#cult to arrange. Although the Abadi-Needham paper on prudent engineering practice for cryptographic protocols =-=[1]-=- does not discuss mixing di#erent protocols, this rule---to try to achieve disjoint encryption---is in the same spirit as those it proposes. In this section, we will prove that, properly formalized, i   </text>
<query_num> 17915 </query_num>
<text>   ths are especially predictable, and in fact every bundle has an equivalent well-behaved bundle. This section will illustrate the advantages of the strand space model over closely related alternatives =-=[40, 43]-=-, at least as a heuristic for discovering results about cryptographic protocols. 4.1 Bundle Equivalence Definition 10. Bundles C, C # on a strand space # are equivalent i# 1. they have the same regula   </text>
<query_num> 17916 </query_num>
<text>   ths are especially predictable, and in fact every bundle has an equivalent well-behaved bundle. This section will illustrate the advantages of the strand space model over closely related alternatives =-=[40, 43]-=-, at least as a heuristic for discovering results about cryptographic protocols. 4.1 Bundle Equivalence Definition 10. Bundles C, C # on a strand space # are equivalent i# 1. they have the same regula cies of type C-S and E-D. We call a bundle normal if it has no redundancies of type C-S and E-D, by analogy with Prawitz&amp;apos;s normal deductions [42]. Many others have noted related properties, including =-=[7, 40, 20]-=-. We may also assume another property of the bundle C. Definition 14. C is directed if for every node n # C, there is a regular node m # C such that n # C m. Every bundle C is equivalent to some direc   </text>
<query_num> 17917 </query_num>
<text>   will achieve a given goal. Abstraction from router configuration files. Section 3 Strand Spaces and Protocol Security Goals : Material derived from [47]. Coauthors: J. Herzog and J. Thayer. Also from =-=[14, 16]-=-. Coauthor: J. Thayer. Contents : Cryptographic protocols and the Dolev-Yao problem. Why cryptographic protocols fail: Unintended services. Powers of the penetrator. Strand space definitions. Authenti edundancies and redundancy elimination. Paths. The normal form lemma. Rising and falling paths, bridges, e#ciency. Proof of the secrecy theorem. Section 5 Authentication Tests : Material derived from =-=[14, 16]-=-. Coauthor: J. Thayer. Authentication tests: proving authentication and secrecy. Proofs of the authentication test results. Application to examples. Section 6 Protocol Independence via Disjoint Encryp t Common sense suggests a rule of thumb when protocols are to be mixed together. This rule is that if the primary protocol uses a particular form of encrypted message as a test to authenticate a peer =-=[14]-=-, then the secondary protocols should not construct a message of that form. If the primary protocol uses a particular form of encrypted component to protect some private value, then the secondary prot   </text>
</top>
<top>
<paper_num> 180 </paper_num>
<paper_title>   Identification with encrypted biometric data.  </paper_title>
<paper_abstract>   Biometrics make human identification possible with a sample of a biometric trait and an associated database. Classical identification techniques lead to privacy concerns. This paper introduces a new method to identify someone using his biometrics in an encrypted way. Our construction combines Bloom Filters with Storage and Locality-Sensitive Hashing. We apply this error-tolerant scheme, in a Hamming space, to achieve biometric identification in an efficient way. This is the first non-trivial identification scheme dealing with fuzziness and encrypted data. Keywords. Identification, Biometrics, Privacy, Searchable Encryption. 1  </paper_abstract>
<query_num> 18001 </query_num>
<text>   ata” is the full version of our work. † This work was partially supported by the french ANR RNRT project BACH. 11.1 Related Works and Motivation Security of biometric systems is widely studied – cf. =-=[31, 33, 4]-=- – and although a lot of vulnerabilities are now well understood and controlled, it is still difficult to achieve an end-to-end system which satisfies all constraints. In particular, biometric templat   </text>
<query_num> 18002 </query_num>
<text>   d the notion of Public-key encryption with Keyword Search (PEKS) [5], in which specific trapdoors are created for the lookup of keywords over public-key encrypted messages. Several other papers, e.g. =-=[24, 2, 15, 35, 43]-=-, have also elaborated solutions in this field. However the main difference between the search for a keyword as understood by Boneh et al. [5, 6] and biometric matching is that an exact match for a gi   </text>
<query_num> 18003 </query_num>
<text>   eas distant data should remain significantly remote. A noticeable example of a LSH family was proposed by Kushilevitz et al. in [37]; see also [36, 30, 1]. 4.2 Bloom Filters As introduced by Bloom in =-=[3]-=-, a set of Bloom filters is a data structure used for answering set membership queries. Definition 4 Let D be a finite subset of Y . For a collection of ν (independent) hash functions H ′ = {h′ 1, . . nappropriate tags, but the choice of relevant hash functions and increasing their number allow to reduce the probability of that event. These properties are summed up in the following lemma. Lemma 1 (=-=[3]-=-) Let (H ′ , T1, . . .,Tm) be a (ν, m)-Bloom filter with storage indexing a set D with tags from a tag set V . Then, for y ∈ D, the following properties hold: • ψ(y) ⊂ T(y) = ⋂ ν j=1 T h ′ j (y), i.e.   </text>
<query_num> 18004 </query_num>
<text>   iently reduce the cost. The best performances are from Gentry and Ramzan [22] and Lipmaa [38] with a communication complexity polynomial in the logarithm of M. Surveys of the subject are available in =-=[21, 40]-=-. Some PIR protocols are called Symmetric Private Information Retrieval, when they comply with the Data Privacy requirement [23]. This condition states that the querier cannot distinguish between a da er do not get more information that what he asked. Private Information Storage (PIS) Protocols PIR protocols enable to retrieve information of a database. A Private Information Storage (PIS) protocol =-=[40]-=- is a protocol that enables to write information in a database with properties that are similar to that of PIR. The goal is to prevent the database from knowing the content of the information that is   </text>
<query_num> 18005 </query_num>
<text>   j. Among the known constructions of computational secure PIR, block-based PIR – i.e. working on block of bits – allows to efficiently reduce the cost. The best performances are from Gentry and Ramzan =-=[22]-=- and Lipmaa [38] with a communication complexity polynomial in the logarithm of M. Surveys of the subject are available in [21, 40]. Some PIR protocols are called Symmetric Private Information Retriev   </text>
<query_num> 18006 </query_num>
<text>   lowing this, we solely focus in this paper on binary biometric data compared with Hamming distance. Most of protocols involving biometric data and cryptography use Secure Sketches or Fuzzy Extractors =-=[19, 34]-=-. It uses error correction to reduce variations between the different measures, and to somehow hide the biometric data behind a random codeword – e.g. [46, 39, 27, 9, 8, 11]. On the other hand, severa   </text>
<query_num> 18007 </query_num>
<text>   nformation about sensitive data. Our works are also influenced by the problem of finding a match on encrypted data. Boneh et al. defined the notion of Public-key encryption with Keyword Search (PEKS) =-=[5]-=-, in which specific trapdoors are created for the lookup of keywords over public-key encrypted messages. Several other papers, e.g. [24, 2, 15, 35, 43], have also elaborated solutions in this field. H ce a new model for error-tolerant search in Sec. 3 and specific functions to take into account fuzziness in Sec. 4.1. The most significant difference here from the primitives introduced previously in =-=[5]-=- is that messages are no longer associated to keywords. Moreover, our primitives enable some imprecision on the message that is looked up. For example, one can imagine a mailing application, where all   </text>
<query_num> 18008 </query_num>
<text>   phic encryption schemes to compute the Hamming distance between two encrypted templates. Some other interesting solutions based on adaptation of known cryptographic protocols are also investigated in =-=[7, 13]-=-. The drawback with all these techniques is that they do not fit well with identification in large databases as the way to run an identification among N data would be to run almost as many authenticat   </text>
<query_num> 18009 </query_num>
<text>   ryptographic ones) that give the same result for near points, as defined in [30]: Definition 3 ([30]) Let B be a metric space, U a set with a smaller dimensionality, r1, r2 ∈ R with r1 &amp;lt; r2, p1, p2 ∈ =-=[0, 1]-=- with p1 &amp;gt; p2. A family H = {h1, . . . , hµ}, hi : B → U, is (r1, r2, p1, p2)-LSH (Locality-Sensitive Hashing), if for all h ∈ H, x, x ′ ∈ B, Pr[h(x) = h(x ′ )] &amp;gt; p1 (resp. Pr[h(x) = h(x ′ )] &amp;lt; p2) if ccurring between similar data with high probability, whereas distant data should remain significantly remote. A noticeable example of a LSH family was proposed by Kushilevitz et al. in [37]; see also =-=[36, 30, 1]-=-. 4.2 Bloom Filters As introduced by Bloom in [3], a set of Bloom filters is a data structure used for answering set membership queries. Definition 4 Let D be a finite subset of Y . For a collection o   </text>
<query_num> 18010 </query_num>
<text>   th this property gives the lemma. ✷ 4.4 Cryptographic Primitives Public Key Cryptosystem Our construction requires a semantically secure public key cryptosystem – as defined in [25], see for instance =-=[20, 42]-=- – to store some encrypted data in the database. Encryption function is noted Enc and decryption function Dec, the use of the keys is implicit. An encryption scheme is said to be semantically secure (   </text>
<query_num> 18011 </query_num>
<text>   the encryptions of a message x0 and a message x1. Private Information Retrieval Protocols A primitive that enables privacyensuring queries to databases is Private Information Retrieval protocol (PIR, =-=[17]-=-). Its goal is to retrieve a specific information from a remote server in such a way that he does not know which data was sent. This is done through a method Query PIR Y,S (a), that allows Y to recove   </text>
<query_num> 18012 </query_num>
<text>   tions between the different measures, and to somehow hide the biometric data behind a random codeword – e.g. [46, 39, 27, 9, 8, 11]. On the other hand, several biometrics verification protocols, e.g. =-=[14, 10, 12, 44, 47]-=-, have proposed to embed the matching directly. They use the property of homomorphic encryption schemes to compute the Hamming distance between two encrypted templates. Some other interesting solution   </text>
<query_num> 18013 </query_num>
<text>   tions between the different measures, and to somehow hide the biometric data behind a random codeword – e.g. [46, 39, 27, 9, 8, 11]. On the other hand, several biometrics verification protocols, e.g. =-=[14, 10, 12, 44, 47]-=-, have proposed to embed the matching directly. They use the property of homomorphic encryption schemes to compute the Hamming distance between two encrypted templates. Some other interesting solution of the identification, IP has the fresh biometric template b ′ along with the address of the candidate reference templates in DB. To reduce the list of identities, we can use a secure matching scheme =-=[12, 44]-=- to run a final secure comparison between b ′ and the candidates. 6.2 Practical Considerations 6.2.1 Choosing the LSH family: an Example Let’s place ourself in the practical setting of human identific   </text>
<query_num> 18014 </query_num>
<text>   ur in a Hamming space is to use a generic construction called locality-sensitive hashing. It looks for hash functions (not cryptographic ones) that give the same result for near points, as defined in =-=[30]-=-: Definition 3 ([30]) Let B be a metric space, U a set with a smaller dimensionality, r1, r2 ∈ R with r1 &amp;lt; r2, p1, p2 ∈ [0, 1] with p1 &amp;gt; p2. A family H = {h1, . . . , hµ}, hi : B → U, is (r1, r2, p1,  ccurring between similar data with high probability, whereas distant data should remain significantly remote. A noticeable example of a LSH family was proposed by Kushilevitz et al. in [37]; see also =-=[36, 30, 1]-=-. 4.2 Bloom Filters As introduced by Bloom in [3], a set of Bloom filters is a data structure used for answering set membership queries. Definition 4 Let D be a finite subset of Y . For a collection o   </text>
<query_num> 18015 </query_num>
<text>   use Secure Sketches or Fuzzy Extractors [19, 34]. It uses error correction to reduce variations between the different measures, and to somehow hide the biometric data behind a random codeword – e.g. =-=[46, 39, 27, 9, 8, 11]-=-. On the other hand, several biometrics verification protocols, e.g. [14, 10, 12, 44, 47], have proposed to embed the matching directly. They use the property of homomorphic encryption schemes to comp   </text>
<query_num> 18016 </query_num>
<text>   wn constructions of computational secure PIR, block-based PIR – i.e. working on block of bits – allows to efficiently reduce the cost. The best performances are from Gentry and Ramzan [22] and Lipmaa =-=[38]-=- with a communication complexity polynomial in the logarithm of M. Surveys of the subject are available in [21, 40]. Some PIR protocols are called Symmetric Private Information Retrieval, when they co . The PIR query complexity at the sensor level depends on the scheme used (recall that the PIR query is made only over the set of buckets and not over the whole database); in the case of Lipmaa’s PIR =-=[38]-=-, this cost κ(PIR) is dominated by the cost of a Damg˚ard-Jurik encryption. The overall sensor complexity of an identification request is O(µν(|T |κ(Dec) + κ(PIR))).7 Conclusion This paper details th   </text>
</top>
<top>
<paper_num> 181 </paper_num>
<paper_title>   Information extraction challenges in managing unstructured data.  </paper_title>
<paper_abstract>   Over the past few years, we have been trying to build an end-to-end system at Wisconsin to manage unstructured data, using extraction, integration, and user interaction. This paper describes the key information extraction (IE) challenges that we have run into, and sketches our solutions. We discuss in particular developing a declarative IE language, optimizing for this language, generating IE provenance, incorporating user feedback into the IE process, developing a novel wikibased user interface for feedback, best-effort IE, pushing IE into RDBMSs, and more. Our work suggests that IE in managing unstructured data can open up many interesting research challenges, and that these challenges can greatly benefit from the wealth of work on managing structured data that has been carried out by the database community. 1.  </paper_abstract>
<query_num> 18101 </query_num>
<text>   e IE languages (e.g., UIMA atresearch.ibm.com/UIMA), xlog builds on the well-founded semantics of Datalog. As such, it can naturally and rigorously handle recursion (which occurs quite commonly in IE =-=[1, 2]-=-). Finally, it can also leverage the wealth of execution and optimization techniques already developed for Datalog. Much work remains, however, as our current xlog version is still rudimentary. We are   </text>
<query_num> 18102 </query_num>
<text>   em, then apply it to a broad variety of applications, including community information management [13], personal information management [3], besteffort/on-the-fly data integration [17], and dataspaces =-=[14]-=- (see www.cs.wisc.edu/~anhai/projects/cimple for more detail on the Cimple project). The rest of this paper is organized as follows. In Sections 2-4 we describe key IE challenges in developing IE prog   </text>
<query_num> 18103 </query_num>
<text>   ion integration, and user interaction. In this paper we briefly describe the key challenges in information extraction (IE) that we have faced, sketch our solutions, and discuss future directions (see =-=[11, 10]-=- for a discussion of non-IE challenges). Our work suggests . that managing unstructured data can open up many interesting IE directions for database researchers. It further suggests that these directi   </text>
<query_num> 18104 </query_num>
<text>   nd the latest community information. In such contexts, applying IE to each corpus snapshot in isolation, from the scratch, as typically done today, is very time consuming. To address this problem, in =-=[5]-=- we have developed a set of techniques to efficiently execute an xlog program over an evolving text corpus. The key idea underlying our solution is to recycle previous IE results, given that consecuti imple project will meet in room CS 105 at 3pm”, from which we have extracted “CS 105” as a room number. Then when we see the above text fragment again in a new snapshot, under certain conditions (see =-=[5]-=-) we can immediately conclude that “CS 105” is a room number, without re-applying the IE program to the text fragment. Overall, our work has suggested that xlog is highly SIGMOD Record, December 2008   </text>
<query_num> 18105 </query_num>
<text>   of the Cimple project. Cimple started out trying to build community information management systems: those that manage data for online communities, using extraction, integration, and user interaction =-=[13]-=-. Over time, however, it became clear that such systems can be used to manage unstructured data in many contexts beyond just online communities. Hence, Cimple now seeks to build such a general-purpose   </text>
<query_num> 18106 </query_num>
<text>   proceedings), then the second plan can significantly outperform the first one. One way to address this choice of plans is to perform cost-based optimization, like in relational query optimization. In =-=[18]-=- we have developed such a cost-based optimizer. Given an xlog program P, the optimizer conceptually generates an execution plan for P, employs a set of rewriting rules (such as pushing down a selectio  promising plan candidates, then selects the candidate with the lowest estimated cost, where the costs are estimated using a cost model (in the same spirit as relational query optimization). The work =-=[18]-=- describes the optimizer in detail, including techniques to efficiently search for the best candidate in the often huge candidate space. Optimizing for Evolving Data: So far we have considered only st  their IE without the use of an RDBMS. A very common method, for example, is to store text data in files, write the IE program as a script, or in a recently developed declarative language (e.g., xlog =-=[18]-=-, AQL of System-T [16], UIMA atresearch.ibm.com/UIMA), then execute this program over these text files, using the file system for all storage. This method indeed offers a good start. But given that IE   </text>
<query_num> 18107 </query_num>
<text>   structured data, and that RDBMSs have had a 30-year history of managing structured data, a natural question arises: Do RDBMSs offer any advantage over file systems for IE applications? In recent work =-=[6, 19]-=-, we have explored this question, provided an affirmative answer, and further explored the natural follow-on questions of How can we best exploit current RDBMS technology to support IE? and How can cu   </text>
<query_num> 18108 </query_num>
<text>   structured data, such as text, Web pages, emails, blogs, and memos, is becoming increasingly pervasive. Hence, it is important that we develop solutions to manage such data. In a recent CIDR-09 paper =-=[12]-=- we have outlined an approach to such a solution. Specifically, we propose building unstructured data management systems (UDMSs). Such systems extract structures (e.g., person names, locations) from t   </text>
<query_num> 18109 </query_num>
<text>   sts “asking for the provenance of a non-answer”. Such nonanswer provenance is important because it can provide more confidence in the answer for the user, and can help developers debug the system. In =-=[15]-=- we have developed an initial approach to providing the provenance of non-answers. In the above example, for instance, our solution can explain that no tuple with talk-title =“Declarative IE”and talk-   </text>
<query_num> 18110 </query_num>
<text>   to build such a general-purpose unstructured data management system, then apply it to a broad variety of applications, including community information management [13], personal information management =-=[3]-=-, besteffort/on-the-fly data integration [17], and dataspaces [14] (see www.cs.wisc.edu/~anhai/projects/cimple for more detail on the Cimple project). The rest of this paper is organized as follows. I   </text>
<query_num> 18111 </query_num>
<text>   up to date, we often must apply an xlog program repeatedly, to consecutive corpus snapshots. Consider, for example, DBLife, a structured portal for the database community that we have been developing =-=[8, 9]-=-. DBLife operates over a text corpus of 10,000+ URLs. Each day it recrawls these URLs to generate a 120+ MB corpus snapshot, and then applies an IE program to this snapshot to find the latest communit   </text>
<query_num> 18112 </query_num>
<text>   use of an RDBMS. A very common method, for example, is to store text data in files, write the IE program as a script, or in a recently developed declarative language (e.g., xlog [18], AQL of System-T =-=[16]-=-, UIMA atresearch.ibm.com/UIMA), then execute this program over these text files, using the file system for all storage. This method indeed offers a good start. But given that IE programs fundamentall   </text>
<query_num> 18113 </query_num>
<text>   use then a user is more likely to find an UI that he or she is comfortable with, and thus is more likely to participate in the interaction. Toward this goal, we have recently developed a wikibased UI =-=[7]-=- (based on the observation that many users increasingly use wikis to collect and correct data). This UI exposes the data to be corrected in a set of wiki pages. Users examine and correct these pages,  can transform P into P ′ )?. Yet another challenge is that once the system has found the intended edit sequence, how can it efficiently propagate this sequence to the underlying data? Our recent work =-=[7]-=- discusses these challenges in detail and proposes initial solutions. Develop Novel Modes of User Interaction: So far we have discussed the following mode of user interaction for UDMSs: a developer U   </text>
<query_num> 18114 </query_num>
<text>   veloping novel interaction modes. We now briefly explain these directions. Generating the Provenance of Query Result: Much work has addressed the problem of generating the provenance of query results =-=[20]-=-. But this work has focused only on positive provenance: it seeks to explain why an answer is produced. In many cases, however, a user may be interested in negative provenance, i.e., why a certain ans   </text>
</top>
<top>
<paper_num> 182 </paper_num>
<paper_title>   Spatio-temporal Reflectance Sharing for Relightable 3D Video.  </paper_title>
<paper_abstract>   Abstract. In our previous work [21], we have shown that by means of a model-based approach, relightable free-viewpoint videos of human actors can be reconstructed from only a handful of multi-view video streams recorded under calibrated illumination. To achieve this purpose, we employ a marker-free motion capture approach to measure dynamic human scene geometry. Reflectance samples for each surface point are captured by exploiting the fact that, due to the person’s motion, each surface location is, over time, exposed to the acquisition sensors under varying orientations. Although this is the first setup of its kind to measure surface reflectance from footage of arbitrary human performances, our approach may lead to a biased sampling of surface reflectance since each surface point is only seen under a limited number of half-vector directions. We thus propose in this paper a novel algorithm that reduces the bias in BRDF estimates of a single surface point by cleverly taking into account reflectance samples from other surface locations made of similar material. We demonstrate the improvements achieved with this spatio-temporal reflectance sharing approach both visually and quantitatively. 1  </paper_abstract>
<query_num> 18201 </query_num>
<text>   , namely of human actors. By means of a marker-free optical motion capture algorithm, it becomes possible to measure both the shape and the motion of a person from multiple synchronized video streams =-=[2]-=-. If the video footage has, in addition, been captured under calibrated lighting conditions, the video frames showing the moving person not only represent texture samples, but actually reflectance sam ed approach to free-viewpoint video of human actors that jointly employs a marker-free motion capture method and a dynamic multi-view texture generation approach to produce novel viewpoint renditions =-=[2, 20]-=-. Unfortunately, none of the aforementioned methods can correctly reproduce 3D video appearance under novel simulated lighting conditions. Only few papers have been published so far that aim at religh view video footage without having to resort to optical markers in the scene. It employs a template human body model consisting of a kinematic skeleton and a single-skin triangle mesh surface geometry =-=[2, 20]-=-. In an initialization step, the shape and proportions of the template are matched to the recorded silhouettes of the actor. After shape initialization, the model is made to follow the motion of the a   </text>
<query_num> 18202 </query_num>
<text>   been developed in recent years. A popular category of algorithms employs the shape-from-silhouette principle to reconstruct dynamic scene geometry by means of voxels, polyhedrals or point primitives =-=[14, 24, 13, 6, 12]-=-. By finding temporal correspondences between pertime-step reconstructions, it becomes feasible to generate novel animations asswell [19]. Another category of approaches reconstructs dynamic scene geo   </text>
<query_num> 18203 </query_num>
<text>   d the light stage device such that it enables capturing of dynamic reflectance fields. Their results are impressive, however it is not possible to change the viewpoint in the scene. Einarsson et. al. =-=[3]-=- extends it further by using a large light stage, a trade-mill where the person walks on, and light field rendering for display. Eventually, human performances can be rendered from novel perspectives   </text>
<query_num> 18204 </query_num>
<text>   eans of voxels, polyhedrals or point primitives [14, 24, 13, 6, 12]. By finding temporal correspondences between pertime-step reconstructions, it becomes feasible to generate novel animations asswell =-=[19]-=-. Another category of approaches reconstructs dynamic scene geometry by means of multi-view stereo [27, 9, 22]. In any case, time-varying textures for rendering are assembled from the input video stre   </text>
<query_num> 18205 </query_num>
<text>   he estimation of reflectance properties of static scenes from images that are captured under calibrated setups of light source and camera. Typically, parameters of a BRDF model are fitted to the data =-=[18, 11]-=- or appearance under novel lighting conditions is created via interpolation between the images themselves [15]. A combination of reflectance estimation and shape-from-shading to refine the geometry of   </text>
<query_num> 18206 </query_num>
<text>   ing conditions is created via interpolation between the images themselves [15]. A combination of reflectance estimation and shape-from-shading to refine the geometry of static scenes is also feasible =-=[25, 17, 1, 4, 5]-=-. In an independent line of research, many methods to capture and render 3D videos of real-world scenes have been developed in recent years. A popular category of algorithms employs the shape-from-sil   </text>
<query_num> 18207 </query_num>
<text>   ition for the subsequent dynamic reflectometry procedure. 3.3 Dynamic Reflectometry Our dynamic reflectance model consists of two components, a static parametric isotropic BRDF for each surface point =-=[16, 10]-=-, as well as a description of the timevarying direction of the normal at each surface location. The first component of the reflectance model is reconstructed from the video frames of the reflectance e   </text>
<query_num> 18208 </query_num>
<text>   light source and camera. Typically, parameters of a BRDF model are fitted to the data [18, 11] or appearance under novel lighting conditions is created via interpolation between the images themselves =-=[15]-=-. A combination of reflectance estimation and shape-from-shading to refine the geometry of static scenes is also feasible [25, 17, 1, 4, 5]. In an independent line of research, many methods to capture   </text>
<query_num> 18209 </query_num>
<text>   s between pertime-step reconstructions, it becomes feasible to generate novel animations asswell [19]. Another category of approaches reconstructs dynamic scene geometry by means of multi-view stereo =-=[27, 9, 22]-=-. In any case, time-varying textures for rendering are assembled from the input video streams. In contrast, the authors in their previous work have proposed a model-based approach to free-viewpoint vi   </text>
</top>
<top>
<paper_num> 183 </paper_num>
<paper_title>   Dynamic Policy-Driven Quality of Service in Service-Oriented Systems.  </paper_title>
<paper_abstract>   Service-oriented architecture (SOA) middleware has emerged as a powerful and popular distributed computing paradigm due to its high-level abstractions for composing systems and hiding platform-level details. Control of some details hidden by SOA middleware is necessary, however, to provide managed quality of service (QoS) for SOA systems that need predictable performance and behavior. This paper presents a policy-driven approach for managing QoS in SOA systems. We discuss the design of several key QoS services and empirically evaluate their ability to provide QoS under CPU overload and bandwidth-constrained situations.  </paper_abstract>
<query_num> 18301 </query_num>
<text>   ation remote communications by adding middleware modules at the OS kernel space and dynamically reserve network resources to provide network QoS for the application remote invocations.Schantz et al. =-=[16]-=- intercept remote communications using middleware proxies and provide network QoS for remote communications by using both DiffServ and IntServ network QoS mechanisms. Yemini et al. [18] provide middle   </text>
<query_num> 18302 </query_num>
<text>   eware. In addition, [2] extends EJB containers to integrate QoS features by providing negotiation interfaces which the application developers need to implement to receive desired QoS support. Synergy =-=[14]-=- describes a distributed stream processing middleware that provides QoS to data streams in real time by efficient reuse of data streams and processing components. [13] presents an algorithm for compos   </text>
<query_num> 18303 </query_num>
<text>   ynamic control and responsiveness to the policy infrastructure. V. RELATED WORK QoS management in middleware and SOA. Prior work focused on adding various QoS capabilities to middleware. For example, =-=[8]-=- describes J2EE container resource management mechanisms that provide CPU availability assurances to applications. Likewise, 2K [19] provides QoS to applications from varied domains using a component-   </text>
</top>
<top>
<paper_num> 184 </paper_num>
<paper_title>   A taxonomy of multicast data origin authentication: Issues and solutions.  </paper_title>
<paper_abstract>   this article we review and classify recent works  dealing with the data origin authentication problem in group communication, and we discuss  and compare them with respect to some relevant performance criteria  </paper_abstract>
<query_num> 18401 </query_num>
<text>   . layer such as TCP and hence packet loss is not a concern for data origin authentication in this kind of application. An example of applications that can be modeled as streaming is executing applets =-=[44]-=-. Applets are organized into modules that are downloaded and executed module by module. In this case instant data origin authentication is required to avoid an attacker sending malicious code, and non ing malicious code. In the following sections we present protocols that fit into this sub-category. Simple Off-line Chaining Overview: The main idea of the solution proposed by Gennaro and Rohatgi in =-=[44, 45]-=- is to divide the stream into blocks and embed some authentication information in the stream itself. The authentication information of the ith block is used to authenticate the (i + 1)st block. In thi onstruct the chain (get-all-before) is too strong of a requirement for practical Internet applications. On-line Signing Overview: The essence of the second solution proposed by Gennaro and Rohatgi in =-=[44, 45]-=- is to make each data block carrying the one-time public key used to one-time sign the following block. Hence, by signing the first one-time public key, the signature is amortized over the stream. The f the packets can be dropped between the sender and the receiver. It has been proven that a packet P i is verifiable if it remains a path (following the hash-links) from P i to a signature packet S j =-=[44]-=-. We designate by verification ratio the number of verifiable packets by the number of received packets. The verification ratio is a good indicator of the verification probability, which means the pro   </text>
<query_num> 18402 </query_num>
<text>   ate and distribute polynomial shares for each message for each receiver. However, this is the price to pay for unconditional security. E x ten s io n s a n d Im p ro vem en ts : Safavi-Naini and Wang =-=[33, 34]-=- generalized the Desmedt et al. polynomial scheme in such a way that instead of a single message, each polynomial can be used to authenticate multiple messages. The authors then proposed to construct   </text>
<query_num> 18403 </query_num>
<text>   ch are efficient to generate and verify. The authors propose to use MACs with a single bit as output. Hence, the authentication information is reduced to l bits. Boneh, Durfee, and Franklin showed in =-=[36]-=- that the Canetti et al. construction has optimal length (up to a small constant factor) for an authenticator that is based purely on pseudo random functions. Each packet carries its authentication in   </text>
<query_num> 18404 </query_num>
<text>   de-off between the authentication information overhead (bandwidth) and tolerance to packet loss (parameter m). However, this scheme requires a high computation over5 IDA has been proposed by Rabin in =-=[56]-=-. It was originally developed to provide safe and reliable storage or transmission of information in distributed systems. The basic idea of IDA is to process the original data denoted by F, by introdu   </text>
<query_num> 18405 </query_num>
<text>   e amounts of data. However, it is possible to use public-key encryption to send a symmetric key, which can then be used to encrypt additional data. An example of asymmetric encryption systems is R SA =-=[12]-=-. NON-REPUDIATION WITH PROOF OF ORIGIN Receiver Message Encryption algorithm Encryption speed (MBytes/s) DES 22.19 IDEA 17.65 AES-128 bits 62.04 Table 3. Computation speed of some encryption algorithm ). If the signature is valid then the message as well as its origin are authentic and non-repudiation is guaranteed. Otherwise, the message is rejected. Examples of digital signature schemes are R SA =-=[12]-=- and DSA [11]. Table 4 lists measurements for commonly used digital signature systems. Certification: To verify a signature, a receiver needs to be assured that the public key used in verifying a sign   </text>
<query_num> 18406 </query_num>
<text>   hat Powerballs is worth almost another ball; that is, using x Powerballs is almost as secure as requiring x + 1 balls to fall into a bin using the original BiBa scheme. Leonid Reyzin and Natan Reyzin =-=[29]-=- also proposed a one-time signature scheme based on one-way hash functions that is as fast as BiBa in verifying. Signing is even faster than verifying and the key and signature sizes are slightly impr   </text>
<query_num> 18407 </query_num>
<text>   reduce the siz e of the authentication information share carried by each packet, and on the other hand to increase the tolerated packet loss threshold. 4 This technique is proposed by Wong and Lam in =-=[52, 53]-=-. (4) (5) Signature verification V PK (d1|d2|d3, S) d1 d2 d3 H H H P1 d2 d3 IEEE Communications Surveys &amp; Tutorials • Third Quarter 2004 S PKs� ��� In the following sections we review proposed solutio dancy of the authentication information by means of replication, or on an implicit redundancy of the authentication information by means of some extra-processing. Tree-Chaining Overview: Wong and Lam =-=[52, 53]-=- have proposed that each packet carries the required authentication information so that it can be individually verifiable. In other words, even if n – 1 out of n packets are lost, the authenticity of   </text>
<query_num> 18408 </query_num>
<text>   th signing and verifying. SAIDA: Signature Amortization using IDA Overview: In order to reduce the size of the authentication information carried by each packet, the approach proposed by P ark et al. =-=[54, 55]-=- uses a famous mechanism called ID A 5 to disperse the n hashes of n block packets, as well as the block signature into n pieces in such a way that the n pieces (and hence the n hashes as well as the  hemes try to find the best trade-off between robustness and required resources and hence have a probabilistic tolerance to packet loss that depends on the packet loss rate in the network. Park et al. =-=[55]-=- showed by using simulations that EMSS, SAIDA, piggybacking, and augmented-chain schemes have a verification rate that decreases almost linearly with respect to packet loss in the network. Namely, whe diation protocols makes it difficult to cope with latencies at both the sender and receivers at the same time. Indeed, 7 These results depend on the settings of the simulations carried by the authors =-=[55]-=-.. IEEE Communications Surveys &amp; Tutorials • Third Quarter 2004 53sexcept for the on-line/off-line and the on-line signing schemes, the other protocols require either that packets are buffered at the   </text>
<query_num> 18409 </query_num>
<text>   the drawbacks of both of these approaches by mixing their underlying asymmetry mechanisms. The BiBa One-Time Signature and Broadcast Authentication Protocol Overview: The main idea proposed by Perrig =-=[28]-=- is that the sender uses a set of keys to authenticate messages. When the sender authenticates a message it discloses only a subset of keys that allow the verification of the packet’s authentication i r some of them are lost. This complies with the time asymmetry approach. Thus, the proposed scheme is a hybrid of the two approaches presented above, eliminating the drawbacks of both of them. Perrig =-=[28]-=- achieved this hybrid asymmetry using a new one-time signature scheme called Bins and Balls (BiBa). In the following section we present the BiBa one-time signature scheme, then show how it is extended may have only this small number of balls that form the signature and hence there is a low probability that a signature for a new message can be forged using this small subset of balls. Indeed, Perrig =-=[28]-=- showed that if the number of thrown balls decreases, then the G h probability that two balls fall into the same bin decreases exponentially. In the following sections we consider balls with the prope   </text>
</top>
<top>
<paper_num> 185 </paper_num>
<paper_title>   Learning to Order Terms: Supervised Interestingness Measures in Terminology Extraction  </paper_title>
<paper_abstract>   Abstract — Term Extraction, a key data preparation step in Text Mining, extracts the terms, i.e. relevant collocation of words, attached to specific concepts (e.g. genetic-algorithms and decisiontrees are terms associated to the concept “Machine Learning ”). In this paper, the task of extracting interesting collocations is achieved through a supervised learning algorithm, exploiting a few collocations manually labelled as interesting/not interesting. From these examples, the ROGER algorithm learns a numerical function, inducing some ranking on the collocations. This ranking is optimized using genetic algorithms, maximizing the trade-off between the false positive and true positive rates (Area Under the ROC curve). This approach uses a particular representation for the word collocations, namely the vector of values corresponding to the standard statistical interestingness measures attached to this collocation. As this representation is general (over corpora and natural languages), generality tests were performed by experimenting the ranking function learned from an English corpus in Biology, onto a French corpus of Curriculum Vitae, and vice versa, showing a good robustness of the approaches compared to the state-of-the-art Support  </paper_abstract>
<query_num> 18501 </query_num>
<text>   0,1), with no false positive and 100% true positive examples. ROC curve has no sensitivities to the ratio of positives and negatives examples [15] as opposed to other accuracy measures such as Fscore =-=[4]-=-. One advantage of ROC curves is to naturally accommodate ill-balanced distributions and costsensitive learning [8]. The area under the ROC curve (AUC) is thus viewed as a global measure of the learni   </text>
<query_num> 18502 </query_num>
<text>   4] � Loglikelihood (L) [9] � Number of occurrences + Loglikelihood (OccL) 1 [18] The choice of an interestingness measure, mostly tackled in the literature through statistical and linguistic criteria =-=[7,17,28]-=- is currently viewed as a decision making problem. Another approach based on learning an interestingness measure, is proposed by Vivaldi et al. [26]. They represent collocations from the values of the   </text>
<query_num> 18503 </query_num>
<text>   UC value along each run. Although these hypotheses cannot be considered truly independent as they are optimized on the same training set, it makes sense to consider their combination [3]. As shown in =-=[10]-=-, the averaging of randomized hypotheses can exponentially amplify their advantage over the default accuracy. Formally, let h1,…,hT denote the T normalized 2 hypotheses constructed along T independent   </text>
<query_num> 18504 </query_num>
<text>   ctives for further research. II. TERMS EXTRACTION MEASURES Different statistical criteria are used in systems of terminology extraction, for instance ACABIT [8] uses loglikelihood measure [9] and KEA =-=[27]-=- uses TF x IDF measure. The statistical criteria (value of the measures and the rank of each collocation) used in our approach are: � Mutual Information (MI) [5] � Mutual Information with cube (MI 3 )   </text>
<query_num> 18505 </query_num>
<text>   d in our approach are: � Mutual Information (MI) [5] � Mutual Information with cube (MI 3 ) [7] � Dice Coefficient (Dice) [24] � Loglikelihood (L) [9] � Number of occurrences + Loglikelihood (OccL) 1 =-=[18]-=- The choice of an interestingness measure, mostly tackled in the literature through statistical and linguistic criteria [7,17,28] is currently viewed as a decision making problem. Another approach bas at the proportion of interesting terms is very high in both datasets, which is why we could ask an expert to label them. The situation is entirely different when rare collocations are considered (see =-=[18]-=-). TABLE I : MOLECULAR BIOLOGY Frequent collocations # collocations relevant irrelevant Biology 1028 90.9% 9.1% CV 376 85.7% 14.3% B. Comparative validations Table II shows the predictive accuracy of   </text>
<query_num> 18506 </query_num>
<text>   he known difficulties of data mining, text mining presents specific difficulties due to the structure of documents and natural language. In particular, the construction of ontologies or terminologies =-=[2,16]-=- which is a central task in text mining, aims at controlling the polysemy and synonymy of words by structuring the words and their meanings in the application domain. A preliminary step for ontology c   </text>
<query_num> 18507 </query_num>
<text>   its recall-precision tradeoff, measured with respect to its Receiver Operating Characteristics (ROC) curve. Accordingly, a ranking function is learned by optimizing the area under the ROC curve (AUC) =-=[11,14]-=- from a few words collocations labelled as relevant/irrelevant by an expert. The paper is organised as follows. Section II briefly reviews the main criteria used in terms extraction. Section III prese n of AUC constitutes a NP-complete problem, which has been undertaken in the literature in a number of ways, from evolutionary programming of neural nets [12] to greedy optimization of decision trees =-=[11]-=-. Recently, this problem was turned into a differentiable International Journal of Computational Intelligence Volume 1 Number 2 99 optimization problem by encapsulating the comparison of any two examp   </text>
<query_num> 18508 </query_num>
<text>   ng problem. Another approach based on learning an interestingness measure, is proposed by Vivaldi et al. [26]. They represent collocations from the values of the statistical criteria and use Adaboost =-=[20]-=- to automatically construct a discriminant hypothesis. The presented work follows [26] with two main differences 1 OccL is defined by ranking terms according to their number of occurrences, and breaki   </text>
<query_num> 18509 </query_num>
<text>   of ranking an interesting collocation below a non-interesting one constitutes an appropriate evaluation for an interestingness measure. The bias and variance of the AUC criterion have been studied by =-=[19]-=- and compared to the criteria of the misclassification error. An analytical and empirical study suggests that though the AUC bias might be higher than for the misclassification cost, its variance is l   </text>
<query_num> 18510 </query_num>
<text>   olysemy and synonymy of words by structuring the words and their meanings in the application domain. A preliminary step for ontology construction is to extract the domain terms, or words collocations =-=[2,16,23]-=-. Terms extraction involves two tasks: detecting “interesting” collocation of words (terms) and classifying them according to classes predefined by an expert. This paper focuses on the detection of in   </text>
<query_num> 18511 </query_num>
<text>   on defining a ranking criteria on the words collocations. Based on [13], this paper formalizes an interestingness measure as a solution of some supervised learning problem (Learning to Order Things, =-=[6]-=-), or LRI, Univ. Paris XI, 91405 Orsay Cedex, France. {aze,roche,yk,sebag}@lri.fr International Journal of Computational Intelligence Volume 1 Number 2 98 optimization problem. Actually, an interestin   </text>
<query_num> 18512 </query_num>
<text>   re through statistical and linguistic criteria [7,17,28] is currently viewed as a decision making problem. Another approach based on learning an interestingness measure, is proposed by Vivaldi et al. =-=[26]-=-. They represent collocations from the values of the statistical criteria and use Adaboost [20] to automatically construct a discriminant hypothesis. The presented work follows [26] with two main diff   </text>
<query_num> 18513 </query_num>
<text>   rpus to another. The paper ends with perspectives for further research. II. TERMS EXTRACTION MEASURES Different statistical criteria are used in systems of terminology extraction, for instance ACABIT =-=[8]-=- uses loglikelihood measure [9] and KEA [27] uses TF x IDF measure. The statistical criteria (value of the measures and the rank of each collocation) used in our approach are: � Mutual Information (MI ves and negatives examples [15] as opposed to other accuracy measures such as Fscore [4]. One advantage of ROC curves is to naturally accommodate ill-balanced distributions and costsensitive learning =-=[8]-=-. The area under the ROC curve (AUC) is thus viewed as a global measure of the learning efficiency. As noted by [14], the area under the ROC curve is equivalent to the Wilcoxon rank statistics, the pr   </text>
<query_num> 18514 </query_num>
<text>   tistical criteria (value of the measures and the rank of each collocation) used in our approach are: � Mutual Information (MI) [5] � Mutual Information with cube (MI 3 ) [7] � Dice Coefficient (Dice) =-=[24]-=- � Loglikelihood (L) [9] � Number of occurrences + Loglikelihood (OccL) 1 [18] The choice of an interestingness measure, mostly tackled in the literature through statistical and linguistic criteria [7   </text>
<query_num> 18515 </query_num>
<text>   ying them according to classes predefined by an expert. This paper focuses on the detection of interesting terms, and more precisely on defining a ranking criteria on the words collocations. Based on =-=[13]-=-, this paper formalizes an interestingness measure as a solution of some supervised learning problem (Learning to Order Things, [6]), or LRI, Univ. Paris XI, 91405 Orsay Cedex, France. {aze,roche,yk,s ber of occurrences, and breaking the ties based on the term likelihoods.si) measures (Dice, OccL) are added to the description of collocations ; ii) the learning problem is one of preference learning =-=[13]-=- instead of discriminant learning. III. OVERVIEW A. Linear ranking function The ROGER algorithm [21,22] tackling the AUC optimization using evolution strategies, is among the most efficient evolutiona   </text>
</top>
<top>
<paper_num> 186 </paper_num>
<paper_title>   Scenes: Abstracting interaction in immersive sensor networks.  </paper_title>
<paper_abstract>   Pervasive computing deployments are increasingly using sensor networks to build instrumented environments that provide local data to immersed mobile applications immersed. These applications demand opportunistic and unpredictable interactions with local devices. While this direct communication has the potential to reduce both overhead and latency, it deviates significantly from existing uses of sensor networks that funnel information to a static central collection point. This pervasive computing driven perspective demands new communication abstractions that enable the required direct communication among mobile applications and embedded sensors. This paper presents the scene abstraction, which allows immersed applications to create dynamic distributed data structures over the immersive sensor network. A scene is created based on application requirements, properties of the underlying network, and properties of the physical environment. This paper details our work on defining scenes, providing an abstract model, an implementation, and an evaluation.  </paper_abstract>
<query_num> 18601 </query_num>
<text>   BasicScene, which provides a prototype of the protocol’s functionality. Other communication styles can be swapped in for the BasicScene (for example one built around TinyDB [21] or directed diffusion =-=[22]-=-). By defining the SceneStrategy interface, we enable developers who are experts in existing communication approaches to create simple plug-ins that use different query communication protocols and yet adigms differ significantly from approaches for even mobile networks in that they include constructs to directly address resource constraints and to enable cooperation among nodes. Directed diffusion =-=[22]-=-, for example provides a decentralized data-centric communication protocol that allows nodes in the network to cooperate to aggregate data as it is funneled back to the requester. However, in directed   </text>
<query_num> 18602 </query_num>
<text>   behavior by all sensors all the time. A few constructs have also begun to address mobility. Mobicast [15] pushes messages to nodes that fall in a dynamic region in front of a moving target. MobiQuery =-=[16]-=- allows a query area to respond to a user’s announced motion profile. These approaches require nodes to have a fine-grained knowledge of their physical locations, which is not reasonable in future per   </text>
<query_num> 18603 </query_num>
<text>   ese three approaches do not directly consider the dynamics of mobility, and they require proactive behavior by all sensors all the time. A few constructs have also begun to address mobility. Mobicast =-=[15]-=- pushes messages to nodes that fall in a dynamic region in front of a moving target. MobiQuery [16] allows a query area to respond to a user’s announced motion profile. These approaches require nodes   </text>
<query_num> 18604 </query_num>
<text>   ion is independent of the particular hardware used to support it, in our initial implementation, these software components have been developed for Crossbow Mica2 motes [23] and are written for TinyOS =-=[24]-=- in the nesC language [25]. Our nesC implementation of the scene abstraction (along with other project information) is availScene Receive StdControl SceneM SendMsg StdControl ContextQuery QueuedSend S   </text>
<query_num> 18605 </query_num>
<text>   ly, future work will investigate the question of how rich and usable the abstraction is with respect to the requirements of a variety of applications. Our own work with intelligent construction sites =-=[29]-=- has demonstrated applicability to a second domain (other than the first responder domain used in this paper), but we plan to perform additional user and application studies to further validate our ex   </text>
<query_num> 18606 </query_num>
<text>   of the SceneStrategy, the BasicScene, which provides a prototype of the protocol’s functionality. Other communication styles can be swapped in for the BasicScene (for example one built around TinyDB =-=[21]-=- or directed diffusion [22]). By defining the SceneStrategy interface, we enable developers who are experts in existing communication approaches to create simple plug-ins that use different query comm   </text>
<query_num> 18607 </query_num>
<text>   particular hardware used to support it, in our initial implementation, these software components have been developed for Crossbow Mica2 motes [23] and are written for TinyOS [24] in the nesC language =-=[25]-=-. Our nesC implementation of the scene abstraction (along with other project information) is availScene Receive StdControl SceneM SendMsg StdControl ContextQuery QueuedSend StdControl GenericCommPromi   </text>
<query_num> 18608 </query_num>
<text>   r differ significantly from other approaches, so direct comparison to existing protocols is not very meaningful. Instead, we measured our protocol’s overhead in varying scenarios, by employing TOSSIM =-=[27]-=-, a simulator that allows direct simulation of code written for TinyOS. TOSSIM therefore allows us to perform large-scale simulations (in this case, of 100 nodes); these simulations are of a scale tha   </text>
<query_num> 18609 </query_num>
<text>   s. In this section, we highlight some existing work for both sensor networks and pervasive computing and examine how well each approach addresses the aforementioned requirements. Network abstractions =-=[9]-=- allows applications to provide metrics over network paths; nodes to which there exists a path satisfying the metric are included in the network context. This approach is overly expressive for pervasi   </text>
<query_num> 18610 </query_num>
<text>   stract Regions [13] define regions of coordination and couple the abstraction with programming constructs that allow applications to issue operations over the regions. Likewise, logical neighborhoods =-=[14]-=- provide a communication infrastructure that logically groups similar nodes. These three approaches do not directly consider the dynamics of mobility, and they require proactive behavior by all sensor   </text>
<query_num> 18611 </query_num>
<text>   upport of pervasive computing. While the networks may remain domainspecific, the deployed applications may not be known a priori and may include varying adaptive behaviors, for example in aware homes =-=[1]-=-, intelligent construction sites [2], and first responder deployments [3]. These applications require on-demand access to local information, which is exactly the vision of pervasive computing [4], in   </text>
</top>
<top>
<paper_num> 187 </paper_num>
<paper_title>   WARPP: a toolkit for simulating high-performance parallel scientific codes.  </paper_title>
<paper_abstract>   There are a number of challenges facing the High Performance  </paper_abstract>
<query_num> 18701 </query_num>
<text>   ) methods, (e.g. LogP [6], LogGP [2], LoPC [9]), modelling based on tool support and simulation (e.g. PACE [11, 4] and DIMEMAS [10, 19]), and a hybrid approach which uses elements of both (e.g. POEMS =-=[1]-=-). Modelling based on tool support has a number of advantages over its more mathematical counterpart: firstly, it is often based on (source) code analysis, which absolves the user from translating len   </text>
<query_num> 18702 </query_num>
<text>   SMP node and the slower, lower bandwidth InfiniBand network. Each of these networks has complex performance properties which must be modelled if the simulation of communications is to be accurate. In =-=[5]-=- the author’s show that in a number of high performance codes the use of local communications (i.e. those within a single node) can be up to 50% of the total messages sent during execution, demonstrat   </text>
<query_num> 18703 </query_num>
<text>   attempt to replicate the behaviour of the code with respect to a set of input parameters such as machine processor count, network latency etc. Examples include the Wisconsin Wind Tunnel [22], PROTEUS =-=[3]-=- and the PACE toolkit [4, 11] also developed at the University of Warwick. The notable problem with this previous research has been that in order for simulations to achieve appreciable levels of accur   </text>
<query_num> 18704 </query_num>
<text>   f systems for the United States Department of Energy [18]. The process of modelling itself can be generalised to three basic approaches; modelling based on analytic (mathematical) methods, (e.g. LogP =-=[6]-=-, LogGP [2], LoPC [9]), modelling based on tool support and simulation (e.g. PACE [11, 4] and DIMEMAS [10, 19]), and a hybrid approach which uses elements of both (e.g. POEMS [1]). Modelling based on   </text>
<query_num> 18705 </query_num>
<text>   h the development of a simulation model using the WARPP toolkit. The modelling process requires four stages: (1) model construction, (2) machine benchmarking using a reliable MPI benchmarking utility =-=[12, 16]-=-, a filesystem I/O benchmark [24] and an instrumented version of the application, (3) the post-execution analysis of machine benchmarking results to produce simulator inputs and finally (4) simulation our, potentially enabling data dependent runtimes to be modelled. 3.2 Developing Simulator Inputs and Modelling Machine Networks Following code instrumentation, the timed code as well as reliable MPI =-=[12, 16, 23, 13]-=- and file I/O benchmarks [24] are executed on the target machine. Only a limited number of processors are required for this purpose, since the timings which are obtained can then be used to produce es   </text>
<query_num> 18706 </query_num>
<text>   imulator will attempt to replicate the behaviour of the code with respect to a set of input parameters such as machine processor count, network latency etc. Examples include the Wisconsin Wind Tunnel =-=[22]-=-, PROTEUS [3] and the PACE toolkit [4, 11] also developed at the University of Warwick. The notable problem with this previous research has been that in order for simulations to achieve appreciable le   </text>
<query_num> 18707 </query_num>
<text>   ised to three basic approaches; modelling based on analytic (mathematical) methods, (e.g. LogP [6], LogGP [2], LoPC [9]), modelling based on tool support and simulation (e.g. PACE [11, 4] and DIMEMAS =-=[10, 19]-=-), and a hybrid approach which uses elements of both (e.g. POEMS [1]). Modelling based on tool support has a number of advantages over its more mathematical counterpart: firstly, it is often based on   as limited support for complex networking models, make these toolkits less plausible solutions for users who require accurate models of large, complex parallel codes. The more recent DIMEMAS project =-=[10, 19]-=- alleviates the instructionbased simulation approach through replay of traces obtained during a run of the application. Evaluation in the context of different machine sizes is supported through the re   </text>
<query_num> 18708 </query_num>
<text>   of larger internal codes which occupy a considerable proportion of parallel runtime on the supercomputing facilities of AWE. The code shares many similarities with the ubiquitous Sweep3D application =-=[15, 17]-=- developed by the Los Alamos National Laboratory (LANL) in the United States, but is considerably larger and more complex in its operation. Unlike Sweep3D, Chimaera employs alternative sweep orderings   </text>
<query_num> 18709 </query_num>
<text>   or the United States Department of Energy [18]. The process of modelling itself can be generalised to three basic approaches; modelling based on analytic (mathematical) methods, (e.g. LogP [6], LogGP =-=[2]-=-, LoPC [9]), modelling based on tool support and simulation (e.g. PACE [11, 4] and DIMEMAS [10, 19]), and a hybrid approach which uses elements of both (e.g. POEMS [1]). Modelling based on tool suppor   </text>
<query_num> 18710 </query_num>
<text>   our, potentially enabling data dependent runtimes to be modelled. 3.2 Developing Simulator Inputs and Modelling Machine Networks Following code instrumentation, the timed code as well as reliable MPI =-=[12, 16, 23, 13]-=- and file I/O benchmarks [24] are executed on the target machine. Only a limited number of processors are required for this purpose, since the timings which are obtained can then be used to produce es sing the Intel MPI Benchmarking utility [16] version 3.0 in order to obtain a set of network profiles suitable for simulation. We note that several more advanced MPI benchmark utilities are available =-=[13, 23]-=- for complex network benchmarking, however, for our purposes the Intel benchmark is sufficient to obtain accurate point-to-point communication times which support the application modelling process. Th   </text>
<query_num> 18711 </query_num>
<text>   ted States Department of Energy [18]. The process of modelling itself can be generalised to three basic approaches; modelling based on analytic (mathematical) methods, (e.g. LogP [6], LogGP [2], LoPC =-=[9]-=-), modelling based on tool support and simulation (e.g. PACE [11, 4] and DIMEMAS [10, 19]), and a hybrid approach which uses elements of both (e.g. POEMS [1]). Modelling based on tool support has a nu   </text>
</top>
<top>
<paper_num> 188 </paper_num>
<paper_title>   Utility based sensor selection.  </paper_title>
<paper_abstract>   ... an environment and communicate using wireless links. The lifetime of these networks is severely curtailed by the limited battery power of the sensors. One line of research in sensor network lifetime management has examined sensor selection techniques, in which applications judiciously choose which sensors&amp;apos; data should be retrieved and are worth the expended energy. In the past, many ad-hoc approaches for sensor selection have been proposed. In this paper, we argue that sensor selection should be based upon a tradeoff between application-perceived benefit and energy consumption of the selected sensor set. We propose  </paper_abstract>
<query_num> 18801 </query_num>
<text>   ed utility-based sensor selection. However, they have analytically examined only the special class of utility functions that depend exclusively on the size of the set. More recently, Isler and Bajcsy =-=[32]-=- used utility functions to model the sensor selection problem for target localization. By contrast to both these pieces of work, we examine a broader class of utility functions. Finally, the work of M   </text>
<query_num> 18802 </query_num>
<text>   eighted supermodular functions, the UOSSP problem is NP-hard. THEOREM 4. UOSSP with set-weighted supermodular functions is NP-hard. Proof. We will give a reduction from the densest k-subgraph problem =-=[17, 18]-=-: Given a graph G = (V,E) and a number k, as well as a density requirement ρ, is there a subset S ⊆ V of size at most k containing at least ρ ·k edges. (Here, we say that S contains an edge e if it co   </text>
<query_num> 18803 </query_num>
<text>   lifetime using energy-aware routing [1, 2, 3, 4, 5, 6, 7, 8]. Utility functions have significant prior history in networking, having been used to model architectural questions relating to network QoS =-=[30]-=- and network pricing [31]. In the sensor networks literature, however, we have found relatively few applications thereof. Perhaps closest to our work is that of Byers and Nasser [9], who first propose   </text>
<query_num> 18804 </query_num>
<text>   ng sensor network lifetime by carefully managing communication and computation resources. One thread has focused on techniques to maximize sensor network lifetime by using routing or topology control =-=[1, 2, 3, 4, 5, 6, 7, 8]-=-. Somewhat orthogonal to this thread is the approach of asking which and how much data can be collected over the lifetime of a network. Work on this question is usually centered around one of two para  that applications, rather than the system, indicate which set of sensors should be active. Also complementary, for a similar reason, is work on increasing network lifetime using energy-aware routing =-=[1, 2, 3, 4, 5, 6, 7, 8]-=-. Utility functions have significant prior history in networking, having been used to model architectural questions relating to network QoS [30] and network pricing [31]. In the sensor networks litera   </text>
<query_num> 18805 </query_num>
<text>   practice in sensor network deployment. Most existing deployments, including the James Reserve habitat monitoring network [10], the Great Duck Island network [11, 12], and the Extreme Scaling network =-=[13]-=- are tiered: they consist of a large number of small battery-powered motes, sending data (perhaps after some local processing) to a smaller number of well endowed upper-tier 32-bit embedded nodes (e.g   </text>
<query_num> 18806 </query_num>
<text>   s network model is consistent with current practice in sensor network deployment. Most existing deployments, including the James Reserve habitat monitoring network [10], the Great Duck Island network =-=[11, 12]-=-, and the Extreme Scaling network [13] are tiered: they consist of a large number of small battery-powered motes, sending data (perhaps after some local processing) to a smaller number of well endowed   </text>
<query_num> 18807 </query_num>
<text>   the literature. For example, there is extensive work on overhearing avoidance and node duty-cycling at the MAC layer [22, 23, 24, 25]. As well, several pieces of work have focused on topology control =-=[26, 27, 28, 29]-=- to conserve energy in dense deployments. Our utility-based sensor selection is largely complementary, in that applications, rather than the system, indicate which set of sensors should be active. Als   </text>
<query_num> 18808 </query_num>
<text>   utility functions applicable to geometric coverage settings. 3. SUBMODULAR UTILITY FUNCTIONS In game theory, submodular functions are frequently studied as a natural restriction of utility functions =-=[16]-=-. Recall that a function u : 2 V → R is submodular if u(S1 ∪ S2) + u(S1 ∩ S2) ≤ u(S1) + u(S2) for all sets S1,S2. The practical importance of submodular functions is more evident from an equivalent ch   </text>
<query_num> 18809 </query_num>
<text>   y-conservation in sensor networks has received a tremendous amount of attention in the literature. For example, there is extensive work on overhearing avoidance and node duty-cycling at the MAC layer =-=[22, 23, 24, 25]-=-. As well, several pieces of work have focused on topology control [26, 27, 28, 29] to conserve energy in dense deployments. Our utility-based sensor selection is largely complementary, in that applic   </text>
</top>
<top>
<paper_num> 189 </paper_num>
<paper_title>   Representing uncertain data: models, properties, and algorithms.  </paper_title>
<paper_abstract>   In general terms, an uncertain relation encodes a set of possible certain relations. There are many ways to represent uncertainty, ranging from alternative values for attributes to rich constraint languages. Among the possible models for uncertain data, there is a tension between simple and intuitive models, which tend to be incomplete, and complete models, which tend to be nonintuitive and more complex than necessary for many applications. We present a space of models for representing uncertain data based on a variety of uncertainty constructs and tuple-existence constraints. We explore a number of properties and results for these models. We study completeness of the models, as well as closure under relational operations, and we give results relating closure and completeness. We then examine whether different models guarantee unique representations of uncertain data, and for those models that do not, we provide complexity results and algorithms for testing equivalence of representations. The next problem we consider is that of minimizing the size of representation of models, showing that minimizing the number of tuples also minimizes the size of constraints. We show that minimization is intractable in general and study the more restricted problem of maintaining minimality incrementally when performing operations. Finally, we present several results on the problem of approximating uncertain data in an insufficiently expressive model.  </paper_abstract>
<query_num> 18901 </query_num>
<text>   Although in general finding best approximations for propositional theories into tractable classes is known to be a hard problem, approximation techniques have been proposed in the past. Specifically, =-=[42]-=- describes techniques for finding 2-CNF lower and upper bounds f1 and f2 for a general propositional formula f, i.e., the satisfying assignments of f1 (f2 respectively) are a subset (superset respecti he number of differing satisfying assignments (which correspond to possible instances in our case), and not the number of variables (which would correspond to the number of tuples). The algorithms in =-=[42]-=- proceed in an online fashion, progressively giving better approximations until finding the best. We can employ these techniques to approximate the set of constraints f in R ′ , and thus obtain subset   </text>
<query_num> 18902 </query_num>
<text>   and we have not seen these problems addressed in the related areas. Approximate query answering, and obtaining ranked results to imprecisely defined queries, is also an active area of research, e.g., =-=[5,30,32,57]-=-. This body of work differs from ours in that we look at modeling uncertainty and querying it exactly, as opposed to modeling exact data and querying it approximately. 11 Conclusions and Future Work T   </text>
<query_num> 18903 </query_num>
<text>   and we have not seen these problems addressed in the related areas. Approximate query answering, and obtaining ranked results to imprecisely defined queries, is also an active area of research, e.g., =-=[5,30,32,57]-=-. This body of work differs from ours in that we look at modeling uncertainty and querying it exactly, as opposed to modeling exact data and querying it approximately. 11 Conclusions and Future Work T s paper suggest a number of areas for future work: – As mentioned in Section 10, the area of probabilistic databases is of great interest. Uncertain databases can be extended to include probabilities =-=[10,13,18,32,43]-=-, so extending the definitions and results in this paper poses a set of interesting open problems. – We considered the problem of maintaining minimality incrementally for only certain models and opera   </text>
<query_num> 18904 </query_num>
<text>   ations R1 and R2 of R, i.e., I(R1) ⊆ I(R) and I(R2) ⊇ I(R). 10 Related Work The study of uncertain databases has a long history, dating back to a series of initial papers from the early 1980’s, e.g., =-=[2,11,15,38,58]-=-, and a great deal of follow-on work, e.g., [10,12,16,23,30,37,38,43,44,46,48,56]. Much of this previous work lays theoretical foundations and considers query answering, e.g., [2,3,35,58]. Systems bas   </text>
<query_num> 18905 </query_num>
<text>   f alternative values can be used to indicate uncertainty about the value of a particular attribute in a tuple [10,25,27,30,31,39,45] or the entire tuple can be comprised of several alternative tuples =-=[12, 25,37,53]-=-. Annotations can be used to indicate uncertainty about whether or not a tuple is present, and constraints on variables or tuple identifiers can be used to correlate uncertainties such as tuple presen -ors and “?”s we can also represent the three-way mutual-exclusion in Example 5 above, but we still do not achieve closure: the mutual inclusion in Example 3 is still not representable. Previous work =-=[12,25,37,53]-=- has used similar constructs to represent possible values for a tuple. 2.2.4 C-Tables Most early work in uncertain databases has been devoted to defining and analyzing data models, and in particular,   10 Related Work The study of uncertain databases has a long history, dating back to a series of initial papers from the early 1980’s, e.g., [2,11,15,38,58], and a great deal of follow-on work, e.g., =-=[10,12,16,23,30,37,38,43,44,46,48,56]-=-. Much of this previous work lays theoretical foundations and considers query answering, e.g., [2,3,35,58]. Systems based around uncertain data are discussed in [6,13,19,40,43,60,59]. In this paper, w   </text>
<query_num> 18906 </query_num>
<text>   f alternative values can be used to indicate uncertainty about the value of a particular attribute in a tuple [10,25,27,30,31,39,45] or the entire tuple can be comprised of several alternative tuples =-=[12, 25,37,53]-=-. Annotations can be used to indicate uncertainty about whether or not a tuple is present, and constraints on variables or tuple identifiers can be used to correlate uncertainties such as tuple presen -ors and “?”s we can also represent the three-way mutual-exclusion in Example 5 above, but we still do not achieve closure: the mutual inclusion in Example 3 is still not representable. Previous work =-=[12,25,37,53]-=- has used similar constructs to represent possible values for a tuple. 2.2.4 C-Tables Most early work in uncertain databases has been devoted to defining and analyzing data models, and in particular,  on is intractable. Then, we show that reducing the number of tuples also reduces21 the size of the minimal constraint, and we study incremental minimality, which are not considered in [7]. Reference =-=[53]-=- studies a limited form of approximation in uncertain databases. They present a new data model, called BID tables, and address the problem of approximating uncertain relations as BID tables. They show   </text>
<query_num> 18907 </query_num>
<text>   he problem of approximating uncertain data in an insufficiently expressive model. 1 Introduction The field of uncertain databases has attracted considerable attention over the last few decades (e.g., =-=[2,3,31,38,43]-=-), This work was supported by the National Science Foundation under grants IIS-0324431 and IIS-0414762, by grants from the Boeing and Hewlett-Packard Corporations, by a Microsoft Graduate Fellowship,   to indicate uncertainty about whether or not a tuple is present, and constraints on variables or tuple identifiers can be used to correlate uncertainties such as tuple presence or alternative values =-=[3,21, 23,34,35,38,43]-=-. In this paper, we consider combinations of tuple-level constructs as well as constraints across tuples, giving us a space of models for uncertain data comprised of a finite set of possible instances et of relation instances has a unique representation in M; otherwise M is non-unique. We analyze uniqueness properties of the different models we study. Previous models for uncertain databases, e.g., =-=[3,31,34,38,43]-=-, are generally very expressive and easily seen to be non-unique. The problem of uniqueness becomes significantly more interesting when we consider simpler models. For the non-unique models, we addres hether the tuple is in the relation. (Previous work has used similar constructs to represent uncertainty in the value of an attribute [10,25,27,30,31,39,45] and uncertainty in the presence of a tuple =-=[3,21,23,34,35,38,43]-=-.) Intuitively, this uncertain relation represents the following set of three possible relation instances, the first containing only a single tuple, and the other two containing two tuples and differi  10 Related Work The study of uncertain databases has a long history, dating back to a series of initial papers from the early 1980’s, e.g., [2,11,15,38,58], and a great deal of follow-on work, e.g., =-=[10,12,16,23,30,37,38,43,44,46,48,56]-=-. Much of this previous work lays theoretical foundations and considers query answering, e.g., [2,3,35,58]. Systems based around uncertain data are discussed in [6,13,19,40,43,60,59]. In this paper, w   </text>
<query_num> 18908 </query_num>
<text>   he problem of approximating uncertain data in an insufficiently expressive model. 1 Introduction The field of uncertain databases has attracted considerable attention over the last few decades (e.g., =-=[2,3,31,38,43]-=-), This work was supported by the National Science Foundation under grants IIS-0324431 and IIS-0414762, by grants from the Boeing and Hewlett-Packard Corporations, by a Microsoft Graduate Fellowship,   to indicate uncertainty about whether or not a tuple is present, and constraints on variables or tuple identifiers can be used to correlate uncertainties such as tuple presence or alternative values =-=[3,21, 23,34,35,38,43]-=-. In this paper, we consider combinations of tuple-level constructs as well as constraints across tuples, giving us a space of models for uncertain data comprised of a finite set of possible instances relation I an instance of an uncertain relation R? (4) Is a given certain relation I the only instance of an uncertain relation R? A considerable amount of past work has studied these problems, e.g., =-=[2,3,34,35,38]-=-. We give complexity results for these problems with respect to the new models we introduce, showing that some of the simpler models permit more efficient membership testing. Uniqueness and Equivalenc et of relation instances has a unique representation in M; otherwise M is non-unique. We analyze uniqueness properties of the different models we study. Previous models for uncertain databases, e.g., =-=[3,31,34,38,43]-=-, are generally very expressive and easily seen to be non-unique. The problem of uniqueness becomes significantly more interesting when we consider simpler models. For the non-unique models, we addres hether the tuple is in the relation. (Previous work has used similar constructs to represent uncertainty in the value of an attribute [10,25,27,30,31,39,45] and uncertainty in the presence of a tuple =-=[3,21,23,34,35,38,43]-=-.) Intuitively, this uncertain relation represents the following set of three possible relation instances, the first containing only a single tuple, and the other two containing two tuples and differi   </text>
<query_num> 18909 </query_num>
<text>   ine the necessary fundamentals for the remainder of the paper (Section 2.3).3 2.1 A Running Example As a running example for the paper, we consider data management for the Christmas Bird Count (CBC) =-=[1,60]-=-. Each year, volunteers and professionals worldwide observe birds for a fixed period of time, recording their observations. The data from year to year is used to understand trends in bird populations, , [10,12,16,23,30,37,38,43,44,46,48,56]. Much of this previous work lays theoretical foundations and considers query answering, e.g., [2,3,35,58]. Systems based around uncertain data are discussed in =-=[6,13,19,40,43,60,59]-=-. In this paper, we introduced a space of uncertain-data models based on tuple-level uncertainty constructs and tuple-existence constraints. We then studied the problems of closure and relative expres oximation. Trio: The problems addressed in this paper arose in the context of the Trio project at Stanford, whose objective is to develop a system that fully integrates data, uncertainty, and lineage =-=[60]-=-. The enumeration of the space of models, their expressiveness hierarchy, and related membership problems were studied in an initial Trio paper on models for uncertainty [25]. Although the approximati   </text>
<query_num> 18910 </query_num>
<text>   not allow for probability distributions; revisiting and extending our results to the probabilistic case is an important direction of future work. Another related area is inconsistent databases, e.g., =-=[4, 8,9,14,20,28,36,61]-=-, in which the possible “minimal repairs” [9,17,61] to an inconsistent database result in a set of possible instances (i.e., an uncertain database). Reasoning with uncertainty in the Artificial Intell   </text>
<query_num> 18911 </query_num>
<text>   one uncertain relation is contained in that of another. The problem of equivalence testing can be solved using containment. However, minimization and approximation are not discussed in [3]. Reference =-=[7]-=- studies the problem of finding maximal decompositions of an uncertain relation represented as a “world-set decomposition” (called the gWSD data model). Maximal decompositions in their setting can be  nts, minimization is intractable. Then, we show that reducing the number of tuples also reduces21 the size of the minimal constraint, and we study incremental minimality, which are not considered in =-=[7]-=-. Reference [53] studies a limited form of approximation in uncertain databases. They present a new data model, called BID tables, and address the problem of approximating uncertain relations as BID t   </text>
<query_num> 18912 </query_num>
<text>   our results to the probabilistic case is an important direction of future work. Another related area is inconsistent databases, e.g., [4, 8,9,14,20,28,36,61], in which the possible “minimal repairs” =-=[9,17,61]-=- to an inconsistent database result in a set of possible instances (i.e., an uncertain database). Reasoning with uncertainty in the Artificial Intelligence context is also related, e.g., [29,50,54]. A   </text>
<query_num> 18913 </query_num>
<text>   porations, by a Microsoft Graduate Fellowship, and by a Stanford Graduate Fellowship from Sequoia Capital. 1 Stanford University, 2 Google Inc., 3 Microsoft Corp. and is experiencing revived interest =-=[6,13,16,19,37,40,56, 59]-=- due to the increasing popularity of applications such as data cleaning and integration, information extraction, scientific and sensor databases, and others. We observe that data models for uncertaint  10 Related Work The study of uncertain databases has a long history, dating back to a series of initial papers from the early 1980’s, e.g., [2,11,15,38,58], and a great deal of follow-on work, e.g., =-=[10,12,16,23,30,37,38,43,44,46,48,56]-=-. Much of this previous work lays theoretical foundations and considers query answering, e.g., [2,3,35,58]. Systems based around uncertain data are discussed in [6,13,19,40,43,60,59]. In this paper, w   </text>
<query_num> 18914 </query_num>
<text>   porations, by a Microsoft Graduate Fellowship, and by a Stanford Graduate Fellowship from Sequoia Capital. 1 Stanford University, 2 Google Inc., 3 Microsoft Corp. and is experiencing revived interest =-=[6,13,16,19,37,40,56, 59]-=- due to the increasing popularity of applications such as data cleaning and integration, information extraction, scientific and sensor databases, and others. We observe that data models for uncertaint , [10,12,16,23,30,37,38,43,44,46,48,56]. Much of this previous work lays theoretical foundations and considers query answering, e.g., [2,3,35,58]. Systems based around uncertain data are discussed in =-=[6,13,19,40,43,60,59]-=-. In this paper, we introduced a space of uncertain-data models based on tuple-level uncertainty constructs and tuple-existence constraints. We then studied the problems of closure and relative expres   </text>
<query_num> 18915 </query_num>
<text>   porations, by a Microsoft Graduate Fellowship, and by a Stanford Graduate Fellowship from Sequoia Capital. 1 Stanford University, 2 Google Inc., 3 Microsoft Corp. and is experiencing revived interest =-=[6,13,16,19,37,40,56, 59]-=- due to the increasing popularity of applications such as data cleaning and integration, information extraction, scientific and sensor databases, and others. We observe that data models for uncertaint f alternative values can be used to indicate uncertainty about the value of a particular attribute in a tuple [10,25,27,30,31,39,45] or the entire tuple can be comprised of several alternative tuples =-=[12, 25,37,53]-=-. Annotations can be used to indicate uncertainty about whether or not a tuple is present, and constraints on variables or tuple identifiers can be used to correlate uncertainties such as tuple presen -ors and “?”s we can also represent the three-way mutual-exclusion in Example 5 above, but we still do not achieve closure: the mutual inclusion in Example 3 is still not representable. Previous work =-=[12,25,37,53]-=- has used similar constructs to represent possible values for a tuple. 2.2.4 C-Tables Most early work in uncertain databases has been devoted to defining and analyzing data models, and in particular,  ertain small set of operations, can represent all finite sets of possible instances of an uncertain relation, i.e., M is complete. Note that a more general similar result was subsequently obtained in =-=[37]-=-. Theorem 4 (Closure ⇒ Completeness) Consider a model M with the following properties: – Basic Uncertainty: M can represent all ordinary (certain) relations as well as independent copies of a unary un  10 Related Work The study of uncertain databases has a long history, dating back to a series of initial papers from the early 1980’s, e.g., [2,11,15,38,58], and a great deal of follow-on work, e.g., =-=[10,12,16,23,30,37,38,43,44,46,48,56]-=-. Much of this previous work lays theoretical foundations and considers query answering, e.g., [2,3,35,58]. Systems based around uncertain data are discussed in [6,13,19,40,43,60,59]. In this paper, w   </text>
<query_num> 18916 </query_num>
<text>   relation. A variety of constructs can be used to represent possible instances [2]: A set of alternative values can be used to indicate uncertainty about the value of a particular attribute in a tuple =-=[10,25,27,30,31,39,45]-=- or the entire tuple can be comprised of several alternative tuples [12, 25,37,53]. Annotations can be used to indicate uncertainty about whether or not a tuple is present, and constraints on variable the two values, and “?” denotes a maybe-tuple, i.e., uncertainty whether the tuple is in the relation. (Previous work has used similar constructs to represent uncertainty in the value of an attribute =-=[10,25,27,30,31,39,45]-=- and uncertainty in the presence of a tuple [3,21,23,34,35,38,43].) Intuitively, this uncertain relation represents the following set of three possible relation instances, the first containing only a   </text>
<query_num> 18917 </query_num>
<text>   relation. A variety of constructs can be used to represent possible instances [2]: A set of alternative values can be used to indicate uncertainty about the value of a particular attribute in a tuple =-=[10,25,27,30,31,39,45]-=- or the entire tuple can be comprised of several alternative tuples [12, 25,37,53]. Annotations can be used to indicate uncertainty about whether or not a tuple is present, and constraints on variable the two values, and “?” denotes a maybe-tuple, i.e., uncertainty whether the tuple is in the relation. (Previous work has used similar constructs to represent uncertainty in the value of an attribute =-=[10,25,27,30,31,39,45]-=- and uncertainty in the presence of a tuple [3,21,23,34,35,38,43].) Intuitively, this uncertain relation represents the following set of three possible relation instances, the first containing only a   10 Related Work The study of uncertain databases has a long history, dating back to a series of initial papers from the early 1980’s, e.g., [2,11,15,38,58], and a great deal of follow-on work, e.g., =-=[10,12,16,23,30,37,38,43,44,46,48,56]-=-. Much of this previous work lays theoretical foundations and considers query answering, e.g., [2,3,35,58]. Systems based around uncertain data are discussed in [6,13,19,40,43,60,59]. In this paper, w and we have not seen these problems addressed in the related areas. Approximate query answering, and obtaining ranked results to imprecisely defined queries, is also an active area of research, e.g., =-=[5,30,32,57]-=-. This body of work differs from ours in that we look at modeling uncertainty and querying it exactly, as opposed to modeling exact data and querying it approximately. 11 Conclusions and Future Work T   </text>
<query_num> 18918 </query_num>
<text>   rs” [9,17,61] to an inconsistent database result in a set of possible instances (i.e., an uncertain database). Reasoning with uncertainty in the Artificial Intelligence context is also related, e.g., =-=[29,50,54]-=-. Again, our work in this paper focuses on a specific set of problems associated with models for representing uncertainty, and we have not seen these problems addressed in the related areas. Approxima   </text>
<query_num> 18919 </query_num>
<text>   s paper suggest a number of areas for future work: – As mentioned in Section 10, the area of probabilistic databases is of great interest. Uncertain databases can be extended to include probabilities =-=[10,13,18,32,43]-=-, so extending the definitions and results in this paper poses a set of interesting open problems. – We considered the problem of maintaining minimality incrementally for only certain models and opera   </text>
<query_num> 18920 </query_num>
<text>   to indicate uncertainty about whether or not a tuple is present, and constraints on variables or tuple identifiers can be used to correlate uncertainties such as tuple presence or alternative values =-=[3,21, 23,34,35,38,43]-=-. In this paper, we consider combinations of tuple-level constructs as well as constraints across tuples, giving us a space of models for uncertain data comprised of a finite set of possible instances hether the tuple is in the relation. (Previous work has used similar constructs to represent uncertainty in the value of an attribute [10,25,27,30,31,39,45] and uncertainty in the presence of a tuple =-=[3,21,23,34,35,38,43]-=-.) Intuitively, this uncertain relation represents the following set of three possible relation instances, the first containing only a single tuple, and the other two containing two tuples and differi   </text>
<query_num> 18921 </query_num>
<text>   to indicate uncertainty about whether or not a tuple is present, and constraints on variables or tuple identifiers can be used to correlate uncertainties such as tuple presence or alternative values =-=[3,21, 23,34,35,38,43]-=-. In this paper, we consider combinations of tuple-level constructs as well as constraints across tuples, giving us a space of models for uncertain data comprised of a finite set of possible instances hether the tuple is in the relation. (Previous work has used similar constructs to represent uncertainty in the value of an attribute [10,25,27,30,31,39,45] and uncertainty in the presence of a tuple =-=[3,21,23,34,35,38,43]-=-.) Intuitively, this uncertain relation represents the following set of three possible relation instances, the first containing only a single tuple, and the other two containing two tuples and differi  10 Related Work The study of uncertain databases has a long history, dating back to a series of initial papers from the early 1980’s, e.g., [2,11,15,38,58], and a great deal of follow-on work, e.g., =-=[10,12,16,23,30,37,38,43,44,46,48,56]-=-. Much of this previous work lays theoretical foundations and considers query answering, e.g., [2,3,35,58]. Systems based around uncertain data are discussed in [6,13,19,40,43,60,59]. In this paper, w   </text>
</top>
<top>
<paper_num> 190 </paper_num>
<paper_title>   Mitigating denial of service attacks: A tutorial.  </paper_title>
<paper_abstract>   This tutorial describes what Denial of Service (DoS) attacks are, how they can be carried out in IP networks, and how one can defend against them. Distributed DoS (DDoS) attacks are included here as a subset of DoS attacks. A DoS attack has two phases: a deployment and an attack phase. A DoS program must first be deployed on one or more compromised hosts before an attack is possible. Mitigation of DoS attacks requires thus defense mechanisms for both phases. Completely reliable protection against DoS attacks is, however, not possible. There will always be vulnerable hosts in the Internet, and many attack mechanisms are based on ordinary use of protocols. Defense in depth is thus needed to mitigate the effect of DoS attacks. This paper describes shortly many defense mechanisms proposed in the literature. The goal is not to implement all possible defenses. Instead, one should optimize the trade-off between security costs and acquired benefits in handling the most important risks. Mitigation of DoS attacks is thus closely related to risk management.  </paper_abstract>
<query_num> 19001 </query_num>
<text>   after which hosts may crash due to unavailability of free memory. Reflector attacks utilize any protocol behavior, where an attack packet triggers a response packet to be sent to the ultimate victim =-=[49]-=-. Reflector attacks can also include the use of a technique called bandwidth or packet amplification. The innocent third party will either reply with a longer packet or with several packets to a singl   </text>
<query_num> 19002 </query_num>
<text>   cker coordinates a flooding or a logic attack against a victim. Both of these phases make use of deficiencies in the design or implementation of applications, protocols, and the Internet architecture =-=[37]-=-. A vulnerability is a flaw in security procedures, software, internal system controls, or implementation of an information system that may affect the integrity, confidentiality, and/or availability o   </text>
<query_num> 19003 </query_num>
<text>   e arbitrary code (e.g., Code Red II [41]). Some worms, however, are designed to propagate fast without any malicious payload (e.g., Slammer [40]) or their full functionality is not known (e.g., Nimda =-=[57]-=-). Viruses can also be used for the deployment phase to build a large DDoS network, but they cannot replicate automatically by themselves. Typically social engineering is required to get a human to st on hosts with different operating systems or applications. 3.1. Modeling worm propagation Propagation of worms can be modeled well with the epidemic model describing the spread of infectious diseases =-=[42,57]-=-. Epidemic model gives the average infection rate in a population where the infected individuals (infectives) contact uninfected individuals (susceptibles) with the average contact rate of β. Accordin an efficient scanning strategy increases the probability of finding new vulnerable hosts, and this decreases the time spent on trying to infect non-vulnerable, non-existing, or already infected hosts =-=[57]-=-. Simple worms use the random scanning strategy which is based on scanning the whole address space in a random order [41]. Awormusingthelocalized scanning strategy tries to infect hosts with the same  ties are appealing due to wide-bandwidth Internet connections and open usage policies [21]. By using threads carefully a worm can send even congestion controlled TCP packets with the full access rate =-=[57]-=-. By using UDP a worm does not experience any delays due tos812 J. Mölsä / Mitigating denial of service attacks: A tutorial connection setup or end-to-end congestion control. With UDP it is even possi loitation of vulnerabilities in Peer-to-Peer (P2P) applications can be easy, because all participants are using the same application-level protocol and there are at most few different implementations =-=[57]-=-. Insertion of infected files or exploitation of vulnerabilities in a P2P application-level protocol can compromise a large number of hosts participating in the same P2P network. A popular P2P system   </text>
<query_num> 19004 </query_num>
<text>   elaying of packets. Packet mistreatment attacks can be detected by running a specific detection protocol inside an Autonomous System to locate and isolate routers dropping or misrouting valid packets =-=[6]-=-. • Ordinary DoS attacks are flooding or logic DoS attacks already described in this paper. The same defense mechanisms can be used, regardless of the victim being an ordinary host or a node in the in   </text>
<query_num> 19005 </query_num>
<text>   enough crashing from time to time. An IDS can be overloaded intentionally by an attacker, so that at least part of the attack traffic is dropped by the IDS. This makes it possible to evade detection =-=[48]-=-. An IDS is also susceptible to a crash attack, in which the attacker knocks down whole or part of an IDS by utilizing some vulnerability [48]. Flooding and logic DoS attacks can thus be used against  xample, an NIDS may reassemble overlapping IP fragments in a different fashion than an end-host. This can prevent an NIDS from seeing complete signatures broken down in multiple overlapping fragments =-=[48]-=-. In an insertion attack an NIDS accepts a packet that an end-system rejects or does not receive. In an evasion attack an NIDS rejects a packet that an end-system accepts. Both insertion and evasion a tarting an application in a more secure form (e.g., with more effective and more expensive checks against buffer overflow attacks) [15], • killing of active network connections (e.g., with a TCP RST) =-=[48]-=-, and • delaying of network connections to new destinations [62]. Worm propagation and DoS tool deployment are typically based on so called stack smashing attacks [14], where an attacker can get unaut Bro, Aggregate-based Congestion Control (ACC) and Cooperative Intrusion Traceback and Response Architecture (CITRA). These systems can be used to defend against both deployment and attack phases. Bro =-=[48]-=- is a system for detecting network intruders in real-time, and it can thus be classified as an NIDS with extra capability for defining policies with a specific language. In addition to monitoring, Bro   </text>
<query_num> 19006 </query_num>
<text>   exploit a format string vulnerability [5]. Stack smashing attacks can be prevented or made more difficult by paying attention to software security which can be enhanced by three different mechanisms =-=[13]-=-: • Software auditing can be used to search vulnerabilities from source code automatically or manually before these vulnerabilities are found and exploited by attackers. • Vulnerability mitigation is   </text>
<query_num> 19007 </query_num>
<text>   false positives [44]. Instead of technical effectiveness one should increase the cost/benefit trade-off, which means focusing the limited computer and human resources on the most damaging intrusions =-=[33]-=-. This is done by tuning the signatures or training the anomaly detection to detect those attacks, for which the environment is most vulnerable, and to prevent alerting on those attacks, for which env mprehensive set of defenses. Implementing and applying every possible defense is, however, not feasible. This would simply cost too much in terms of resources, like humans, equipment, money, and time =-=[33]-=-. It is not even possible to achieve perfect network security: new vulnerabilities in computer systems are continuously found, security of one site is dependent on the security of other sites in the I hese assets should be protected from certain threats. With the help of a security policy it is possible to concentrate consistently on the major threats and implement a cost-effective set of defenses =-=[33]-=-. Legislation, standards, best current practices, and other documents may dictate parts of security policy and risk management. Legislation may specify requirements for availability of public services   </text>
<query_num> 19008 </query_num>
<text>   fective and more expensive checks against buffer overflow attacks) [15], • killing of active network connections (e.g., with a TCP RST) [48], and • delaying of network connections to new destinations =-=[62]-=-. Worm propagation and DoS tool deployment are typically based on so called stack smashing attacks [14], where an attacker can get unauthorized access (like a root shell) to a victim host. These attac ows with identical characteristics, one entering a host and the other leaving the same host [65]. Worm propagation can be restricted by limiting (delaying) the rate of connections to new destinations =-=[62]-=-. An infected host will try to connect to as many different hosts as fast as possible, but an uninfected host typically makes connections to locally correlated destinations at a lower rate. The slower   </text>
<query_num> 19009 </query_num>
<text>   ity for defining policies with a specific language. In addition to monitoring, Bro can terminate connections by sending RST packets or ask a router to drop traffic involving a particular address. ACC =-=[35]-=- tries to prevent general network congestion by detecting some of the most wide-bandwidth aggregates and rate-limiting them. ACC does not make any difference for the origin or reason of the congestion   </text>
<query_num> 19010 </query_num>
<text>   l damage or block specific behavior known to be dangerous (access control). Stack smashing attacks can also be prevented or made more difficult by looking at short sequences of operating system calls =-=[20]-=-. For example, the exploit of a buffer overflow vulnerability in a sendmail-program causes it to issue abnormal sequences of system calls, when a sendmail-process starts to execute a root shell. If de   </text>
<query_num> 19011 </query_num>
<text>   lt of a TCP packet. The majority of identified direct DoS attacks are thus TCP-based, probably TCP SYN floods. The median attack duration was 10 minutes. Fragmentation in real networks was studied in =-=[56]-=-. Bugs in the fragment handling software are exploited in many logic DoS attacks, and the results of this study still indicate the presence of these kind of DoS attacks in the Internet. 5. Handling Do   </text>
<query_num> 19012 </query_num>
<text>   on hosts with different operating systems or applications. 3.1. Modeling worm propagation Propagation of worms can be modeled well with the epidemic model describing the spread of infectious diseases =-=[42,57]-=-. Epidemic model gives the average infection rate in a population where the infected individuals (infectives) contact uninfected individuals (susceptibles) with the average contact rate of β. Accordin n routers. If a (global) defense infrastructure is available, the amount of infected hosts can be reasonably restricted by installing filtering rules in the most important routers of the Internet. In =-=[42]-=- it was studied how the reaction time affects the size of the infected population in case of the Code Red I v2 worm. To keep the ratio of susceptibles infected within 24 hours below 10%, simple addres ddresses of the infected hosts) should be installed within 20 minutes, but more generic content filtering (filtering based on the signature of the worm) allows almost three hours for installation. In =-=[42]-=- it was also estimated that in case of the Code Red I v2 worm the infected population could have been restricted to less than 20% of the population really infected. This would have required the instal   </text>
<query_num> 19013 </query_num>
<text>   ss ad hoc networks have their additional vulnerabilities, but these kind of wireless networks are not the subject of this paper. A DoS attack can be carried out either as a flooding or a logic attack =-=[43]-=-. A flooding DoS attack is based on brute force. Real-looking but unnecessary data is sent as much as possible to a victim. As a result, network bandwidth is wasted, disk space is filled with unnecess electing victims with higher access rates will make a worm more virulent. Home users with wide-bandwidth Digital Subscriber Lines (xDSL) are seldom security-conscious which makes them easy to exploit =-=[43]-=-. Universities are appealing due to wide-bandwidth Internet connections and open usage policies [21]. By using threads carefully a worm can send even congestion controlled TCP packets with the full ac most all DDoS agents were Unix or Linux hosts, some of which resided in university networks. The backscatter analysis was used to assess the number, duration, and focus of DoS attacks in the Internet =-=[43]-=-. Backscatter is called the unsolicited response traffic which the victim sends in response to direct attack packets with spoofed IP source address. The results indicate more than 12 000 attacks again ntations of Mobile IP. TCP SYN flooding is a widely used flooding DoS attack mechanism. Most available DoS tools support this attack type and studies also indicate that most DoS attacks are TCP-based =-=[43]-=-. The effect of TCP SYN flooding attacks can be mitigated by applying the following defenses [52]: • improve end-system configurations (reduction of the timeout period for halfopen connections, increa   </text>
<query_num> 19014 </query_num>
<text>   tools were deployed manually, but now worms are typically used for that. Worms are self-propagating malicious software which often either include directly the DDoS attack capability (e.g., Code Red I =-=[41]-=- and Slapper [3]) or contain the possibility to execute arbitrary code (e.g., Code Red II [41]). Some worms, however, are designed to propagate fast without any malicious payload (e.g., Slammer [40])  nt on trying to infect non-vulnerable, non-existing, or already infected hosts [57]. Simple worms use the random scanning strategy which is based on scanning the whole address space in a random order =-=[41]-=-. Awormusingthelocalized scanning strategy tries to infect hosts with the same address prefix with a higher probability [41]. This strategy infects quickly networks with many vulnerable hosts. The pro for less popular software can thus be utilized effectively for worm propagation. 3.3. Worm propagation in real-life The real-life propagation of the random scanning Code Red I v2 worm was analyzed in =-=[41]-=-. This worm exploited a vulnerability in the Microsoft IIS web server. On July 19, 2001, this latency-limited worm managed to infect more than 359 000 hosts within 24 hours. At maximum the worm genera   </text>
</top>
<top>
<paper_num> 191 </paper_num>
<paper_title>   Music Recommendation Based on Acoustic Features and User Access Patterns.  </paper_title>
<paper_abstract>   Abstract—Music recommendation is receiving increasing attention as the music industry develops venues to deliver music over the Internet. The goal of music recommendation is to present users lists of songs that they are likely to enjoy. Collaborative-filtering and content-based recommendations are two widely used approaches that have been proposed for music recommendation. However, both approaches have their own disadvantages: collaborative-filtering methods need a large collection of user history data and content-based methods lack the ability of understanding the interests and preferences of users. To overcome these limitations, this paper presents a novel dynamic music similarity measurement strategy that utilizes both content features and user access patterns. The seamless integration of them significantly improves the music similarity measurement accuracy and performance. Based on this strategy, recommended songs are obtained by a means of label propagation over a graph representing music similarity. Experimental results on a real data set collected from  </paper_abstract>
<query_num> 19101 </query_num>
<text>   . Since acoustic contents are susceptible to feature extraction, music recommendation is considered different from movie recommendation, in which meta-data is generally the only available information =-=[24]-=-. In music recommendation, the reflective and consistent acoustic features can represent song-specific characteristics such as genre, timbre, pitch, and rhythm. Comparing with the acoustic features, a   </text>
<query_num> 19102 </query_num>
<text>   [34]. Content-based methods provide recommendations based on the meta-data such as genre, styles, artists, and lyrics [28], [31], [41], and/or the acoustic features extracted from audio samples [15], =-=[17]-=-, [19], [20]. Since acoustic contents are susceptible to feature extraction, music recommendation is considered different from movie recommendation, in which meta-data is generally the only available   </text>
<query_num> 19103 </query_num>
<text>   atings are unknown. label propagation is like a random walk on a song graph [36]. Using diffusion kernel [18], [35], the label propagation is like a diffusive process of the labeled information [42], =-=[43]-=-. Zhu et al. [43] utilizes the harmonic nature of the diffusive function, Zhou et al. [42] emphasize the spread of label information in a consistent and iterative way. Motivated from the previous rese   </text>
<query_num> 19104 </query_num>
<text>   been developed, and user demographic information, music contents, user listening history, and the discography (e.g., Last.fm, Goombah, and Pandora) have been used for music recommendations [3], [4], =-=[22]-=-, [26]–[29], [38]. These approaches can be generally divided into two groups: collaborative-filtering methods and content-based methods. Collaborative-filtering methods recommend songs by identifying   </text>
<query_num> 19105 </query_num>
<text>   colored (shaded) nodes represent the rated items with their corresponding ratings. The others are the unrated items, whose ratings are unknown. label propagation is like a random walk on a song graph =-=[36]-=-. Using diffusion kernel [18], [35], the label propagation is like a diffusive process of the labeled information [42], [43]. Zhu et al. [43] utilizes the harmonic nature of the diffusive function, Zh   </text>
<query_num> 19106 </query_num>
<text>   d are similar to each other in that they are accessed by users and , but not by and . Similar ideas have been successfully applied to image retrieval to improve the accuracy of similarity measurement =-=[12]-=-, [13], [25]. (1) Let be the number of content features. The summation in (1) is rewritten as follows: B. Dynamic Weighting Schemes 1) Introduction: A simple approach capable of combining acoustic fea   </text>
<query_num> 19107 </query_num>
<text>   developed, and user demographic information, music contents, user listening history, and the discography (e.g., Last.fm, Goombah, and Pandora) have been used for music recommendations [3], [4], [22], =-=[26]-=-–[29], [38]. These approaches can be generally divided into two groups: collaborative-filtering methods and content-based methods. Collaborative-filtering methods recommend songs by identifying simila   </text>
<query_num> 19108 </query_num>
<text>   e rated items with their corresponding ratings. The others are the unrated items, whose ratings are unknown. label propagation is like a random walk on a song graph [36]. Using diffusion kernel [18], =-=[35]-=-, the label propagation is like a diffusive process of the labeled information [42], [43]. Zhu et al. [43] utilizes the harmonic nature of the diffusive function, Zhou et al. [42] emphasize the spread   </text>
<query_num> 19109 </query_num>
<text>   ent the rated items with their corresponding ratings. The others are the unrated items, whose ratings are unknown. label propagation is like a random walk on a song graph [36]. Using diffusion kernel =-=[18]-=-, [35], the label propagation is like a diffusive process of the labeled information [42], [43]. Zhu et al. [43] utilizes the harmonic nature of the diffusive function, Zhou et al. [42] emphasize the   </text>
<query_num> 19110 </query_num>
<text>   eta-data are thus very time-consuming to obtain and not capable of providing adequate information for describing listeners’ preferences [19]. Recently probabilistic models and hybrid algorithms [16], =-=[30]-=-, [41] have been proposed to overcome the aforementioned limitations by combining contents and user ratings. Yoshii et al. [41] attempted to integrate both rating and content data. They utilized Bayes   </text>
<query_num> 19111 </query_num>
<text>   experiments on a real data set constructed by anonymous users at http://www. A. Audio Similarity Extraction of audio features for music similarity search has been well studied in the literature [10], =-=[21]-=-, [23]. The use of acoustic features is justified by the fact that similar music pieces use similar instruments and possess similar sound textures [8]. The music features are vectors in a multidimensi ND INDEXING A. Feature Extraction There has been a considerable amount of research in extracting descriptive features from music signals for music genre classification and artist identification [10], =-=[21]-=-, [23], [37]. In this paper, we employ timbral features and wavelet coefficient histograms for feature extraction. The extracted feature set consists of the following three components and total 80 fea echies Wavelet Coefficient Histograms (DWCH): Daubechies wavelet filters are a set of filters that are widely used in image retrieval (see [6]). Daubechies Wavelet Coefficient Histograms, proposed in =-=[21]-=-, are features extracted in the following manners: first, the Daubechies-8 filter with seven levels of decomposition (or seven subbands) is applied to 30 s of monaural audio signals; then, the histogr , are calculated from each subband; in addition, the subband energy, defined as the mean of the absolute value of the coefficients, is computed from each subband. More details of DWCH can be found in =-=[21]-=-. B. Music Indexing Once the features/signatures for each song are obtained, efficient data structures can be built for similarity search. In this study, min-wise hashing [2] is used to speed up simil   </text>
<query_num> 19112 </query_num>
<text>   groups: collaborative-filtering methods and content-based methods. Collaborative-filtering methods recommend songs by identifying similar users or items based on ratings of items given by users [1], =-=[5]-=-, [14]. If the rating of an item by a user is unavailable, collaborative-filtering methods estimate it by computing a weighted average of known ratings of the items from similar users. Thus, for colla   </text>
<query_num> 19113 </query_num>
<text>   hose ratings are unknown. label propagation is like a random walk on a song graph [36]. Using diffusion kernel [18], [35], the label propagation is like a diffusive process of the labeled information =-=[42]-=-, [43]. Zhu et al. [43] utilizes the harmonic nature of the diffusive function, Zhou et al. [42] emphasize the spread of label information in a consistent and iterative way. Motivated from the previou   </text>
<query_num> 19114 </query_num>
<text>   igh Recommendation Novelty: Good novelty is defined as rich artist variety and well-balanced music content variety. Music content represents the information of genre, timbre, pitch, rhythm, and so on =-=[37]-=-. Well-balance means that the music content is diverse and informative while not diverging much from the user’s preferences. Various music recommendation approaches have been developed, and user demog A. Feature Extraction There has been a considerable amount of research in extracting descriptive features from music signals for music genre classification and artist identification [10], [21], [23], =-=[37]-=-. In this paper, we employ timbral features and wavelet coefficient histograms for feature extraction. The extracted feature set consists of the following three components and total 80 features. 1) Me using MFCC. It consists of the following five types of features: spectral centroid, spectral rolloff, spectral flux, zero crossings, and low energy. More detailed descriptions of STFT can be found in =-=[37]-=-. Spectral Centroid is the centroid of the magnitude spectrum of short-term Fourier transform and is a measure of spectral brightness. Spectral Rolloff is the frequency below which 85% of the magnitud   </text>
<query_num> 19115 </query_num>
<text>   ion vectors characterizes and quantifies the closeness between two pieces of music. Traditionally, there are two popular distance functions for measuring similarity in multimedia retrieval [9], [23], =-=[32]-=-: (weighted) Minkowski distance and cosine similarity. The assumption of using Minkowski distance function is that the similar objects should be close in all dimensions as all the dimensions are treat   </text>
<query_num> 19116 </query_num>
<text>   nt of user-rating data are required. This is a major limitation [33], [34]. Content-based methods provide recommendations based on the meta-data such as genre, styles, artists, and lyrics [28], [31], =-=[41]-=-, and/or the acoustic features extracted from audio samples [15], [17], [19], [20]. Since acoustic contents are susceptible to feature extraction, music recommendation is considered different from mov ta are thus very time-consuming to obtain and not capable of providing adequate information for describing listeners’ preferences [19]. Recently probabilistic models and hybrid algorithms [16], [30], =-=[41]-=- have been proposed to overcome the aforementioned limitations by combining contents and user ratings. Yoshii et al. [41] attempted to integrate both rating and content data. They utilized Bayesian ne   </text>
<query_num> 19117 </query_num>
<text>   o two groups: collaborative-filtering methods and content-based methods. Collaborative-filtering methods recommend songs by identifying similar users or items based on ratings of items given by users =-=[1]-=-, [5], [14]. If the rating of an item by a user is unavailable, collaborative-filtering methods estimate it by computing a weighted average of known ratings of the items from similar users. Thus, for   </text>
<query_num> 19118 </query_num>
<text>   of DWCH can be found in [21]. B. Music Indexing Once the features/signatures for each song are obtained, efficient data structures can be built for similarity search. In this study, min-wise hashing =-=[2]-=- is used to speed up similarity computation for large data sets, especially in online calculation. The key idea is that we can create a small signature for each song and the resemblance of any pair of emblance between and can be then measured by the proportion of the number of matches between and to , the number of components. The min-wise hashing estimator is unbiased. An error bound was given in =-=[2]-=- and the accuracy increases with the resemblance value. Note that the number of matches between two signatures can be computed in time and that is independent of the size of database. IV. MUSIC RECOMM   </text>
<query_num> 19119 </query_num>
<text>   oped, and user demographic information, music contents, user listening history, and the discography (e.g., Last.fm, Goombah, and Pandora) have been used for music recommendations [3], [4], [22], [26]–=-=[29]-=-, [38]. These approaches can be generally divided into two groups: collaborative-filtering methods and content-based methods. Collaborative-filtering methods recommend songs by identifying similar use   </text>
<query_num> 19120 </query_num>
<text>   re-dependent and user-dependent. 2) Problem Formulation: Thus, the calculation of appropriate similarity measures can be cast as a learning problem aimed to assign approximate weights to each feature =-=[39]-=-. To automatically determine the weights for audio features, the metric learning approach [13], [40], which learns appropriate similarity metrics based on the correlation between acoustic features and   </text>
<query_num> 19121 </query_num>
<text>   representation vectors characterizes and quantifies the closeness between two pieces of music. Traditionally, there are two popular distance functions for measuring similarity in multimedia retrieval =-=[9]-=-, [23], [32]: (weighted) Minkowski distance and cosine similarity. The assumption of using Minkowski distance function is that the similar objects should be close in all dimensions as all the dimensio   </text>
<query_num> 19122 </query_num>
<text>   rity measures can be cast as a learning problem aimed to assign approximate weights to each feature [39]. To automatically determine the weights for audio features, the metric learning approach [13], =-=[40]-=-, which learns appropriate similarity metrics based on the correlation between acoustic features and user access patterns of music, needs to be explored. Given that human perception of music is well a   </text>
<query_num> 19123 </query_num>
<text>   rough experiments on a real data set constructed by anonymous users at http://www. A. Audio Similarity Extraction of audio features for music similarity search has been well studied in the literature =-=[10]-=-, [21], [23]. The use of acoustic features is justified by the fact that similar music pieces use similar instruments and possess similar sound textures [8]. The music features are vectors in a multid TION AND INDEXING A. Feature Extraction There has been a considerable amount of research in extracting descriptive features from music signals for music genre classification and artist identification =-=[10]-=-, [21], [23], [37]. In this paper, we employ timbral features and wavelet coefficient histograms for feature extraction. The extracted feature set consists of the following three components and total   </text>
<query_num> 19124 </query_num>
<text>   similar to each other in that they are accessed by users and , but not by and . Similar ideas have been successfully applied to image retrieval to improve the accuracy of similarity measurement [12], =-=[13]-=-, [25]. (1) Let be the number of content features. The summation in (1) is rewritten as follows: B. Dynamic Weighting Schemes 1) Introduction: A simple approach capable of combining acoustic features  similarity measures can be cast as a learning problem aimed to assign approximate weights to each feature [39]. To automatically determine the weights for audio features, the metric learning approach =-=[13]-=-, [40], which learns appropriate similarity metrics based on the correlation between acoustic features and user access patterns of music, needs to be explored. Given that human perception of music is   </text>
</top>
<top>
<paper_num> 192 </paper_num>
<paper_title>   Local Discriminant Wavelet Packet Coordinates for Face Recognition.  </paper_title>
<paper_abstract>   Face recognition is a challenging problem due to variations in pose, illumination, and expression. Techniques that can provide effective feature representation with enhanced discriminability are crucial. Wavelets have played an important role in image processing for its ability to capture localized spatial-frequency information of images. In this paper, we propose a novel local discriminant coordinates method based on wavelet packet for face recognition to compensate for these variations. Traditional wavelet-based methods for face recognition select or operate on the most discriminant subband, and neglect the scattered characteristic of discriminant features. The proposed method selects the most discriminant coordinates uniformly from all spatial frequency subbands to overcome the deficiency of traditional wavelet-based methods. To measure the discriminability of coordinates, a new dilation invariant entropy and a maximum a posterior logistic model are put forward. Moreover, a new triangle square ratio criterion is used to improve classification using the Euclidean distance and the cosine criterion. Experimental results show that the proposed method is robust for face recognition under variations in illumination, pose and expression.  </paper_abstract>
<query_num> 19201 </query_num>
<text>   n method which largely relies on the representation of the training samples. On the other hand, wavelet-based methods with no special focus on the training data have been used for feature extraction (=-=Mallat, 1989; Coifman et al., 1992-=-). The decomposition of the data into different frequency ranges allows us to isolate the frequency components introduced by intrinsic deformations due to expression or extrinsic   </text>
<query_num> 19202 </query_num>
<text>   nce, the relative entropy, or the JDivergence. Then φ1(B j) will be a measure of efficacy of the subband B j for classification, and local discriminant basis are selected by the best-basis algorithm (=-=Coifman and Wicherhauser, 1992-=-) using the following criterion: Ψ = argmax B j ∈D φ1(B j). (4) The final step is to construct traditional discriminant analysis (e.g., LDA, CT) with features derived from the LDB feature extraction.  re function will get a contrastive term to ensure not only the within-class difference is low, but also the betweenclass difference is large. Our LDC algorithm does not need the best-basis algorithm (=-=Coifman and Wicherhauser, 1992-=-) used in LDB, it ensures that we can select the most discriminant features without any impact of the best-basis algorithm. Subsequently, the LDC algorithm uses the complete linear discriminant analys   </text>
</top>
<top>
<paper_num> 193 </paper_num>
<paper_title>   A Framework for Classifying and Comparing Software Architecture Evaluation Methods.  </paper_title>
<paper_abstract>   Software architecture evaluation has been proposed as a means to achieve quality attributes such as maintainability and reliability in a system. The objective of the evaluation is to assess whether or not the architecture will lead to the desired quality attributes. Recently, there have been a number of evaluation methods proposed. There is, however, little consensus on the technical and non-technical issues that a method should comprehensively address and which of the existing methods is most suitable for a particular issue. This paper presents a set of commonly known but informally described features of an evaluation method and organizes them within a framework that should offer guidance on the choice of the most appropriate method for an evaluation exercise. In this paper, we use this framework to characterise eight SA evaluation methods.  </paper_abstract>
<query_num> 19301 </query_num>
<text>   e methods will also adopt the same definition as the working definition. However, this is not a very prudent approach when it is generally acknowledged fact that there is no standard definition of SA =-=[30]-=-. 3.3 Process support SA evaluation requires a number of activities to be performed. It needs a number of inputs, generates important artefacts, and requires stakeholders’ participation. Like any proc   </text>
<query_num> 19302 </query_num>
<text>   enariobased Architecture Analysis Method (SAAM) [6], Architecture Tradeoff Analysis Method (ATAM) [7], Active Reviews for Intermediate Design (ARID) [8], SAAM for Evolution and Reusability (SAAMER) 1 =-=[9]-=-, Architecture-Level Modifiability Analysis (ALMA) [10] , Architecture-Level Prediction of Software Maintenance (ALPSM) 1 [11], Scenario-Based Architecture Reengineering (SBAR) 1 [12], SAAM for Comple   </text>
<query_num> 19303 </query_num>
<text>   informally classified evaluation approaches. Their categories of assessment techniques include scenario-based, simulation or prototyping, mathematical modelling, and experiencebased. Bosch and Molin =-=[44]-=- have classified quality attributes into two main categories: development (e.g. maintainability, flexibility) and operational (e.g. performance, reliability). Scenario-based and experiencedbased techn   </text>
<query_num> 19304 </query_num>
<text>   maturation, we have conjectured that existing evaluation methods may be in any of the following four maturity phases: 1) inception (method recently surfaced and has not been validated yet for example =-=[27]-=-), 2) development (methods is continuously being validated with results appearing in credible literature), 3) refinement (method has been validated in various domains and refinements are continuously   </text>
<query_num> 19305 </query_num>
<text>   rchitecture Tradeoff Analysis Method (ATAM) [7], Active Reviews for Intermediate Design (ARID) [8], SAAM for Evolution and Reusability (SAAMER) 1 [9], Architecture-Level Modifiability Analysis (ALMA) =-=[10]-=- , Architecture-Level Prediction of Software Maintenance (ALPSM) 1 [11], Scenario-Based Architecture Reengineering (SBAR) 1 [12], SAAM for Complex Scenarios (SAAMCS) 1 [13], and integrating SAAM in do ents. Abowd et al. [5] also provided a four dimensional framework to further categorise them. The four dimensions are generality, level of detail, phase, and what is evaluated. Bosch [32] and PerOlof =-=[10]-=- have also informally classified evaluation approaches. Their categories of assessment techniques include scenario-based, simulation or prototyping, mathematical modelling, and experiencebased. Bosch   </text>
<query_num> 19306 </query_num>
<text>   s (ADLs) with little consensus on a universally accepted definition of ADL. A common understanding is that an ADL is a formal language used to represent the architecture of a softwareintensive system =-=[21]-=-. Since architectural description has a vital role in the architecture assessment process, an evaluation method should guide its user on which ADL is most appropriate to be used. The role of architect   </text>
<query_num> 19307 </query_num>
<text>   s of a stakeholder. The importance of multiple views as an effective means of separating the concerns during architectural design and analysis has been realised by SA researchers and practitioners in =-=[20, 41, 42]-=-. However, there are different opinions on the number and nature of the architectural views (for example Hofmeiser et al. [41] suggest conceptual, module, execution, and code views; Kruchten [42] gave   </text>
<query_num> 19308 </query_num>
<text>   s of the evaluation methods and their usefulness. Two research groups have attempted to address informally some of the above mentioned issues by providing a conceptual framework for method comparison =-=[1, 18]-=-. Each of these efforts enhances our understanding of the available methods; however, for reasons mentioned in section 2, each has significant limitations. Therefore, 1 The abbreviated names are not w  other previous surveys. We have made every effort to find and examine all the survey work done on SA evaluation methods during the last decade. To the best of our knowledge there is only one attempt =-=[18]-=- that provides a comprehensive treatment of the topic. None of the other published surveys or comparison of SA evaluation methods provides an explicit framework for comparing the methods. Instead quit ir comparison framework does not include some important features that an evaluation method should have, e.g., architecture definition, architectural description, tool support, and so forth. We regard =-=[18]-=- as a first comprehensive attempt to provide a taxonomy of this growing area of research and practice; and it is quite valuable in better comprehending the existing SA evaluation methods. However, thi searchers and practitioners. To identify the components of our framework, we have also drawn upon a number of other sources including previously developed comparison frameworks for evaluation methods =-=[1, 18]-=-, an extensive survey of SA literature, and an analysis of heuristics of experienced software architects and software engineers. We do not claim that we have produced an exhaustive list of features th   </text>
<query_num> 19309 </query_num>
<text>   uding he himself. A later classification of software quality is provided in [24].sThere are a number of definitions of scenarios as well. We will use as our working definition of scenario provided in =-=[25]-=-: ���A scenario is a brief description of a single interaction of a stakeholder with a system.” The following subsections motivate and elaborate on each of the components (given in right hand-side of Ta pproaches in the industry [49]. That is why the SA community has been emphasising the need to automate as many of the tedious, expensive, and error-prone tasks of SA design and evaluation as possible =-=[25]-=-. A tool can also capture the design artefacts along with the decision rationale, evaluation outcomes, measurement and administrative information that are invaluable assets for future. Despite the rec   </text>
<query_num> 19310 </query_num>
<text>   y related issues at the SA level. The SA evaluation methods specifically studied in this paper are: Scenariobased Architecture Analysis Method (SAAM) [6], Architecture Tradeoff Analysis Method (ATAM) =-=[7]-=-, Active Reviews for Intermediate Design (ARID) [8], SAAM for Evolution and Reusability (SAAMER) 1 [9], Architecture-Level Modifiability Analysis (ALMA) [10] , Architecture-Level Prediction of Softwar  conflict between configurability and performance [38]; performance also impacts modifiability, availability affects safety, security conflicts with usability, and each quality attribute impacts cost =-=[7]-=-. That is why it is important to find an appropriate balance of quality attributes in order to develop a successful product. One of the most significant features of method differentiation and classifi   </text>
</top>
<top>
<paper_num> 194 </paper_num>
<paper_title>   Minimum Bayes-Risk Decoding for Statistical Machine Translation.  </paper_title>
<paper_abstract>   We present Minimum Bayes-Risk (MBR) decoding over translation lattices that compactly encode a huge number of translation hypotheses. We describe conditions on the loss function that will enable efficient implementation of MBR decoders on lattices. We introduce an approximation to the BLEU score (=-=Papineni et al., 2001-=-) that satisfies these conditions. The MBR decoding under this approximate BLEU is realized using Weighted Finite State Automata. Our experiments show that the Lattice MBR decoder yields moderate, consistent gains in translation performance over N-best MBR decoding on Arabicto-English, Chinese-to-English and Englishto-Chinese translation tasks. We conduct a range of experiments to understand why Lattice MBR improves upon N-best MBR and study the impact of various parameters on MBR performance. 1  </paper_abstract>
<query_num> 19401 </query_num>
<text>   al 2 iterations of Model-4 are performed for zhen and enzh pairs. Word Alignments in both source-to-target and target-to-source directions are obtained using the Maximum A-Posteriori (MAP) framework (=-=Matusov et al., 2004-=-). An inventory of phrase-pairs up to length 5 is then extracted from the union of source-target and target-source alignments. Several feature functions are then computed over the phrasepairs. 5-gram   </text>
<query_num> 19402 </query_num>
<text>   e approach has been shown to give improvements over the MAP classifier in many areas of natural language processing including automatic speech recognition (=-=Goel and Byrne, 2000-=-), machine translation (=-=Kumar and Byrne, 2004; Zhang and Gildea, 2008-=-), bilingual word alignment (=-=Kumar and Byrne, 2002-=-), and parsing (=-=Goodman, 1996; Titov and Henderson, 2006; Smith and Smith, 2007-=-). In statistical machine translation, MBR deco n such a loss function L(E, E ′ ) between an automatic translation E ′ and the reference E, and an underlying probability model P (E|F ), the MBR decoder has the following form (=-=Goel and Byrne, 2000; KumaKumar and Byrne, 2004-=-): Ê = argmin E ′ R(E ∈E ′ ) = argmin E ′ ∑ L(E, E ∈E ′ )P (E|F ), E∈E where R(E ′) denotes the Bayes risk of candidate translation E ′ under the loss function L. If the loss function between any two   </text>
<query_num> 19403 </query_num>
<text>   e number of translation hypotheses. We describe conditions on the loss function that will enable efficient implementation of MBR decoders on lattices. We introduce an approximation to the BLEU score (=-=Papineni et al., 2001-=-) that satisfies these conditions. The MBR decoding under this approximate BLEU is realized using Weighted Finite State Automata. Our experiments show that the Lattice MBR decoder yields moderate, con osteriori (MAP) decision rule which optimizes the 0-1 loss function. In contrast, these systems are evaluated using metrics based on string-edit distance (Word Error Rate), ngram overlap (BLEU score (=-=Papineni et al., 2001-=-)), or precision/recall relative to human annotations. Minimum Bayes-Risk (MBR) decoding (=-=Bickel and Doksum, 1977-=-) aims to address this mismatch by selecting the hypothesis that minimizes the expected  Kumar and Byrne (2004) show that MBR decoding gives optimal performance when the loss function is matched to the evaluation criterion; in particular, MBR under the sentence-level BLEU loss function (=-=Papineni et al., 2001-=-) gives gains on BLEU. This is despite the fact that the sentence-level BLEU loss function is an approximation to the exact corpus-level BLEU. A different MBR inspired decoding approach is pursued in   syntaxbased systems would work in this framework. We will introduce conditions on the loss functions that can be incorporated in Lattice MBR decoding. We describe an approximation to the BLEU score (=-=Papineni et al., 2001-=-) that will satisfy these conditions. Our Lattice MBR decoding is realized using Weighted Finite State Automata. We expect Lattice MBR decoding to improve upon N-best MBR primarily because lattices co  be rewritten in terms of a gain function G(E, E ′) = Lmax − L(E, E ′): Ê = argmax E ′ ∈E ∑ G(E, E ′ )P (E|F ). (1) E∈E We are interested in performing MBR decoding under a sentence-level BLEU score (=-=Papineni et al., 2001-=-) which behaves like a gain function: it varies between 0 and 1, and a larger value reflects a higher similarity. We will therefore use Equation 1 as the MBR decoder. We note that E represents the spa   </text>
<query_num> 19404 </query_num>
<text>   hat there is no scaling. This is an important parameter that needs to be tuned on a development set. There has been prior work in MBR speech recognition and machine translation (=-=Goel and Byrne, 2000; Ehling et al., 2007-=-) which has shown the need for tuning this factor. Our MT system parameters are trained with Minimum Error Rate Training which assigns a very high posterior probability to the MAP translation. As a re   </text>
<query_num> 19405 </query_num>
<text>   n (=-=Goel and Byrne, 2000-=-), machine translation (=-=Kumar and Byrne, 2004; Zhang and Gildea, 2008-=-), bilingual word alignment (=-=Kumar and Byrne, 2002-=-), and parsing (=-=Goodman, 1996; Titov and Henderson, 2006; Smith and Smith, 2007-=-). In statistical machine translation, MBR decoding is generally implemented by re-ranking an N-best list of translations produced by a first-pass decoder; this list typically contains between 100 and   </text>
<query_num> 19406 </query_num>
<text>   of the number of times the ngram occurs in the input (#w(E)), the n-gram factor θw from Equation 6, and the posterior probability p(w|E). The automaton corresponds to the weighted regular expression (=-=Karttunen et al., 1996-=-): ¯w(w/(θwp(w|E)) ¯w) ∗ . We successively intersect each of these automata with an automaton that begins as an unweighted copy of the lattice Eh. This automaton must also incorporate the factor θ0 of   </text>
<query_num> 19407 </query_num>
<text>   parallel and monolingual data consists of all the allowed training sets in the constrained track. For each language pair, we use two development sets: one for Minimum Error Rate Training (=-=Och, 2003; Macherey et al., 2008-=-), and the other for tuning the scale factor for MBR decoding. Our development sets consists of the NIST 2004/2003 evaluation sets for both aren and zhen, and NIST 2006 (NIST portion)/2003 evaluation   </text>
<query_num> 19408 </query_num>
<text>   sing including automatic speech recognition (=-=Goel and Byrne, 2000-=-), machine translation (=-=Kumar and Byrne, 2004; Zhang and Gildea, 2008-=-), bilingual word alignment (=-=Kumar and Byrne, 2002-=-), and parsing (=-=Goodman, 1996; Titov and Henderson, 2006; Smith and Smith, 2007-=-). In statistical machine translation, MBR decoding is generally implemented by re-ranking an N-best list of translations produced by a first-pass dec   </text>
<query_num> 19409 </query_num>
<text>   weights of all paths in the lattices Ew and E respectively. 4 WFSA MBR Computations We now show how the Lattice MBR Decision Rule (Equation 6) can be implemented using Weighted Finite State Automata (=-=Mohri, 1997-=-). There are four steps involved in decoding starting from weighted finite-state automata representing the candidate outputs of a translation system. We will describe these 2 in the log semiring, wher   </text>
<query_num> 19410 </query_num>
<text>   wn to give improvements over the MAP classifier in many areas of natural language processing including automatic speech recognition (=-=Goel and Byrne, 2000-=-), machine translation (=-=Kumar and Byrne, 2004;Zhang and Gildea, 2008-=-), bilingual word alignment (=-=Kumar and Byrne, 2002-=-), and parsing (=-=Goodman, 1996; Titov and Henderson, 2006; Smith and Smith, 2007-=-). In statistical machine translation, MBR decoding is generally implem   </text>
</top>
<top>
<paper_num> 195 </paper_num>
<paper_title>   Feature-based similarity search in graph structures.  </paper_title>
<paper_abstract>   Similarity search of complex structures is an important operation in graph-related applications since exact matching is often too restrictive. In this article, we investigate the issues of substructure similarity search using indexed features in graph databases. By transforming the edge relaxation ratio of a query graph into the maximum allowed feature misses, our structural filtering algorithm can filter graphs without performing pairwise similarity computation. It is further shown that using either too few or too many features can result in poor filtering performance. Thus the challenge is to design an effective feature set selection strategy that could maximize the filtering capability. We prove that the complexity of optimal feature set selection is Ω(2 m) in the worst case, where m is the number of features for selection. In practice, we identify several criteria to build effective feature sets for filtering, and demonstrate that combining features with similar size and selectivity can improve the filtering and search performance significantly within a multi-filter composition framework. The proposed feature-based filtering concept can be generalized and applied to searching approximate non-consecutive sequences, trees, and other structured data as well. Categories and Subject Descriptors: H.2.4 [Database Management]: Systems – Query process-This is a preliminary release of an article accepted by ACM Transactions on Database Systems. The definitive version is currently in production at ACM and, when released, will supersede this version.  </paper_abstract>
<query_num> 19501 </query_num>
<text>   h wildcards, i.e., don’t care nodes and edges. Different from their similarity model, we do not fix the positions of wildcards, thus allowing a general and flexible search scheme. In our recent work [=-=Yan et al. 2005-=-], we introduced the basic concept of a featurebased indexing and filtering methodology for substructure similarity search. It was shown that using either too few or too many features can result in po   </text>
<query_num> 19502 </query_num>
<text>   m, researchers also studied the approximate tree search problem. Wang et al. [=-=Wang et al. 1994-=-] designed an interactive system that allows a user to search inexact matchings of trees. Kailing et al. [=-=Kailing et al. 2004-=-] presented new filtering methods based on tree height, node degree and label information. ACM Transactions on Database Systems, Vol. V, No. N, June 2006.s6 · Xifeng Yan et al. The structural filterin   </text>
<query_num> 19503 </query_num>
<text>   ol. V, No. N, June 2006.sFeature-based Similarity Search in Graph Structures · 3 ilarity search: find structures that are similar to the query graph [=-=Petrakis and Faloutsos 1997, Willett et al. 1998, Raymond et al. 2002-=-]. These kinds of queries are very useful. For example, in substructure search, a user may not know the exact composition of the full structure he wants, but requires that it contain a set of small fu s been studied in various fields. Willett et al. [=-=Willett et al. 1998-=-] summarized the techniques of fingerprint-based and graph-based similarity search in chemical compound databases. Raymond et al. [=-=Raymond et al. 2002-=-] proposed a three-tier algorithm for full structure similarity search. Recently, substructure search has attracted lots of attention in the database research community. Shasha et al. [=-=Shasha et al-=-. 2 onal to the size of the candidate set. Quite a lot of work has been done at calculating the pairwise substructure similarity. Readers are referred to the related work in [=-=Nilsson 1980, Hagadone 1992,Raymond et al. 2002-=-]. In the step of feature miss estimation, we calculate the number of features in the query graph. One feature may have multiple embeddings in a graph; thus, we use the number of embeddings of a featu  a single filter: one using individual edges as features (denoted as Edge) and the other using all features of a query graph (denoted as Allfeature). Many similarity search algorithms [=-=Hagadone 1992,Raymond et al. 2002-=-] can only apply the edge-based filtering mechanism since the mapping between edge deletion/relabeling and feature misses was not established before this study. In fact, the edge-based filtering appro hile Grafil can prune 99.7%. If a user wants to check whether there are real matches in the remaining 0.3% of the dataset, they can apply the similarity computation tools developed in [=-=Hagadone 1992,Raymond et al. 2002-=-] to check them. If they are not satisfied with the result, the user can relax the edge loss to 3. The Edge approach will return 18% of the dataset and Grafil will return 11% of the dataset. The runni   </text>
<query_num> 19504 </query_num>
<text>   research. Due to the complexity of graph data and the diversity of their applications, graphs are generally key entities in widely used databases in chem-informatics and bioinformatics, such as PDB [=-=Berman et al. 2000-=-] and KEGG [=-=Kanehisa and Goto 2000-=-]. In chemistry, the structures and properties of newly discovered or synthesized chemical molecules are studied, classified, and recorded for scientific and commerci   </text>
<query_num> 19505 </query_num>
<text>   rro [=-=Navarro 2001-=-]. The well-known q-gram method was initially developed by Ullmann [=-=Ullmann 1977-=-]. Ukkonen [=-=Ukkonen 1992-=-] independently discovered the q-gram approach, which was further extended in [=-=Gravano et al. 2001-=-] against large scale sequence databases. These q-gram algorithms work for consecutive sequences, not structures. Our work generalized the q-gram method to fit structural patterns of various sizes. 3.   </text>
<query_num> 19506 </query_num>
<text>   sign and other scientific activities. Nevertheless, the usage of a graph database as well as its query system is not confined to chemical informatics only. In computer vision and pattern recognition [=-=Petrakis and Faloutsos 1997, Messmer and Bunke 1998, Beretti et al. 2001-=-], graphs are used to represent complex structures such as hand-drawn symbols, fingerprints, 3D objects, and medical images. Researchers extract graph mode hemidplus. ACM Transactions on Database Systems, Vol. V, No. N, June 2006.sFeature-based Similarity Search in Graph Structures · 3 ilarity search: find structures that are similar to the query graph [=-=Petrakis and Faloutsos 1997, Willett et al. 1998, Raymond et al. 2002-=-]. These kinds of queries are very useful. For example, in substructure search, a user may not know the exact composition of the full structure he wants, but   </text>
<query_num> 19507 </query_num>
<text>   tics, such as clustering-based feature set selection developed in our solution. Besides the full-scale graph search problem, researchers also studied the approximate tree search problem. Wang et al. [=-=Wang et al. 1994-=-] designed an interactive system that allows a user to search inexact matchings of trees. Kailing et al. [=-=Kailing et al. 2004-=-] presented new filtering methods based on tree height, node degree and lab   </text>
<query_num> 19508 </query_num>
<text>   tion for two structures. Hagadone [=-=Hagadone 1992-=-] recognized the importance of substructure similarity search in a large set of graphs. He used the atom and edge label to do screening. Holder et al. [=-=Holder et al. 1994-=-] adopted the principle of minimum description length for approximate graph matching. Messmer and Bunke [=-=Messmer and Bunke 1998-=-] studied the reverse substructure similarity search problem in computer   </text>
<query_num> 19509 </query_num>
<text>   two graphs, which is often costly to compute. However, since this measure takes structure connectivity fully into consideration, it is more accurate than the feature-based measure. Bunke and Shearer [=-=Bunke and Shearer 1998-=-] used the maximum common subgraph to measure full structure similarity. Researchers also developed the concept of graph edit distance by simulating the graph matching process in a way similar to the   </text>
</top>
<top>
<paper_num> 196 </paper_num>
<paper_title>   Best Effort Query Processing in DHT-based P2P Systems.  </paper_title>
<paper_abstract>   Structured P2P systems in the form of distributed hash tables (DHT) are a promising approach for building massively distributed data management platforms. However, for many applications the supported key lookup queries are not sufficient. Instead, techniques for managing and querying (relational) structured data are required. In this paper, we argue that in order to cope with the dynamics in large-scale P2P systems such query techniques should be work in a best effort manner. We describe such operations (namely grouping/aggregation, similarity and nearest neighbor search) and discuss appropriate query evaluation strategies. 1.  </paper_abstract>
<query_num> 19601 </query_num>
<text>   (e.g. relational data) using a DHT and second to process queries consisting of operations such as join, range selections and grouping / aggregation. These challenges were recently addressed, e.g. in =-=[11, 20]-=- by distributing tuples in a DHT and implement relational algebra operators exploiting the hashing scheme. However, the approaches lack the comprehensive consideration of three important problems: • B ject. [11] presents a system architecture based on CAN, and discusses several distributed join strategies. A detailed discussion of query processing strategies for Chord-based P2P systems is given in =-=[20]-=-. Here, the authors propose a framework of operators and their implementation by using modified Chord search algorithms. Another P2Pbased query processing approach is AmbientDB [4]. It exploits Chord-   </text>
<query_num> 19602 </query_num>
<text>   (e.g. relational data) using a DHT and second to process queries consisting of operations such as join, range selections and grouping / aggregation. These challenges were recently addressed, e.g. in =-=[11, 20]-=- by distributing tuples in a DHT and implement relational algebra operators exploiting the hashing scheme. However, the approaches lack the comprehensive consideration of three important problems: • B mmunity has started to adapt techniques from distributed and parallel database systems [5, 10] to (rather simple) DHTs. An approach with a stronger relation to modern DHT systems is the PIER project. =-=[11]-=- presents a system architecture based on CAN, and discusses several distributed join strategies. A detailed discussion of query processing strategies for Chord-based P2P systems is given in [20]. Here   </text>
<query_num> 19603 </query_num>
<text>   al static query optimization and execution techniques can lead to less effective results in such environments. Therefore, we have investigated a more adaptive approach by following the idea of eddies =-=[3]-=-. An eddy is a pipeline for routing tuples to operators of the plan based on runtime statistics such as queue length, selectivity or load. In our implementation an instance of an eddy can run at each   </text>
<query_num> 19604 </query_num>
<text>   ent DHT appliations are distributed search engines, directories for the semantic web or genome data management (cf. [7] for an exhaustive list). Examples of DHT are Content-Addressable Networks (CAN) =-=[14]-=-, Chord [18], Pastry [15] or P-Grid [1]. The proposals mainly differ in the topic of the key space and the contact selection, i.e., how to distribute the (key,value)-pairs among the peers and which ar egies for query processing. Similar work is described in [9] based on locality sensitive hashing. 3. CAN &amp; Data Fragmentation A variant of distributed hash tables are ContentAddressable Networks (CAN =-=[14]-=-). Each CAN node is responsible for a certain part of the key space, its zone. This means, the node stores all (key, value)-pairs whose keys fall into its zone. The key space is a torus of Cartesian c   </text>
<query_num> 19605 </query_num>
<text>   esigned to perform operations of a hashtable, e.g., put(key, value) or value = get(key). But soon the database community has started to adapt techniques from distributed and parallel database systems =-=[5, 10]-=- to (rather simple) DHTs. An approach with a stronger relation to modern DHT systems is the PIER project. [11] presents a system architecture based on CAN, and discusses several distributed join strat   </text>
<query_num> 19606 </query_num>
<text>   h allows heuristics such as “selection first”), • input queue length (each operator contains an input queue; the length of this queue is a measure for processing costs), • selectivity (as proposed in =-=[19]-=- the selectivity of an operator is learned using a ticket approach), and • next-join (if tuples have to be reinserted e.g. for performing a join, one can choose the join first for which the distance o   </text>
<query_num> 19607 </query_num>
<text>   iations are distributed search engines, directories for the semantic web or genome data management (cf. [7] for an exhaustive list). Examples of DHT are Content-Addressable Networks (CAN) [14], Chord =-=[18]-=-, Pastry [15] or P-Grid [1]. The proposals mainly differ in the topic of the key space and the contact selection, i.e., how to distribute the (key,value)-pairs among the peers and which are the nodes   </text>
<query_num> 19608 </query_num>
<text>   idered. Instead, one can choose only a subset of nonoverlapping q-grams, i.e. for a threshold e only e + 1 q-grams are required [12], where q-grams with a high selectivity are chosen for preselection =-=[17]-=-. 2. For each q-sample si send a selection request to the peer responsible for the zone containing the point h(R, si). This is part of the basic functionality of the DHT. 3. Each receiving peer whose   </text>
<query_num> 19609 </query_num>
<text>   istributed search engines, directories for the semantic web or genome data management (cf. [7] for an exhaustive list). Examples of DHT are Content-Addressable Networks (CAN) [14], Chord [18], Pastry =-=[15]-=- or P-Grid [1]. The proposals mainly differ in the topic of the key space and the contact selection, i.e., how to distribute the (key,value)-pairs among the peers and which are the nodes a peer commun   </text>
<query_num> 19610 </query_num>
<text>   iven similarity threshold is reached the tuple is returned as a result of the similarity search. Another possible approach is to sent all q-grams and apply a count filtering technique as described in =-=[6]-=-. Nearest Neighbor Search. In order to be able to process NN queries, the keys of the tuples stored in the DHT have to be generated with a hash function that preserves the neighborhood between similar   </text>
<query_num> 19611 </query_num>
<text>   more sophisticated query algebra. In order to address the problem of range operations, [2] proposes space filling curves and DHT flooding strategies for query processing. Similar work is described in =-=[9]-=- based on locality sensitive hashing. 3. CAN &amp; Data Fragmentation A variant of distributed hash tables are ContentAddressable Networks (CAN [14]). Each CAN node is responsible for a certain part of th   </text>
<query_num> 19612 </query_num>
<text>   ode is addressed with a subtree-ID. For each level of the tree, each node maintains a reference to another peer in the same subtree, but whose ID branches to a different subtree in the deeper levels. =-=[8]-=- features a detailed survey of the various approaches. Our DHTbased query processor is applicable to any DHT. But key space topologies different from CAN require an adoption of the data fragmentation   </text>
</top>
<top>
<paper_num> 197 </paper_num>
<paper_title>   Optimal Algorithm for Finding DNA Motifs with Nucleotide Adjacent Dependency.  </paper_title>
<paper_abstract>   Finding motifs and the corresponding binding sites is a critical and challenging problem in studying the process of gene expression. String and matrix representations are two popular models to represent a motif. However, both representations share an important weakness by assuming that the occurrence of a nucleotide in a binding site is independent of other nucleotides. More complicated representations, such as HMM or regular expression, exist that can capture the nucleotide dependency. Unfortunately, these models are not practical (with too many parameters and require many known binding sites). Recently, Chin and Leung introduced the SPSP representation which overcomes the limitations of these complicated models. However, discovering novel motifs in SPSP representation is still a NP-hard problem. In this paper, based on our observations in real binding sites, we propose a simpler model, the Dependency Pattern Sets (DPS) representation, which is simpler than the SPSP model but can still capture the nucleotide dependency. We develop a branch and bound algorithm (DPS-Finder) for finding optimal DPS motifs. Experimental results show that DPS-Finder can discover a length-10 motif from 22 length-500 DNA sequences within a few minutes and the DPS representation has a similar performance as SPSP representation.  </paper_abstract>
<query_num> 19701 </query_num>
<text>   (l 2 ) motifs for a length-l substring) length-O(X) sequences (input sequences with some positions represented by brackets) to represent the O(Xl 2 ) possible motifs. A l-factor tree is a suffix tree =-=[23]-=- of height l where each path from the root to a leaf represents a length-l substring occurring in the input sequence. Figure 1 shows a factor tree of height-8 for the sequence “CA(…)(…)GGATGGCA(…)(…)G   </text>
<query_num> 19702 </query_num>
<text>   independent. When there are sufficient number of known binding sites of a transcription factor, people can use some complex representations, e.g. the hidden Markov model (HMM) [24], Bayesian network =-=[3]-=- or enhanced PWM [9], to represent nucleotide interdependence. However, when we want to discover novel motif or describe a motif with only a few known binding sites, the input data may not contain eno   </text>
<query_num> 19703 </query_num>
<text>   motif with only a few known binding sites, the input data may not contain enough information for deriving the hidden motif. Chin and Leung overcame the problem by introducing the SPSP representation =-=[7]-=-, a generalized model of string representation and matrix representation, that can model the adjacent dependency of nucleotides with much less parameters than HMM and regular expression. Since the SPS on that the hidden motif should have an unexpectedly large number of binding sites, a motif P with small p2svalue is likely to be the hidden motif. The p-value of a motif can be calculated as follows =-=[7]-=-. Let B be the background model for the non-binding region of the DNA sequences T and B(σ) be the probability that a length-l string σ occurs in a particular position in T. B can be a Markov Chain or  ard pattern sets, many candidate motifs can be pruned out. 4 Experimental Results We compared the performance of some popular motif discovering algorithms, i.e. Weeder [19], MEME [13] and SPSP-Finder =-=[7]-=-, with DSP-Finder on the yeast data set in SCPD [25]. SCPD contains information of the motif patterns and the binding sites for a set of transcription factors of yeast. For each transcription factor,   </text>
<query_num> 19704 </query_num>
<text>   the top-k patterns in its wildcard pattern sets, many candidate motifs can be pruned out. 4 Experimental Results We compared the performance of some popular motif discovering algorithms, i.e. Weeder =-=[19]-=-, MEME [13] and SPSP-Finder [7], with DSP-Finder on the yeast data set in SCPD [25]. SCPD contains information of the motif patterns and the binding sites for a set of transcription factors of yeast.   </text>
<query_num> 19705 </query_num>
<text>   ∗ The research was supported in parts by the RGC grant HKU 7120/06E. 1s2 and PSSMs is huge, which consists of 4l real numbers, and thus, algorithms generally either produce a sub-optimal motif matrix =-=[2,8,12,13]-=- or take too long to run when the motif is longer than 10 [15]. However, both the string and the matrix representations share an important common weakness: they assume that the occurrence of each nucl   </text>
</top>
<top>
<paper_num> 198 </paper_num>
<paper_title>   Interactive rendering of translucent deformable objects.  </paper_title>
<paper_abstract>   Realistic rendering of materials such as milk, fruits, wax, marble, and so on, requires the simulation of subsurface  scattering of light. This paper presents an algorithm for plausible reproduction of subsurface scattering effects.  </paper_abstract>
<query_num> 19801 </query_num>
<text>   .al 15, 13 . The shade of a surface point xo on a translucent object, observed from a direction ωo, is computed with the following integral: L → � � (xo,ωo) = L ← (xi,ωi)S(xi,ωi;xo,ωo)(ωi · Ni)dωidxi =-=(1)-=- S Ω + (xi) L → (xo,ωo) and L ← (xi,ωi) denote exitant/incident radiance respectively. S is the object’s surface, Ω+(xi) is the hemisphere above xi in normal direction Ni, and S(xi,ωi;xo,ωo) is the bi   </text>
<query_num> 19802 </query_num>
<text>   1 1 � √ = 1 − S2 +t 2 s0 1 � 2t0Δt 2 s2 + 0 � Δt � � � 2 + ... (14) s0 1 R2 1 = +t 2 r2 � � 2t0Δt 1 − 0 r2 + 0 � Δt � � � 2 + ... (15) r0 e −σs = e −σs0 � ��� 2t0Δt 1 − σs0 s2 + ( 0 Δt ) s0 2� � =-=+ ... (16) H-=-ere, s0 := s(t0) and r0 := r(t0). The product of the Taylor approximations results in a simple polynomial and its integration is straightforward: I2 = zRe−σs0 r2 0s0 � � L 1 + ... . (17) The higher  n the sum is positive if it is a back facing edge for xo, otherwise it is negative. 4.1. Error Analysis The error due to ignoring second and higher order terms in the Taylor series expansions (14) to =-=(16)-=- is straightforward to analyze if one makes sure that the first order term is sufficiently smaller than 1. The series expansions converge rapidly in that case. F( A,x) 0.6 0.5 0.4 0.3 0.2 0.1 Form Fac   </text>
<query_num> 19803 </query_num>
<text>   We solve the integral in polar coordinates in A’s supporting plane, using the orthogonal projection of d onto this plane, called d ′ , as the pole: � θmax � rmax(θ) I = z(1 + σs) rmin(θ) e−σs r dr dθ =-=(9)-=- s3 θmin Consider, to start with, the case that one of A’s vertices coincides with d ′ so that rmin(θ) = 0 for all polar c○ The Eurographics Association 2003. θ 2 τ supporting plane of triangle d sing   </text>
<query_num> 19804 </query_num>
<text>   emonstrated to work for participating media, but it assumes static models. We show that our goal can be reached with a hierarchical boundary element method to solve the subsurface scattering integral =-=(4)-=- of §2, in the spirit of hierarchical radiosity with clustering 9, 28, 26, 31 . An outline of the method is given in §3. The success of the method is in part due to an efficient and accurate semi-anal ring reflectance function: S(xi,ωi;xo,ωo) = 1 π Ft(η,ωo)Rd(xi;xo)Ft(η,ωi). (2) Substituting this into Equation 1, we get: L → (xo,ωo) = 1 Ft(η,ωo)B(xo) π � (3) B(xo) = E(xi) = E(xi)Rd(xi,xo)dxi S � Ω =-=(4)-=- + L (xi) ← (xi,ωi)Ft(η,ωi)(Ni · ωi)dωi (5) In order to render translucent objects efficiently, one needs an efficient way to solve this equation at every surface point. 2.2. Dipole Source Approximati de on a model for the BSSRDF, and thus a model for Rd. This model has to fulfill two main criteria: it should be adaptable to different materials, and should provide an efficient solution of equation =-=(4)-=-. An accurate method to determine Rd(xi;xo) is to use a full simulation. Many different techniques have been proposed (coming from the area of participating media), e.g. 8, 26, 19, 14 . While these te as well as a hierarchical evaluation algorithm allowing interactive rendering speeds. 3. Outline of the New Method In this section, we outline a hierarchical boundary element method to solve equation =-=(4)-=- efficiently, in the style of hierarchical radiosity with clustering. Before elaborating the details of the method in next sections, we compare with related work at the end of this section. 3.1. Discr n calculations. If we break the surfaces of our object to be rendered into regions Ak, and if we assume small lighting variation over each region, making E(xi) = Ek constant for all xi ∈ Ak, equation =-=(4)-=- reduces to the following sum: Mertens et al. / Interactive Rendering of Translucent Deformable Objects B(xo) = ∑EkF(Ak,xo) k � (6) F(Ak,xo) = Rd(xi,xo)dxi (7) We compute the factors F(A,xo) for the m   </text>
<query_num> 19805 </query_num>
<text>   er each region, making E(xi) = Ek constant for all xi ∈ Ak, equation (4) reduces to the following sum: Mertens et al. / Interactive Rendering of Translucent Deformable Objects B(xo) = ∑EkF(Ak,xo) k � =-=(6)-=- F(Ak,xo) = Rd(xi,xo)dxi (7) We compute the factors F(A,xo) for the midpoint of each element w.r.t. all other elements. The irradiance Ek is also evaluated at the midpoints. The sum (6) can then be us   </text>
<query_num> 19806 </query_num>
<text>   ition is stronger than the previ� ous one since r0 ≤ s0 = r2 0 + h2 . For the last factor we need to take |Δt|σ &amp;lt; 1 5 . With L = 2|Δt|, we have the following conditions: L s0 &amp;lt; L &amp;lt; r0 2 5 and L &amp;lt; 2 . =-=(19)-=- 5σ If these conditions are satisfied, the relative error on the integrand of I2 is maximally roughly 50%. This apparently high error can only occur near the end-points of the edges. On most part of t   </text>
<query_num> 19807 </query_num>
<text>   main factors in the integrand allows to obtain a good approximation: 1 1 � √ = 1 − S2 +t 2 s0 1 � 2t0Δt 2 s2 + 0 � Δt � � � 2 + ... (14) s0 1 R2 1 = +t 2 r2 � � 2t0Δt 1 − 0 r2 + 0 � Δt � � � 2 + =-=... (15) r-=-0 e −σs = e ���σs0 � � 2t0Δt 1 − σs0 s2 + ( 0 Δt ) s0 2� � + ... (16) Here, s0 := s(t0) and r0 := r(t0). The product of the Taylor approximations results in a simple polynomial and its integration is   </text>
<query_num> 19808 </query_num>
<text>   o sources in a dipole configuration. The terms are similar so that we will focus on the integral of the contribution of a single source point d of the dipole (see Table 1): � I = z(1 + σs) A e−σs dx=-=. (8)-=- s3 s denotes the distance between the point x on the polygon A to the considered dipole source d. We first assume that A is a triangle. We shall see that our result straightforwardly extends to the c   </text>
<query_num> 19809 </query_num>
<text>   tens et al. / Interactive Rendering of Translucent Deformable Objects t(θ) = Rtanθ and dθ = Rdt R2 +t 2 (11) � θ2 e I2 = z θ1 −σs Rdt s R2 +t 2 (12) � t2 1 1 = zR √ t1 S2 +t 2 R2 +t 2 e���σ√ S2 +t2 dt=-=. (13)-=- t1 = Rtan(θ1) and t2 = Rtan(θ2) denote the distances from the edge endpoints to d ′′ . We did not find an analytic solution to this integral as such, but Taylor series expansion w.r.t. the midpoint o   </text>
<query_num> 19810 </query_num>
<text>   the material, we get the following subMertens et al. / Interactive Rendering of Translucent Deformable Objects surface scattering reflectance function: S(xi,ωi;xo,ωo) = 1 π Ft(η,ωo)Rd(xi;xo)Ft(η,ωi). =-=(2)-=- Substituting this into Equation 1, we get: L → (xo,ωo) = 1 Ft(η,ωo)B(xo) π � (3) B(xo) = E(xi) = E(xi)Rd(xi,xo)dxi S � Ω (4) + L (xi) ← (xi,ωi)Ft(η,ωi)(Ni · ωi)dωi (5) In order to render translucent   </text>
</top>
<top>
<paper_num> 199 </paper_num>
<paper_title>   Managing Multiple Knowledge Sources in Constraint-Based Parsing of Spoken Language.  </paper_title>
<paper_abstract>   In this paper, we describe a system which is capable of utilizing a variety of knowledge sources to select the most appropriate parse for a spoken sentence. These knowledge sources include syntax, semantics, and contextual information. We discuss one way to utilize contextual information when determining the parse for a sentence. At its simplest level, the system can be thought of as a generalpurpose query answering system for multiple topical databases. The user&amp;apos;s input would be processed by the language processor which interfaces to the databases with the goal of interacting with the correct database in order to provide a reasonable answer to the user&amp;apos;s spoken request. Initially, it analyzes a word graph of sentence hypotheses provided by a speech recognizer using general syntactic and semantic rules. Then, if the utterance is still ambiguous, it utilizes contextspecific constraints to further refine the analysis. This brings us closer to developing a more general purpose interface f...  </paper_abstract>
<query_num> 19901 </query_num>
<text>   . There has been considerable interest in the development of parsers for grammars that are more expressive than the class of context-free grammars, but less expressive than context-sensitive grammars =-=[20, 36, 37]-=-. The running time of the CDG parser compares quite favorably to the running times of parsers for languages which are beyond context-free. For example, the parser for tree adjoining grammars (TAG) has   </text>
<query_num> 19902 </query_num>
<text>   CREW P-RAM model (Concurrent Read, Exclusive Write, Parallel Random Access Machine), but requires O(n 6 ) processors. In contrast, we have devised a parallelization for the single sentence CDG parser =-=[15, 16]-=- which uses O(n 4 ) processors to parse in O(k) time for a CRCW P-RAM model (Concurrent Read, Concurrent Write, Parallel Random Access Machine), where n is the number of words in the sentence, and k,   </text>
<query_num> 19903 </query_num>
<text>   e each application defines its own context. We have introduced a method for incorporating context-specific constraints into the parsing process. However, we have not discussed how discourse structure =-=[8, 9]-=- would impact the system. It would be very helpful for tracking interactions between a user and the general purpose interface to multiple applications and is a topic which we plan to study in the futu   </text>
<query_num> 19904 </query_num>
<text>   e values for the role are selected from the finite set L 1 \Theta fnil,1,2,: : :,ng, CDG parsing can be viewed as a constraint satisfaction problem over a finite domain. Hence, constraint propagation =-=[23, 28, 38]-=- can be used to develop the parse of a sentence. Enumeration of the individual parses for a highly ambiguous sentence is intractable. Therefore, a CDG parser generates all parses for a sentence in a c   </text>
<query_num> 19905 </query_num>
<text>   into our parser. In fact, they were added to the grammar without modifying a single syntactic rule. Preliminary work also suggests that prosodic constraints should be as simple to add to our grammar =-=[12, 42]-=-. Syntactic and semantic constraints are very useful for pruning syntactically or semantically anomalous word nodes from an SLCN. However, they do not, in many cases, sufficiently constrain the SLCN t   </text>
<query_num> 19906 </query_num>
<text>   n language understanding will require effective utilization of this information. To use the variety of knowledge sources needed to disambiguate language, we have constructed a constraint-based system =-=[11, 14, 43]-=- which is an extension to Constraint Dependency Grammar (CDG) parsing as defined by Maruyama [24, 25, 26]. This system is capable of propagating a wide variety of constraints, including syntactic, lex  word graph represents a disjunction of all possible sentence candidates that the speech recognizer provides. Word graphs are typically more compact and more expressive than N-best sentence lists. In =-=[43]-=-, word graphs were constructed from three different lists of sentence hypotheses. The word graphs provided an 83% reduction in storage, and in all cases, they encoded more possible sentence hypotheses   </text>
<query_num> 19907 </query_num>
<text>   nd reduces its understandability. Semantic grammars have been effective for limited domains, but they do not scale up well to larger systems [1]. The second method, the interleaved semantics approach =-=[4, 40, 41]-=-, separates syntactic and semantic processing into two modules, where the semantic module eliminates constituents that are semantically anomalous from further consideration by the syntactic parser. Th   </text>
<query_num> 19908 </query_num>
<text>   ntactic representation of a sentence can constrain the possible antecedents for a referential noun phrase, while the antecedent of a pronoun can also constrain the sentence&amp;apos;s syntactic representation =-=[10]-=-. One way to resolve ambiguity is to utilize a wide variety of knowledge sources. The knowledge sources commonly used in speech understanding are shown in Figure 1. EfAcoustic/ Phonetic Syntactic Sema   </text>
<query_num> 19909 </query_num>
<text>   owledge sources needed to disambiguate language, we have constructed a constraint-based system [11, 14, 43] which is an extension to Constraint Dependency Grammar (CDG) parsing as defined by Maruyama =-=[24, 25, 26]-=-. This system is capable of propagating a wide variety of constraints, including syntactic, lexical, semantic, prosodic, and contextual constraints. The central data structure for this system is a wor lecting the correct computer application with which to interact. 2 The Theoretical Basis of the SLCN Parser Our system uses Constraint Dependency Grammar (CDG) grammar, originally defined by Maruyama =-=[24, 25, 26]-=-, to process sentences. In the following subsections, we will describe the CDG grammar formalism, the CDG parsing algorithm, and the benefits of a constraint-based system. 2.1 Elements of a CDG Gramma   </text>
<query_num> 19910 </query_num>
<text>   processors [30]. To parse a free-order language like Latin, CFGs require that additional rules containing the permutations of the right-hand side of a production be explicitly included in the grammar =-=[29]-=-. Unordered CFGs do not have this combinatorial explosion of rules, but the recognition problem for this class of grammars is NP-complete. A free-order language can easily be handled by a CDG parser b   </text>
<query_num> 19911 </query_num>
<text>   ted in the constraint network [14]. Additional knowledge sources are quite easy to add given the uniform framework provided by constraints, as we demonstrate in this paper. Tight coupling of prosodic =-=[3]-=- and semantic rules with CFG grammar rules typically increases the size and complexity of the grammar and reduces its understandability. Semantic grammars have been effective for limited domains, but   we combine two ordered knowledge sources together in a single module, the resulting system can be difficult to understand and can often become intractable. For example, combining prosodic processing =-=[3]-=- with CFG grammar rules typically increases the size and complexity of the grammar and reduces its understandability. A similar problem holds for semantic grammars [1]. The ordering of knowledge sourc   </text>
<query_num> 19912 </query_num>
<text>   veloped an algorithm to achieve arc consistency in an SLCN by using the properties of the directed acyclic graph representing the word network to filter role values that can never appear in any parse =-=[11, 17]-=-. This algorithm, described in the next section, can process single sentences or word graphs. 3.1 SLCN Arc Consistency When we create a constraint network representing multiple alternative sentence hy N. Consider Figure 21, which shows the roles that are adjacent to role i in an SLCN. Because every sentence in the SLCN which 9 The only condition for a supporting b in the general MUSE CSP algorithm =-=[17, 18]-=- is R2 (i; a; j; b) because not all values assigned to variables in CSP have modifiees. Notation Meaning (i; j) An ordered pair of roles. N fi; j; : : :g is the set of all roles, with jN j = nsp. L fa r of word nodes in network. By comparison, the running time for CN arc consistency is O(n 4 ), assuming that there are n words in a sentence. The proof of correctness of this algorithm is detailed in =-=[17, 18]-=-, but we will summarize it below. A role value is eliminated from a domain by SLCN arc consistency only if its LocalPrev -Support or its Local-Next-Support set becomes empty. Therefore, we must show t   </text>
</top>
<top>
<paper_num> 200 </paper_num>
<paper_title>   Tuning Strassen's Matrix Multiplication for Memory Efficiency.  </paper_title>
<paper_abstract>   Strassen&amp;apos;s algorithm for matrix multiplication gains its lower arithmetic complexity at the expense of reduced locality of reference, which makes it challenging to implement the algorithm efficiently on a modern machine with a hierarchical memory system. We report on an implementation of this algorithm that uses several unconventional techniques to make the algorithm memory-friendly. First, the algorithm internally uses a non-standard array layout known as Morton order that is based on a quad-tree decomposition of the matrix. Second, we dynamically select the recursion truncation point to minimize padding without affecting the performance of the algorithm, which we can do by virtue of the cache behavior of the Morton ordering. Each technique is critical for performance, and their combination as done in our code multiplies their effectiveness. Performance comparisons of our implementation with that of competing implementations show that our implementation often outperforms the alternati...  </paper_abstract>
<query_num> 20001 </query_num>
<text>   avoiding self-interference misses [17]. Given the hierarchical nature of the algorithm (the decomposition is by quadrants within quadrants within : : :), hierarchical layouts such as Morton ordering =-=[8]-=- naturally suggest themselves for storing the matrices. An operational definition of Morton ordering is as follows. Divide the original matrix into four quadrants, and lay out these quadrants in memor epresentations of matrices. Morton ordering has also appeared in the parallel computing literature, where is has been used for load balancing of irregular problems [20]. Most recently, Frens and Wise =-=[8]-=- discuss an implementation of a recursive O(n 3 ) matrix multiplication algorithm using hierarchical matrix layout, in which they sequence the recursive calls in an unusual manner to get better reuse   </text>
<query_num> 20002 </query_num>
<text>   condition. Our top-level conversion between the column-major layout at the interface level and the Morton ordering used internally can be viewed as a logical extension of this proposal. Ghosh et al. =-=[9]-=- present an analytical representation of cache misses for perfect loop nests, which they use to guide selected code optimization problems. Their work, like all of the other work cited above, relies on   </text>
<query_num> 20003 </query_num>
<text>   cursion to the level of single matrix elements as they do, but truncate the recursion when we reach tile sizes that fit in the upper levels of the memory hierarchy. 5.3 Cache behavior Several authors =-=[17, 15, 4, 21]-=- discuss loop transformations such as tiling that attempt to reduce the number of cache misses incurred by a loop nest and thus improve its performance. While these loop transformations are not specif   </text>
<query_num> 20004 </query_num>
<text>   fects Our initial efforts to gain further insight into the performance variability of our implementation begins with analysis of its cache behavior. Here, we present preliminary results. We used ATOM =-=[22]-=- to perform cache simulations of a 16KB direct-mapped cache with 32 byte blocks of both the DGEFMM, and MODGEMM implementations. Figure 9 shows the miss ratios of each implementation for matrix sizes   </text>
<query_num> 20005 </query_num>
<text>   ing tile sizes in the range shown in Figure 2. Second, to achieve performance stability as T varies, it is also important to have the tile contiguous in memory, thus avoiding self-interference misses =-=[17]-=-. Given the hierarchical nature of the algorithm (the decomposition is by quadrants within quadrants within : : :), hierarchical layouts such as Morton ordering [8] naturally suggest themselves for st cursion to the level of single matrix elements as they do, but truncate the recursion when we reach tile sizes that fit in the upper levels of the memory hierarchy. 5.3 Cache behavior Several authors =-=[17, 15, 4, 21]-=- discuss loop transformations such as tiling that attempt to reduce the number of cache misses incurred by a loop nest and thus improve its performance. While these loop transformations are not specif   </text>
<query_num> 20006 </query_num>
<text>   ix multiplication as a building block in numerical codes has generated a significant amount of research into techniques for improving the performance of this basic operation. Several of these efforts =-=[3, 6, 12, 13, 14, 19]-=- focus on algorithms whose arithmetic complexity This work supported in part by DARPA Grant DABT63-98-1-0001, NSF Grants CDA-97-2637 and CDA-9512356, Duke University, and an equipment donation through   </text>
<query_num> 20007 </query_num>
<text>   ix multiplication as a building block in numerical codes has generated a significant amount of research into techniques for improving the performance of this basic operation. Several of these efforts =-=[3, 6, 12, 13, 14, 19]-=- focus on algorithms whose arithmetic complexity This work supported in part by DARPA Grant DABT63-98-1-0001, NSF Grants CDA-97-2637 and CDA-9512356, Duke University, and an equipment donation through educes performance. This overhead is generally limited by stopping the recursion early and performing a conventional matrix multiplication on submatrices that are below the recursion truncation point =-=[13]-=-. Second, the division step must efficiently handle odd-sized matrices. This can be solved by one of several schemes: by embedding the matrix inside a larger one (called static padding), by decomposin at the leaf of the recursion tree and limits the amount of static padding. We measured execution times of our implementation (MODGEMM) and two alternative implementations, DGEFMM uses dynamic peeling =-=[13]-=- and DGEMMW uses dynamic overlap [6], on both a DEC Alpha and a SUN UltraSPARC II. Our results show wide variability in the performance of all three implementations. On the Alpha, our implementation (  products can be computed by recursively invoking Strassen&amp;apos;s algorithm on smaller subproblems, and switching to the conventional algorithm at some matrix size T (called the recursion truncation point =-=[13]-=-) at which Strassen&amp;apos;s construction is no longer advantageous. If one were to estimate running time by counting arithmetic operations, the recursion truncation point would be around 16. However, the em  in Section 3.3. ffl A second solution, dynamic peeling, peels off the extra row or column at each level, and separately adds their contributions to the overall solution in a later fix-up computation =-=[13]-=-. This eliminates the need for extra padding, but reduces the portion of the matrix to which Strassen&amp;apos;s algorithm applies, thus reducing the potential benefits of the recursive strategy. The fix-up co   </text>
<query_num> 20008 </query_num>
<text>   lgorithmic advantages of quad-tree representations of matrices. Morton ordering has also appeared in the parallel computing literature, where is has been used for load balancing of irregular problems =-=[20]-=-. Most recently, Frens and Wise [8] discuss an implementation of a recursive O(n 3 ) matrix multiplication algorithm using hierarchical matrix layout, in which they sequence the recursive calls in an   </text>
<query_num> 20009 </query_num>
<text>   lower than DGEFMM (8%), which matches our expectations. The second observation is the unexpected dramatic drop in MODGEMM&amp;apos;s miss ratio at a matrix size of 513. Preliminary investigations using CProf =-=[18]-=- reveal that this drop is due to a reduction in conflict misses. 500.0 510.0 520.0 530.0 Matrix Size 0.020 0.040 0.060 0.080 0.100 Miss Ratio MODGEMM DGEFMM Figure 9: Cache Miss Ratios for 16KB Direct   </text>
</top>
<top>
<paper_num> 201 </paper_num>
<paper_title>   Swarm: Hyper Awareness, Micro Coordination, and Smart Convergence through Mobile Group Text Messaging.  </paper_title>
<paper_abstract>   Text messaging has become an integral part of mobile communication, with a significant impact on cultural and social norms in many countries. Our goal is to explore how group-based text messaging enables continuous social awareness, group coordination and smart convergence on social events. We implemented a group text messaging system, Swarm, and deployed it to several communication groups for ten months. Through usage logs, questionnaires, text analysis, and direct observation we examined how people integrate group text messaging into their day to day social lives. Swarm was used primarily for lightweight distribution of context information (availability, location, and event status) in order to facilitate social convergence. We discuss the lessons learned from our deployment, and the implications for the design of mobile, group communication systems. 1.  </paper_abstract>
<query_num> 20101 </query_num>
<text>   Service (SMS), has played a vital role in the evolution of interpersonal communication, impacting relationships, politics, health care, and even religious practices in much of Europe, Asia and Africa =-=[12,10,21,23]-=-. Understanding how people use mobile devices to communicate, and learning how to best make interactions through mobile technology as lightweight and unobtrusive as possible, is essential to the ongoi   </text>
<query_num> 20102 </query_num>
<text>   Service (SMS), has played a vital role in the evolution of interpersonal communication, impacting relationships, politics, health care, and even religious practices in much of Europe, Asia and Africa =-=[12,10,21,23]-=-. Understanding how people use mobile devices to communicate, and learning how to best make interactions through mobile technology as lightweight and unobtrusive as possible, is essential to the ongoi s and group identity have become tightly coupled with mobile phones and SMS [3, 24]. Not surprisingly, SMS is heavily used for establishing and maintaining both friendships and romantic relationships =-=[3, 10, 11]-=-. Mobile phones and SMS foster a particularly strong sense of intimacy for teenagers because it allows a continuous connection to friends outside of parental control. Teenagers find it very important   </text>
<query_num> 20103 </query_num>
<text>   and timely contextual information is key to successful micro coordination and smart convergence. One approach is to create systems that infer the user’s context and automatically share it with others =-=[1, 2, 14]-=-. Current location, potential destinations and interrupt ability are just a few examples of context information that can be collected [6]. Unfortunately, many of these contextsensing capabilities rely   </text>
<query_num> 20104 </query_num>
<text>   and timely contextual information is key to successful micro coordination and smart convergence. One approach is to create systems that infer the user’s context and automatically share it with others =-=[1, 2, 14]-=-. Current location, potential destinations and interrupt ability are just a few examples of context information that can be collected [6]. Unfortunately, many of these contextsensing capabilities rely t to the user. The latter approach may provide users with enough flexibility, expressiveness and control over context sharing to efficiently communicate and coordinate plans. Work by Barkhuus and Dey =-=[1]-=- has investigated active versus passive approaches to context distribution. The awareness and coordination activities enabled are very similar to the &amp;quot;passive context-awareness&amp;quot; systems which were pre that the act of explicitly sending location information communicates both where the user is and her desire to socialize, often with just one word. In addition, many of the potential automatic methods =-=[1, 6, 15]-=- raise serious privacy concerns. However, it might be valuable for a user to be able to broadcast location information with a simple touch of one key that expresses to an entire group of people “I am   </text>
<query_num> 20105 </query_num>
<text>   come an integral part of their social lives. Studies of teens, primarily in Japan and the UK, have shown how social practices and group identity have become tightly coupled with mobile phones and SMS =-=[3, 24]-=-. Not surprisingly, SMS is heavily used for establishing and maintaining both friendships and romantic relationships [3, 10, 11]. Mobile phones and SMS foster a particularly strong sense of intimacy f   </text>
<query_num> 20106 </query_num>
<text>   e user’s context and automatically share it with others [1, 2, 14]. Current location, potential destinations and interrupt ability are just a few examples of context information that can be collected =-=[6]-=-. Unfortunately, many of these contextsensing capabilities rely on technology that is not widely available on mobile phones and require users to explicitly configure whom they share information with i that the act of explicitly sending location information communicates both where the user is and her desire to socialize, often with just one word. In addition, many of the potential automatic methods =-=[1, 6, 15]-=- raise serious privacy concerns. However, it might be valuable for a user to be able to broadcast location information with a simple touch of one key that expresses to an entire group of people “I am   </text>
<query_num> 20107 </query_num>
<text>   hibited [13]. Hyper awareness strengthens social ties and creates new opportunities for social exchange [25]. Hyper awareness has been a goal of many ubiquitous computing systems. For example ParcTab =-=[22]-=- provided location information and in/out status for its users, and Portholes [7] proposed the use of a shared image database to support shared awareness and build a sense of community. 3.2 Micro Coor   </text>
<query_num> 20108 </query_num>
<text>   hile in public areas or in transit, such as in Japanese subways where mobile phone calls are prohibited [13]. Hyper awareness strengthens social ties and creates new opportunities for social exchange =-=[25]-=-. Hyper awareness has been a goal of many ubiquitous computing systems. For example ParcTab [22] provided location information and in/out status for its users, and Portholes [7] proposed the use of a   </text>
<query_num> 20109 </query_num>
<text>   people coordinate the time, place and details of an event from moment to moment [17]. This fluid form of negotiation allows groups of friends to coordinate more spontaneously and converge more often =-=[8, 17, 18]-=-. This type of distributed, moment-to-moment coordination is a fundamental application of mobile technologies. As mentioned earlier, voice communication is not necessarily the most efficient or approp   </text>
<query_num> 20110 </query_num>
<text>   s and group identity have become tightly coupled with mobile phones and SMS [3, 24]. Not surprisingly, SMS is heavily used for establishing and maintaining both friendships and romantic relationships =-=[3, 10, 11]-=-. Mobile phones and SMS foster a particularly strong sense of intimacy for teenagers because it allows a continuous connection to friends outside of parental control. Teenagers find it very important   </text>
<query_num> 20111 </query_num>
<text>   s for social exchange [25]. Hyper awareness has been a goal of many ubiquitous computing systems. For example ParcTab [22] provided location information and in/out status for its users, and Portholes =-=[7]-=- proposed the use of a shared image database to support shared awareness and build a sense of community. 3.2 Micro Coordination and Smart Convergence Micro coordination is the process by which people   </text>
</top>
<top>
<paper_num> 202 </paper_num>
<paper_title>   Learning Verb Subcategorization from Corpora: Counting Frame Subsets.  </paper_title>
<paper_abstract>   We present some novel machine learning techniques for the identification of subcategorization information for verbs in Czech. We compare three different statistical techniques applied to this problem. We show how the learning algorithm can be used to discover previously unknown subcategorization frames from the Czech Prague Dependency Treebank. The algorithm can then be used to label dependents of a verb in the Czech treebank as either arguments or adjuncts. Using our techniques, we are able to achieve 88 % accuracy on unseen parsed text.  </paper_abstract>
<query_num> 20201 </query_num>
<text>   998) give several reasons why subcategorization information is important for a natural language parser. Machine-readable dictionaries are not comprehensive enough to provide this lexical information (=-=Manning 1993, Briscoe 1997-=-). Furthermore, such dictionaries are available only for very few languages. We need some general method for the automatic extraction of subcategorization information from text corpora.  =-=bster and Marcus 1989-=-), and (=-=Ushioda et al. 1993-=-). (=-=Brent 1993-=-) uses standard hypotheses testing method for filtering frames observed with a verb. Brent applied his method to very few verbs however. (=-=Manning 1993-=-) applies Brent&amp;apos;s method to parsed data and obtains a subcategorization dictionary for a larger set of verbs. (=-=Briscoe and Carroll 1997-=-) and (=-=Carroll 1998-=-) differ from earlier work in that a substanti   </text>
<query_num> 20202 </query_num>
<text>   automatic extraction of subcategorization information from text corpora. Several techniques and results have been reported on learning subcategorization frames (SFs) from text corpora (=-=Webster 1989,Brent 1991, Brent 1993, Brent 1994, Ushioda 1993, Manning 1993, Ersan 1996, Briscoe 1997, Carroll 1998-=-). All of this work deals with English. In this paper we report on techniques that automatically extract SFs ing that the data is binomially distributed, we can look for frames that co-occur with a verb more often than chance. This is the method used by several earlier papers on SF extraction starting with (=-=Brent 1991, 1993, 1994-=-). Let us consider probability p f which is the probability that a given verb is observed with a frame but this frame is not a valid SF for this verb. p is the error probability on identif  number of correct suggestions divided by the number of known verb complements (the number of &amp;quot;questions&amp;quot;). 5. Comparison with related work Preliminary work on SF extraction from corpora was done by (=-=Brent 1991, 1993, 1994-=-), (=-=Webster and Marcus 1989-=-), and (=-=Ushioda et al. 1993-=-). (=-=Brent 1993-=-) uses standard hypotheses testing method for filtering frames observed with a verb. Brent applied his method to very fe   </text>
<query_num> 20203 </query_num>
<text>   btains a subcategorization dictionary for a larger set of verbs. (=-=Briscoe and Carroll 1997-=-) and (=-=Carroll 1998-=-) differ from earlier work in that a substantially larger set of SF types are considered; (=-=Carroll and Rooth 1998-=-) use an iterative EM algorithm to learn subcategorization as a result of parsing, and, in turn, to improve parsing accuracy by applying the verb SFs obtained. A complete comparison of all the previou   </text>
<query_num> 20204 </query_num>
<text>   es for extracting SF information from data can be used along with other research, which aims to discover relationships between different SFs of a verb (=-=Stevenson and Merlo 1999, Lapata and Brew 1999,Lapata 1999, Stevenson et al. 1999-=-). The statistical models in this paper were based on the assumption that given a verb, different SFs occur independently. This assumption is used to justify the use of the bino   </text>
<query_num> 20205 </query_num>
<text>   ment information to the treebank. Also, techniques for extracting SF information from data can be used along with other research, which aims to discover relationships between different SFs of a verb (=-=Stevenson and Merlo 1999, Lapata and Brew 1999, Lapata 1999, Stevenson et al. 1999-=-). The statistical models in this paper were based on the assumption that given a verb, different SFs occur independently. This assumption is   </text>
<query_num> 20206 </query_num>
<text>   ounts in the training data. Using the values computed above: # # # # # # Q N S Q N S = = # # # # Q Q N N S + + = Taking these probabilities to be binomially distributed, the log likelihood statistic (=-=Dunning 1993-=-) is given by: N4 od do (2) N4 v na (1) N4 v (1+1) N4 od (2) v na (0) N4 na (0) od do (0) N4 do (0) N4 po (1) N4 (2+2+1) od (0) do (0) v (0) na (0) po (0) empty (0) ( ) ( ) ( ( ) ( )) # # # # # # # #   </text>
<query_num> 20207 </query_num>
<text>   s observed with a verb. Brent applied his method to very few verbs however. (=-=Manning 1993-=-) applies Brent&amp;apos;s method to parsed data and obtains a subcategorization dictionary for a larger set of verbs. (=-=Briscoe and Carroll 1997-=-) and (=-=Carroll 1998-=-) differ from earlier work in that a substantially larger set of SF types are considered; (=-=Carroll and Rooth 1998-=-) use an iterative EM algorithm to learn subcategorization as a resu  in common. They all assume that they know the set of possible SF types in advance. Their task can be viewed as assigning one or more of the (known) SF types to a given verb. In addition, except for (=-=Briscoe and Carroll 1997-=-) and (=-=Carroll and Minnen 1998-=-), only a small number of SF types is considered. Using a dependency treebank as input to our learning algorithm has both advantages and drawbacks. There are two main adv   </text>
<query_num> 20208 </query_num>
<text>   xtraction of subcategorization information from text corpora. Several techniques and results have been reported on learning subcategorization frames (SFs) from text corpora (=-=Webster 1989, Brent 1991,Brent 1993, Brent 1994, Ushioda 1993, Manning 1993, Ersan 1996, Briscoe 1997, Carroll 1998-=-). All of this work deals with English. In this paper we report on techniques that automatically extract SFs for Czech,  the number of &amp;quot;questions&amp;quot;). 5. Comparison with related work Preliminary work on SF extraction from corpora was done by (=-=Brent 1991, 1993, 1994-=-), (=-=Webster and Marcus 1989-=-), and (=-=Ushioda et al. 1993-=-). (=-=Brent 1993-=-) uses standard hypotheses testing method for filtering frames observed with a verb. Brent applied his method to very few verbs however. (=-=Manning 1993-=-) applies Brent&amp;apos;s method to parsed data and obtain   </text>
</top>
<top>
<paper_num> 203 </paper_num>
<paper_title>   SMV: Selective Multi-Versioning STM.  </paper_title>
<paper_abstract>   We present Selective Multi-Versioning (SMV), a new STM that reduces the number of aborts, especially those of long read-only transactions. SMV keeps old object versions as long as they might be useful for some transaction to read. It is able to do so while still allowing reading transactions to be invisible by relying on automatic garbage collection to dispose of obsolete versions. SMV is most suitable for read-dominated workloads, for which it achieves much better performance than previous solutions. It has an up to ×11 throughput improvement over a single-version STM and more than a two-fold improvement over an STM keeping a constant number of versions per object. We show that unlike STMs keeping a constant number of versions, SMV operates successfully even in systems with limited memory. 1.  </paper_abstract>
<query_num> 20301 </query_num>
<text>   36: return true Commit (lines 15–28) consists of the following steps: 1. Lock the objects in the write set (line 15). Deadlocks can be detected using standard mechanisms (e.g., timeouts or Dreadlocks =-=[21]-=-), or may be avoided if acquired in the same order by every transaction. 2. Validate the read set (function validateReadSet, lines 34–36). 3. Insert strong references to the over-written versions to T   </text>
<query_num> 20302 </query_num>
<text>   akes them GCable by an automatic garbage collector as soon as they are not. For infrequently updated objects, SMV typically keeps a single version. Other previous suggestions for multi-versioned STMs =-=[2, 5, 20, 22, 23]-=- were based on cycle detection in the conflict graph, a data structure representing all data dependencies among transactions. Cycle detection incurs a high cost (quadratic in the number of transaction   </text>
<query_num> 20303 </query_num>
<text>   d of multi-versioning, STMs can avoid aborts by reading uncommitted values and then having the reader block until the writer commits [24], or by using read-write locks to block in case of concurrency =-=[1, 10]-=-. These approaches differ from SMV, where transactions never block and may always progress independently. Moreover, reads, which are invisible in SMV, must be visible in these “blocking” approaches. I   </text>
<query_num> 20304 </query_num>
<text>   gement of old object versions (see Section 2). Instead of keeping a variable number of versions based on demand, multi-versioned STMs existing today keep a constant number of versions for each object =-=[7, 25, 26]-=-. As we show below, this approach does not provide enough of a performance benefit for read-only transactions, and, even worse, it causes severe memory problems in long executions. The challenge is, t yle algorithm — a single-versioned STM whose basic operation is like that of TL2 [11]; and to a k-versioned variant of the same algorithm, which keeps k previous versions per object, similarly to LSA =-=[25]-=-. We find that SMV is extremely efficient for read-dominated workloads with long-running transactions. For example, in STMBench7 with 64 threads, the throughput of SMV is eleven times higher than that locking of updated objects, and a global version clock for consistency checking. In a way, SMV can be seen as a multi-versioned extension of TL2. Among multi-versioned STMs, the closest to SMV is LSA =-=[25]-=-. LSA, as well as its snapshot-isolation variation [26], implements a simple solution to garbage collection: it keeps a constant number of versions for each object. However, this approach leads to sto vate memory, and do not affect global memory. Invisible reads have been argued to be important for performance, especially in multi-core systems, where updates to global memory cause caches to thrash =-=[12, 25]-=-. Multi-versioning for reads. Read-only transactions always commit. We achieve this by allowing read-only transaction Ti to observe a consistent snapshot corresponding to Ti’s start time — when Ti rea  data is GCable, i.e., once the version of the object cannot be read by any transaction, this version is not strongly referenced by any live memory object. Global version clock. Like TL2 [11] and LSA =-=[25]-=-, SMV uses a global version clock to detect conflicts. Each transaction reads the clock when it begins, and update transactions increment the clock upon commit. Each object is tagged with the version   </text>
<query_num> 20305 </query_num>
<text>   orted by the Israeli Ministry of Science Knowledge Center on CMP, and by the Hasso Plattner Institute. 1 http://www.azulsystems.com/blog/cliff-click/ 2008-05-27-clojure-stms-vs-locks sistent snapshot =-=[3]-=- of the objects it accesses, e.g., values that reflect updates by transactions that committed before it began and no partial updates of concurrent transactions. This way, multiple versions have the po   </text>
<query_num> 20306 </query_num>
<text>   s that can safely read them. In Section 4 we evaluate different aspects of SMV’s performance. We implement SMV in Java 2 and study its behavior using STMBench7 [15] as well as a Java port of Vacation =-=[8]-=-. We compare SMV to a TL2-style algorithm — a single-versioned STM whose basic operation is like that of TL2 [11]; and to a k-versioned variant of the same algorithm, which keeps k previous versions p etup The algorithms are implemented in Java. In order to evaluate their performance, we use two types of benchmarks: 1) the Java version of STMBench7 [15]; and 2) Vacation, which is part of the STAMP =-=[8]-=- benchmark suite. STMBench7. STMBench7 aims to simulate different behaviors of real-world programs by invoking both read-only and update transactions of different lengths over large data structures, t   </text>
<query_num> 20307 </query_num>
<text>   s, and exploits the automatic GC available in languages with managed memory. The idea of keeping information as long is it might be needed by on-going transactions was recently used by Bronson et al. =-=[6]-=- for implementing concurrent collections. However, that paper deals with specific data structures rather than general purpose STMs. Instead of multi-versioning, STMs can avoid aborts by reading uncomm   </text>
<query_num> 20308 </query_num>
<text>   tant number of versions per object. We show that unlike STMs keeping a constant number of versions, SMV operates successfully even in systems with limited memory. 1. Introduction Transactional memory =-=[16, 28]-=- is an increasingly popular paradigm for concurrent computing in multi-core architectures. Most transactional memory implementations today are software toolkits, or STMs for short. STMs speculatively   </text>
<query_num> 20309 </query_num>
<text>   with the version clock of the transaction that wrote it. Though the global version clock is a contention-point, many practical optimizations were introduced to reduce the overhead associated with it =-=[11, 27]-=-. Such optimizations are orthogonal to our work, and are therefore beyond the scope of this paper. Perelman et al. [23] show that such global contention is essential for any multi-versioned STM that a   </text>
</top>
<top>
<paper_num> 204 </paper_num>
<paper_title>   Supporting Framework Use via Automatically Extracted Concept-Implementation Templates.  </paper_title>
<paper_abstract>   Abstract. Application frameworks provide reusable concepts that are instantiated in application code through potentially complex implementation steps such as subclassing, implementing callbacks, and making calls. Existing applications contain valuable examples of such steps, except that locating them in the application code is often challenging. We propose the notion of concept implementation templates, which summarize the necessary implementation steps, and an approach to automatic extraction of such templates from traces of sample applications. We demonstrate the feasibility of the template extraction with high precision and recall through an empirical study with twelve realistic concepts from four widely-used frameworks. Finally, we report on a user experiment with twelve subjects in which the choice of templates vs. documentation had much less impact on development time than the concept complexity. 1  </paper_abstract>
<query_num> 20401 </query_num>
<text>   Specification mining is concerned with automatically discovering the protocols or rules that a program must follow when interacting with an API. Existing techniques can be classified into static [17]=-=[18]-=- and dynamic [19][20][21] ones. Examples of static approaches include inferring ordering patterns among method calls [17] or detecting function precedence protocols [18]. Examples of dynamic approache   </text>
<query_num> 20402 </query_num>
<text>   ace. Specification mining is concerned with automatically discovering the protocols or rules that a program must follow when interacting with an API. Existing techniques can be classified into static =-=[17]-=-[18] and dynamic [19][20][21] ones. Examples of static approaches include inferring ordering patterns among method calls [17] or detecting function precedence protocols [18]. Examples of dynamic appro   </text>
<query_num> 20403 </query_num>
<text>   ag Berlin Heidelberg 2009FUDA: Framework Understanding through Dynamic Analysis 345 Several tools have been proposed to address this challenge. Framework usage comprehension tools such as Strathcona =-=[2]-=- and FrUiT [3] apply static analysis to the source code of sample applications and allow retrieving code snippets or usage rules for a particular API element. These tools can be very helpful to unders  by allowing users to generate implementation templates when the framework documentation is missing. Framework usage comprehension is supported by several approaches such as XSnippet [14], Strathcona =-=[2]-=-, Prospector [15], PARSEWeb [16], and FrUiT [3]. Both XSnippet and Strathcona are context-sensitive code assistant tools that allow developers to query a repository of code snippets that are relevant   </text>
<query_num> 20404 </query_num>
<text>   document framework-provided concepts as hierarchies of mandatory and optional features and actively support users in instantiating the concepts through roundtrip engineering. Further, reuse contracts =-=[12]-=- and collaboration contracts [13] help ensure that frameworks are used correctly. Nonetheless, the main difficulty of these approaches is that framework documentation requires manual effort and, conse   </text>
<query_num> 20405 </query_num>
<text>   e affected common nesting and dependency facts are updated accordingly. Template Generation. This section sketches the main steps of the template generation algorithm. Interested readers can refer to =-=[6]-=- for further details. The inputs to theFUDA: Framework Understanding through Dynamic Analysis 355 a1 a2 a3 a4 a5 a6 a7 a8 a9 a10 a11 a12 ↑jface.action.Action.&amp;lt;init&amp;gt; ↑jface.action.IAction.setText ↑jfa divided by the pooled standard deviation. Experimental Results. Due to space constraints, this section summarizes the most important results. A complete description of the data can be found elsewhere =-=[6]-=-[10]. Discussion of Quantitative Results. Figure 7 shows the time measured for each implementation as a function of the documentation aid and concept complexity. Bold labels identify experienced subje   </text>
<query_num> 20406 </query_num>
<text>   e application code. 7 Related Work Framework documentation and completion approaches support framework users passively or actively or both. For instance, framework-specific modeling languages (FSMLs) =-=[11]-=- document framework-provided concepts as hierarchies of mandatory and optional features and actively support users in instantiating the concepts through roundtrip engineering. Further, reuse contracts   </text>
<query_num> 20407 </query_num>
<text>   elberg 2009FUDA: Framework Understanding through Dynamic Analysis 345 Several tools have been proposed to address this challenge. Framework usage comprehension tools such as Strathcona [2] and FrUiT =-=[3]-=- apply static analysis to the source code of sample applications and allow retrieving code snippets or usage rules for a particular API element. These tools can be very helpful to understand concept i emplates when the framework documentation is missing. Framework usage comprehension is supported by several approaches such as XSnippet [14], Strathcona [2], Prospector [15], PARSEWeb [16], and FrUiT =-=[3]-=-. Both XSnippet and Strathcona are context-sensitive code assistant tools that allow developers to query a repository of code snippets that are relevant to the programming task at hand. Given two API   </text>
<query_num> 20408 </query_num>
<text>   implementation templates when the framework documentation is missing. Framework usage comprehension is supported by several approaches such as XSnippet [14], Strathcona [2], Prospector [15], PARSEWeb =-=[16]-=-, and FrUiT [3]. Both XSnippet and Strathcona are context-sensitive code assistant tools that allow developers to query a repository of code snippets that are relevant to the programming task at hand.   </text>
<query_num> 20409 </query_num>
<text>   in concept or functionality is implemented in the source code of an application. Existing approaches can be mainly categorized into static (e.g., [4]), dynamic (e.g., [22]), and hybrid ones (e.g., [5]=-=[23]-=-). One can refer to [5] for a good literature overview. We focus only on the most related dynamic and hybrid techniques. Most of these techniques use two or more traces to filter out irrelevant events   </text>
<query_num> 20410 </query_num>
<text>   is concerned with automatically discovering the protocols or rules that a program must follow when interacting with an API. Existing techniques can be classified into static [17][18] and dynamic [19]=-=[20]-=-[21] ones. Examples of static approaches include inferring ordering patterns among method calls [17] or detecting function precedence protocols [18]. Examples of dynamic approaches contain mining temp   </text>
<query_num> 20411 </query_num>
<text>   ning is concerned with automatically discovering the protocols or rules that a program must follow when interacting with an API. Existing techniques can be classified into static [17][18] and dynamic =-=[19]-=-[20][21] ones. Examples of static approaches include inferring ordering patterns among method calls [17] or detecting function precedence protocols [18]. Examples of dynamic approaches contain mining   </text>
<query_num> 20412 </query_num>
<text>   rs to generate implementation templates when the framework documentation is missing. Framework usage comprehension is supported by several approaches such as XSnippet [14], Strathcona [2], Prospector =-=[15]-=-, PARSEWeb [16], and FrUiT [3]. Both XSnippet and Strathcona are context-sensitive code assistant tools that allow developers to query a repository of code snippets that are relevant to the programmin   </text>
<query_num> 20413 </query_num>
<text>   ss helpful if the developer has only a high-level idea of the concept that needs to be implemented or if the concept spans multiple classes or both. Concept location tools such as SNIAFL [4] or SITIR =-=[5]-=- can be used to locate the code implementing a concept of interest identified by higher-level characteristics such as usage scenarios or domain terms. These tools do not focus on framework API usage,  rtain concept or functionality is implemented in the source code of an application. Existing approaches can be mainly categorized into static (e.g., [4]), dynamic (e.g., [22]), and hybrid ones (e.g., =-=[5]-=-[23]). One can refer to [5] for a good literature overview. We focus only on the most related dynamic and hybrid techniques. Most of these techniques use two or more traces to filter out irrelevant ev fic content from those traces through the event generalization. Furthermore, we are unaware of other techniques using the combination of API trace marking with API trace slicing. In particular, SITIR =-=[5]-=- uses the runtime trace marking to reduce the size of the traces, but it misses the relevant events to the implementation of the desired concept that are not marked at runtime. FUDA is able to identif   </text>
<query_num> 20414 </query_num>
<text>   to fill this gap by allowing users to generate implementation templates when the framework documentation is missing. Framework usage comprehension is supported by several approaches such as XSnippet =-=[14]-=-, Strathcona [2], Prospector [15], PARSEWeb [16], and FrUiT [3]. Both XSnippet and Strathcona are context-sensitive code assistant tools that allow developers to query a repository of code snippets th   </text>
</top>
<top>
<paper_num> 205 </paper_num>
<paper_title>   Location of Mobile Devices Using Networked Surfaces.  </paper_title>
<paper_abstract>   Networked Surfaces are a novel technology, using contact with physical surfaces such as desks to provide network connectivity for mobile devices.  </paper_abstract>
<query_num> 20501 </query_num>
<text>   allow users to work while moving around, an example being &amp;quot;desktop teleporting&amp;quot; [19]. Thirdly, applications can be made location-dependent, e.g. the Stick-E note system [5] and the Cyberguide project =-=[1]-=-. Finally, &amp;quot;intelligent environments&amp;quot; can be constructed, including interactive desks [16] and walls [18], offices [2], and home environments [6]. In addition to the existing applications above, Netwo   </text>
<query_num> 20502 </query_num>
<text>   dings of WaveLAN cards received at multiple well known locations to gain an estimate for the position of the card. The location of devices can be estimated to within three metres. The &amp;quot;Nibble&amp;quot; system =-=[7]-=- also uses WaveLAN signal strength, but with the prototype system having far more wireless access points (10-14 in one experiment). It is therefore able to achieve high accuracies for room-grain locat   </text>
<query_num> 20503 </query_num>
<text>   k is implemented for peripherals or sensors placed in the user&amp;apos;s environment. Also, a number of high bit-rate networks are provided for objects like laptops. Preliminary measurements are described in =-=[10]-=-. Improvements made since then show end-to-end data rates of up to 5Mbit/s across the high speed data networks. 2.2 Topology The layout of surface strips and object pads is called the topology. The to aking protocol, during which the object instructs the tiles to connect it to the desired functions. The protocol was designed to connect objects as quickly as possible; early experiments described in =-=[10]-=- show that in 98% of the cases connection is achieved in under one second. Recent work has improved this time to half a second. For the purpose of locating devices it is important to note that, after   </text>
<query_num> 20504 </query_num>
<text>   lse to estimate the distance to the beacon. The system can estimate postitions to within two metres, and by using multiple receivers can determine orientation to an accuracy of 3s. The Dolphin system =-=[9]-=- is similar to the Active Bat system, but uses broadband ultrasonic transmissions. This allows it to achieve robustness in the presence of ultrasonic noise, and to allow multiple transmitters to send   </text>
<query_num> 20505 </query_num>
<text>   novel context-aware applications specific to Networked Surfaces. Firs fly, visualisation applications have been developed to display location information to interested users in text [8] or graphical =-=[2]-=- form. Next, &amp;quot;follow-me&amp;quot; applications allow users to work while moving around, an example being &amp;quot;desktop teleporting&amp;quot; [19]. Thirdly, applications can be made location-dependent, e.g. the Stick-E note   </text>
<query_num> 20506 </query_num>
<text>   of technology they use. Systems from each of the prominent groups are reviewed below, and later compared to location using Networked Surfaces. Infrared-Based Location Systems. The Active Badge system =-=[23, 8]-=- uses small badges that periodically send infrared pulses containing a unique identification number. The badges can be worn by users or they can be attached to equipment like workstations. Receivers i   </text>
<query_num> 20507 </query_num>
<text>   om record the sightings of the infrared pulses and pass them on to a central server which converts the readings into a location. The system&amp;apos;s resolution is of room-scale granularity. The Locust Swarm =-=[22]-=- system is another infrared-based system. In this system the beaconing transmitters are located in the environment, and the receivers are located on the tagged objects. This configuration allows for p   </text>
<query_num> 20508 </query_num>
<text>   rdly, applications can be made location-dependent, e.g. the Stick-E note system [5] and the Cyberguide project [1]. Finally, &amp;quot;intelligent environments&amp;quot; can be constructed, including interactive desks =-=[16]-=- and walls [18], offices [2], and home environments [6]. In addition to the existing applications above, Networked Surfaces can lend themselves to new types of application, in which both the networkin   </text>
<query_num> 20509 </query_num>
<text>   rs are located on the tagged objects. This configuration allows for privacy guarantees to be made, since the objects infer their own location without the need for a central server. The ParcTab system =-=[20]-=- explores how devices like Personal Digital Assistants can be enhanced to include location-aware applications. The infrared based location technology gives room-scale accuracy. The system also provide   </text>
<query_num> 20510 </query_num>
<text>   the Stick-E note system [5] and the Cyberguide project [1]. Finally, &amp;quot;intelligent environments&amp;quot; can be constructed, including interactive desks [16] and walls [18], offices [2], and home environments =-=[6]-=-. In addition to the existing applications above, Networked Surfaces can lend themselves to new types of application, in which both the networking and location-aware aspects of the technology are used   </text>
<query_num> 20511 </query_num>
<text>   und a room. Colour histograms are used to keep track of each user separately. The system locates users with 10cm accuracy along the 2D plane of the floor. RF-Based Location Systems. The RADAR project =-=[4]-=- uses signal strength readings of WaveLAN cards received at multiple well known locations to gain an estimate for the position of the card. The location of devices can be estimated to within three met   </text>
<query_num> 20512 </query_num>
<text>   worked Surfaces Although a detailed discussion of the Networked Surface implementation is not the subject of this paper, an overview of the system will be presented here. More details can be found in =-=[21]-=- 3 . 2.1 System Architecture Figure 1 shows a diagram of the system architecture. The data networks and power cabling are distributed within the Surface in the form of function buses. The function bus faces can lend themselves to new types of application, in which both the networking and location-aware aspects of the technology are used. While such applications are not the topic of this paper (see =-=[21]-=-), two applications are outlined below. Firs fly, Networked Surfaces may be used to automatically configure connections between devices, based on their physical placement. For example, a peripheral (e   </text>
</top>
<top>
<paper_num> 206 </paper_num>
<paper_title>   Surface and Volume Discretization of Functionally Based Heterogeneous Objects.  </paper_title>
<paper_abstract>   The presented approach to discretization of functionally defined heterogeneous objects is oriented towards applications associated with numerical simulation procedures, for example, finite element analysis (FEA). Such applications impose specific constraints upon the resulting surface and volume meshes in terms of their topology and metric characteristics, exactness of the geometry approximation, and conformity with initial attributes. The function representation of the initial object is converted into the resulting cellular representation described by a simplicial complex. We consider in detail all phases of the discretization algorithm from initial surface polygonization to final tetrahedral mesh generation and its adaptation to special FEA needs. The initial object attributes are used at all steps both for controlling geometry and topology of the resulting object and for calculating new attributes for the resulting cellular representation. Categories and Subject Descriptors  </paper_abstract>
<query_num> 20601 </query_num>
<text>   cussed in [11]. Voxel arrays in volume modeling and graphics can be considered as discrete attribute models with the default geometry represented by a bounding box. Constructive Volume Geometry (CVG) =-=[12]-=- utilizes voxel arrays and continuous scalar fields for representing both geometry and photometric attributes (opacity, color, etc.). Issues of functionally based modeling of volumetric distribution o   </text>
<query_num> 20602 </query_num>
<text>   e node positions in respect to the underlying surface for remeshing. Issues of finite element mesh generation are discussed in detail in [5]. Unstructured mesh generation methods are also surveyed in =-=[31, 32]-=-. Tetrahedrization is one of the widely used methods of 3D discretization. The main approaches to automatic tetrahedral (triangular) mesh generation include spatial decomposition based methods, Delaun   </text>
<query_num> 20603 </query_num>
<text>   imation. The main disadvantages of the mentioned approaches are smoothing or cutting sharp features of the surfaces. Algorithms of sharp features extraction are presented in [24-26]. The optimization =-=[26]-=- is based on the special vertex relocation strategy and triangles subdivision and allows for extraction of sharp edges and peaks taking into account the surface curvature. However, in the process of o  Accordingly, the requirements 1 ~ and 2 from 2.4 are satisfied for the cellular model Bc1 . To extract sharp features from an implicit surface coarse triangulation, we use the algorithm described in =-=[26]-=-. This algorithm is based on combining the application of the following mesh optimization procedures: �� curvature-weighted vertices resampling; �� dual/primal mesh optimization that involves projecti   </text>
<query_num> 20604 </query_num>
<text>   ined using real-valued functions. This procedure is an implementation of the functional to cellular models conversion operation in the cellular-functional modeling framework for heterogeneous objects =-=[2,6]-=-. In contrast to previous works on implicit surface polygonization and volume mesh generation, the main motivation of this work is generation of meshes suitable for finite element analysis with constr ious types. A practical example of FEM generation for a mixing tank impeller was given. All examples have been prepared using software tools developed by the authors. In the cellular-functional model =-=[2,6]-=-, heterogeneous objects are represented as hypervolumes or multidimensional point set with multiple attributes. A multidimensional point sets can include elements of different dimensions, which can be   </text>
<query_num> 20605 </query_num>
<text>   is true, then vertex P lies on a sharp edge, but the case Sh(P)&amp;gt;2 corresponds to a corner point. To single out those spikes that do not lie on sharp edges, we use the heuristic estimation proposed in =-=[24]-=-. T mi n( m , n( ti)) i � � T and mi n( M , N ( ti)) � � . i Here n(ti) are unit normals of triangles ti adjacent to point P, N (ti) are the implicit surface normals at the central points of ti, m � n   </text>
<query_num> 20606 </query_num>
<text>   ng objects along with a set of special requirements. We consider the problem of discretization of functionally based heterogeneous objects within a hybrid cellular-functional representation framework =-=[2,6]-=- in which objects are treated as hypervolumes (multidimensional point sets with multiple attributes) [1]. 2.1 Initial heterogeneous object Let D be an initial heterogeneous object – a hypervolume expr x 2, x 3) � Ω �E 3 , F(X) = 0 }. i Each attribute Ai is defined by its set of values N along � � with a map function S i(X): Ω-&amp;gt;N i and can be represented by any of the attribute models introduced in =-=[2]-=- that differ from each other in the way of defining S i(X). For instance, the function representation (FRep) can be used for defining attributes representing electric or thermal field distribution as  ng space Ω. Such an attribute can be given by the user or can be calculated on the basis of other given attributes A i. There is a promising way of defining A r through FRep on the basis of “sources” =-=[8,2]-=-. As to defining A r through CRep, it is appropriate when the size distribution function is defined in a discrete manner and its values are known at the vertices of a background geometric complex. Suc eometry G, and attributes ( A A G � ~ ~ ~ A1,..., Am ) i m ~ describe the object’s properties. Formally, G can be expressed as a particular case of a model based on the cellular representation (CRep) =-=[2,6]-=-: G c ~ ={ X | X� Ω � E 3 , X� |K 3 |}, 3 r where K � { Ci ; r � 0, 1, 2, 3; i � 1,..., Ir} is a three-dimensional r polyhedral complex consisting of cells Ci . In conventional terms, such a discrete  alued functions. Distance fields are used to model varying material properties satisfying different types of constraints predefined on the initial object geometry. The approach of [1] was extended in =-=[2]-=- to dimensionally heterogeneous objects with multiple attributes by combining the functionally based and cellular representations into a single hybrid model. In this paper, we describe one of the impo   </text>
<query_num> 20607 </query_num>
<text>   to be also applicable to these models. Numerical FEA methods use discrete models (surface and volume meshes) of geometric objects, although meshfree analysis and simulation methods are also emerging =-=[4]-=-. Algorithms for finite element mesh generation are well developed for boundary and spatial enumeration representations [5]. Meshes are also actively used now in visualization, animation, computationa   </text>
</top>
<top>
<paper_num> 207 </paper_num>
<paper_title>   The Efficacy of Software Prefetching and Locality Optimizations on Future Memory Systems.  </paper_title>
<paper_abstract>   Software prefetching and locality optimizations are techniques for overcoming the speed gap between processor and memory. In this paper, we provide a comprehensive summary of current software prefetching and locality optimization techniques, and evaluate the impact of memory trends on the effectiveness of these techniques for three types of applications: regular scientific codes, irregular scientific codes, and pointer-chasing codes. We find that for many applications, software prefetching outperforms locality optimizations when there is sufficient memory bandwidth, but locality optimizations outperform software prefetching under bandwidth-limited conditions. The break-even point (for 1 GHz processors) occurs at roughly 2.26 GBytes/sec on today’s memory systems, and will increase on future memory systems. We also study the interactions between software prefetching and locality optimizations when applied in concert. Naively combining the techniques provides robustness to changes in memory bandwidth and latency, but does not yield additional performance gains. We propose and evaluate several algorithms to better integrate software prefetching and locality optimizations, including a modified tiling algorithm, padding for prefetching, and index prefetching. Finally, we investigate the interactions of stride-based hardware prefetching with our software techniques. We find that combining hardware and software prefetching yields similar performance to software prefetching alone, and that locality optimizations enable stride-based hardware prefetching for benchmarks that do not normally exhibit striding. 1.  </paper_abstract>
<query_num> 20701 </query_num>
<text>   Optimizations studied in [40, 45, 41, 46, 47] are similarly limited to affine array accesses, but use hardware to identify the access pattern automatically. Prefetch engines for affine array accesses =-=[48, 49, 50, 51]-=- provide hardware support for prefetching, but rely on the programmer or compiler to identify the access pattern. Prefetching for pointer-chasing traversals uses one of several possible approaches. Th   </text>
<query_num> 20702 </query_num>
<text>   [73] and devised techniques to tile 3D scientific computations. Researchers have examined irregular computations mostly in the context of parallel computing, using run-time [26, 27, 28] and compiler =-=[74]-=- support to support accesses on message-passing multiprocessors. A few have also looked at techniques for improving locality [22, 23, 24, 25]. Metrics such as reference affinity have been used to guid   </text>
<query_num> 20703 </query_num>
<text>   ads in front of the main computation. The helper threads trigger cache misses on behalf of the main thread, prefetching them into cache. Another recent technique is Content-Directed Prefetching (CDP) =-=[62]-=-. CDP observes data as it is moved from memory into the L2 cache, and issues prefetches for any values that resemble memory addresses in anticipation of their access by the processor. A less hardware-   </text>
<query_num> 20704 </query_num>
<text>   cache miss [66] (the MIT Alewife [67] is a machine that makes use of such an idea), or they can be made on every cycle [68]. More recently, researchers have investigated Simultaneous Multi-Threading =-=[69]-=- in which instructions from multiple threads can be chosen for execution within a single cycle on a multi-issue processor. One drawback of multi-threading techniques, however, is they cannot improve s   </text>
<query_num> 20705 </query_num>
<text>   chasing traversals uses one of several possible approaches. The first approach inserts additional pointers, called jump pointers, into dynamic data structures to connect non-consecutive link elements =-=[52, 7, 8, 11]-=-, as described in Section 3.3. Another approach uses only natural pointers for prefetching [53, 8, 9, 10]. These techniques prefetch pointer chains sequentially, but schedule each prefetch as early as   </text>
<query_num> 20706 </query_num>
<text>   cycle on a multi-issue processor. One drawback of multi-threading techniques, however, is they cannot improve single-thread performance; they only increase overall processor throughput. Burger et al =-=[70]-=- forecasted the increasing importance of memory latency tolerance techniques, such as prefetching and multi-threading, given the widening gap between processor and memory performance. In [70], they st   </text>
<query_num> 20707 </query_num>
<text>   d issues prefetches for any values that resemble memory addresses in anticipation of their access by the processor. A less hardware-centric approach compared to CDP is Guided Region Prefetching (GRP) =-=[63]-=-. GRP relies on the compiler to convey hints to the memory system that identify pointer accesses, indexed array accesses, and accesses that exhibit spatial reuse. Special prefetch hardware triggers pr   </text>
<query_num> 20708 </query_num>
<text>   data accesses are more likely to be to the same cache line. These compiler and run-time transformations can be automated using an inspector-executor approached developed for message-passing machines =-=[26]-=-, where the compiler identifies index accesses and inserts calls to run-time libraries to analyze and reorder data and loop iterations. Figure 6 Part(B) illustrate how a code with indexed accesses may existing tiling techniques [73] and devised techniques to tile 3D scientific computations. Researchers have examined irregular computations mostly in the context of parallel computing, using run-time =-=[26, 27, 28]-=- and compiler [74] support to support accesses on message-passing multiprocessors. A few have also looked at techniques for improving locality [22, 23, 24, 25]. Metrics such as reference affinity have   </text>
<query_num> 20709 </query_num>
<text>   data and iteration reordering heuristics will result in the best locality for a given dataset [76, 77]. Few researchers have investigated data layout transformations for pointer-based data structures =-=[78, 79]-=-. Chilimbi et al. investigated allocation-time and run-time techniques to improve locality for linked lists and trees [30]. In this paper, we propose further extensions. Calder et al. use profiling to   </text>
<query_num> 20710 </query_num>
<text>   e spatial locality [75]. Strout et al. have examined models for determining which combination of run-time data and iteration reordering heuristics will result in the best locality for a given dataset =-=[76, 77]-=-. Few researchers have investigated data layout transformations for pointer-based data structures [78, 79]. Chilimbi et al. investigated allocation-time and run-time techniques to improve locality for   </text>
<query_num> 20711 </query_num>
<text>   ely to be obtained if the data is larger than the cache. Fortunately, recent research has demonstrated run-time data and computation transformations can improve the locality of irregular computations =-=[22, 23, 24, 25]-=-. Because many irregular computations typically perform reduction (commutative and associative) operations such as sum and max, loop iterations can be safely reordered to bring accesses to the same da es (data elements) so that nearby circles access identical or nearby squares. Several data and computation locality transformations exist to solve the problem of improving the locality of such graphs =-=[22, 23, 25, 24]-=-. Our evaluation uses a technique called gpart that relies on hierarchical clustering to improve locality [27, 28]. gpart works in three steps. First, the graph formed by the index array computation h ontext of parallel computing, using run-time [26, 27, 28] and compiler [74] support to support accesses on message-passing multiprocessors. A few have also looked at techniques for improving locality =-=[22, 23, 24, 25]-=-. Metrics such as reference affinity have been used to guide algorithms for splitting data structures and regrouping arrays to improve spatial locality [75]. Strout et al. have examined models for det   </text>
<query_num> 20712 </query_num>
<text>   ful for linear algebra codes [14, 16, 13] and multiple loop nests across time-step loops [20]. In comparison we apply tiling to 3D stencil codes which cannot be tiled with existing methods. Rivera in =-=[18, 19]-=- studied existing tiling techniques [73] and devised techniques to tile 3D scientific computations. Researchers have examined irregular computations mostly in the context of parallel computing, using   </text>
<query_num> 20713 </query_num>
<text>   g. In addition to these pointer prefetching techniques, there have been other techniques more recently that target pointer-chasing and other irregular applications. One such technique is preexecution =-=[56, 57, 58, 59, 60, 61]-=-. Pre-execution uses spare hardware contexts in a multithreaded processor to run one or more helper threads in front of the main computation. The helper threads trigger cache misses on behalf of the m   </text>
<query_num> 20714 </query_num>
<text>   have been shown to be useful in eliminating conflict misses and improving spatial locality [21]. Several cache miss estimation techniques have been proposed to help guide data locality optimizations =-=[72, 13]-=-. Tiling has been proven useful for linear algebra codes [14, 16, 13] and multiple loop nests across time-step loops [20]. In comparison we apply tiling to 3D stencil codes which cannot be tiled with   </text>
<query_num> 20715 </query_num>
<text>   ial reuse. Special prefetch hardware triggers prefetches on L2 misses based on the compiler hints. Lastly, another recent technique is to migrate a prefetching mechanism into the memory system itself =-=[64, 65]-=-. Such memory-side prefetchers reduce the round-trip latency to main memory, increasing the throughput of serialized pointer chasing references [65], and can implement large correlation tables for poi   </text>
<query_num> 20716 </query_num>
<text>   it is intuitive that a compiler can insert prefetches for list nodes further down the list using the size of a node and the location of the first node. This approach, which we call index prefetching =-=[1, 37]-=-, was originally proposed in [8]. With index prefetching, the jump pointers can be removed, thus eliminating all the overhead associated with jump pointer prefetching. To quantify this benefit, we cre   </text>
<query_num> 20717 </query_num>
<text>   lel applications, particularly for scientific programs making regular memory accesses [4, 5, 6, 3]. Recently, techniques have also been developed to apply prefetching to pointer-based data structures =-=[7, 8, 9, 10, 11]-=-. In this section of the paper, we present three software prefetching algorithms proposed previously in the literature for the three types of memory access patterns discussed in the previous section.  hain, thus limiting their effectiveness. ����������������� ����������������� ��������������� ������������������������������������������ �������������������������������������� Jump pointer prefetching =-=[8, 11]-=- is a promising approach for addressing the pointer-chasing problem. In jump pointer prefetching, additional pointers are inserted into a dynamic data structure to connect non-consecutive link element ough a prefetch array. Before prefetching can commence, the prefetch pointers must be set. Figure 5 Part(B) shows an example of prefetch pointer initialization code which uses a history pointer array =-=[8]-=- to set the prefetch pointers. The history pointer array, called “history” in Figure 5, is a circular queue that records the last PD link nodes traversed by the initialization code. Whenever a new lin  insert prefetches for list nodes further down the list using the size of a node and the location of the first node. This approach, which we call index prefetching [1, 37], was originally proposed in =-=[8]-=-. With index prefetching, the jump pointers can be removed, thus eliminating all the overhead associated with jump pointer prefetching. To quantify this benefit, we created index prefetching versions  chasing traversals uses one of several possible approaches. The first approach inserts additional pointers, called jump pointers, into dynamic data structures to connect non-consecutive link elements =-=[52, 7, 8, 11]-=-, as described in Section 3.3. Another approach uses only natural pointers for prefetching [53, 8, 9, 10]. These techniques prefetch pointer chains sequentially, but schedule each prefetch as early as   </text>
<query_num> 20718 </query_num>
<text>   lel applications, particularly for scientific programs making regular memory accesses [4, 5, 6, 3]. Recently, techniques have also been developed to apply prefetching to pointer-based data structures =-=[7, 8, 9, 10, 11]-=-. In this section of the paper, we present three software prefetching algorithms proposed previously in the literature for the three types of memory access patterns discussed in the previous section.  s, called jump pointers, into dynamic data structures to connect non-consecutive link elements [52, 7, 8, 11], as described in Section 3.3. Another approach uses only natural pointers for prefetching =-=[53, 8, 9, 10]-=-. These techniques prefetch pointer chains sequentially, but schedule each prefetch as early as possible to maximize memory latency overlap. A third approach uses a correlation-based predictor in hard   </text>
<query_num> 20719 </query_num>
<text>   lel applications, particularly for scientific programs making regular memory accesses [4, 5, 6, 3]. Recently, techniques have also been developed to apply prefetching to pointer-based data structures =-=[7, 8, 9, 10, 11]-=-. In this section of the paper, we present three software prefetching algorithms proposed previously in the literature for the three types of memory access patterns discussed in the previous section.  �������� ������������������������� ���������������������������������� ��������������������������������������������������� Figure 5: Example pointer prefetching using jump pointers and prefetch arrays =-=[7]-=-. ����� performed for pointer traversal must dereference a series of pointers sequentially. The memory serialization in pointer chasing prevents conventional prefetching techniques from overlapping ca  PD link nodes in a linked list because there are no jump pointers that point to these early nodes. To enable prefetching of early nodes, jump pointer prefetching can be extended with prefetch arrays =-=[7]-=-. In this technique, an array of prefetch pointers is added to every linked list to point to the first PD link nodes. Hence, prefetches can be issued through the memory addresses in the prefetch array chasing traversals uses one of several possible approaches. The first approach inserts additional pointers, called jump pointers, into dynamic data structures to connect non-consecutive link elements =-=[52, 7, 8, 11]-=-, as described in Section 3.3. Another approach uses only natural pointers for prefetching [53, 8, 9, 10]. These techniques prefetch pointer chains sequentially, but schedule each prefetch as early as   </text>
<query_num> 20720 </query_num>
<text>   multiple times before it is flushed. Tiling is very effective with linear algebra codes [14, 15, 16, 17, 18], and has been been extended to handle stencil codes used in iterative PDE solvers as well =-=[19, 20, 13]-=-. A major problem with tiling is that limited cache associativity may cause data in a tile to be mapped onto the same cache lines, even though there is sufficient space in the cache. Conflict misses w on techniques have been proposed to help guide data locality optimizations [72, 13]. Tiling has been proven useful for linear algebra codes [14, 16, 13] and multiple loop nests across time-step loops =-=[20]-=-. In comparison we apply tiling to 3D stencil codes which cannot be tiled with existing methods. Rivera in [18, 19] studied existing tiling techniques [73] and devised techniques to tile 3D scientific   </text>
<query_num> 20721 </query_num>
<text>   n order and data layout of the program at compile and run time using locality optimizations. These optimizations try to improve data locality, the ability of an application to reuse data in the cache =-=[13]-=-. Reuse may be in the form of temporal locality, where the same cache line is accessed multiple times, or spatial locality, where nearby data is accessed together on the same cache line. Previous rese . One useful program transformation is tiling (blocking), which combines strip-mining with loop permutation to form small tiles of loop iterations which are executed together to exploit data locality =-=[13]-=-. Figure 6 Part(A) demonstrates how the 3D Jacobi code can be tiled. By rearranging the loop structure so that the innermost loops can fit in cache (due to fewer iterations), tiling allows reuse to be  multiple times before it is flushed. Tiling is very effective with linear algebra codes [14, 15, 16, 17, 18], and has been been extended to handle stencil codes used in iterative PDE solvers as well =-=[19, 20, 13]-=-. A major problem with tiling is that limited cache associativity may cause data in a tile to be mapped onto the same cache lines, even though there is sufficient space in the cache. Conflict misses w y bottleneck problem by improving data locality. Computation-reordering transformations such as 29sBadawy, Aggarwal, Yeung, &amp; Tseng loop permutation and tiling are the primary optimization techniques =-=[13]-=-; loop fission (distribution) and loop fusion have also been found to be helpful [71]. Data layout optimizations such as padding and transpose have been shown to be useful in eliminating conflict miss improving spatial locality [21]. Several cache miss estimation techniques have been proposed to help guide data locality optimizations =-=[72, 13]-=-. Tiling has been proven useful for linear algebra codes [14, 16, 13] and multiple loop nests across time-step loops [20]. In comparison we apply tiling to 3D stencil codes which cannot be tiled with existing methods. Rivera in [18, 19] studied existing tiling techniqu   </text>
<query_num> 20722 </query_num>
<text>   ng forces link nodes to be traversed sequentially. Fortunately, researchers have developed cache-conscious heap allocation and transformation techniques to improve locality for pointer-based programs =-=[29, 30, 31]-=-. These techniques improve locality by assigning or changing the locations of dynamically allocated memory in ways designed to improve spatial locality. Examples of cacheconscious algorithms include r d memory allocator which allocates memory in a location near to a user-specified address. ccmalloc is a heuristic that reserves space for future allocation requests when allocating new blocks of data =-=[30]-=-. Using this memory allocator, multiple members of a linked list are thus more likely to be in adjacent memory locations. Not only does this take advantage of hardware prefetching of long cache lines, , and allocating current and future nodes close to it whenever possible. To increase the probability of nearby placement, ccmalloc reserves space for future data blocks when allocating the first node =-=[30]-=-. Figure 9 displays an example of cache-conscious memory allocation to improve locality for pointer-chasing codes. 11sBadawy, Aggarwal, Yeung, &amp; Tseng Note one problem with ccmalloc is that dynamic da nvestigated data layout transformations for pointer-based data structures [78, 79]. Chilimbi et al. investigated allocation-time and run-time techniques to improve locality for linked lists and trees =-=[30]-=-. In this paper, we propose further extensions. Calder et al. use profiling to guide layout of global and stack variables to avoid conflicts [29]. Carlisle et al. investigate parallel performance of p   </text>
<query_num> 20723 </query_num>
<text>   ng forces link nodes to be traversed sequentially. Fortunately, researchers have developed cache-conscious heap allocation and transformation techniques to improve locality for pointer-based programs =-=[29, 30, 31]-=-. These techniques improve locality by assigning or changing the locations of dynamically allocated memory in ways designed to improve spatial locality. Examples of cacheconscious algorithms include r hniques to improve locality for linked lists and trees [30]. In this paper, we propose further extensions. Calder et al. use profiling to guide layout of global and stack variables to avoid conflicts =-=[29]-=-. Carlisle et al. investigate parallel performance of pointer-based codes in Olden [80]. 9. Conclusion Software prefetching and locality optimizations are two promising techniques for addressing the p   </text>
<query_num> 20724 </query_num>
<text>   ns such as 29sBadawy, Aggarwal, Yeung, &amp; Tseng loop permutation and tiling are the primary optimization techniques [13]; loop fission (distribution) and loop fusion have also been found to be helpful =-=[71]-=-. Data layout optimizations such as padding and transpose have been shown to be useful in eliminating conflict misses and improving spatial locality [21]. Several cache miss estimation techniques have   </text>
<query_num> 20725 </query_num>
<text>   olumbian health care system, MST computes a minimum spanning tree, and EM3D simulates electromagnetic wave propagation through 3D objects. Health, MST, and EM3D are all from the olden benchmark suite =-=[34]-=-. For each application, we applied software prefetching and locality optimizations by hand, first in isolation, then in combination. We followed the algorithms described in Sections 3 and 4, applying   </text>
<query_num> 20726 </query_num>
<text>   ormations exist to solve the problem of improving the locality of such graphs [22, 23, 25, 24]. Our evaluation uses a technique called gpart that relies on hierarchical clustering to improve locality =-=[27, 28]-=-. gpart works in three steps. First, the graph formed by the index array computation hierarchically partitioned into roughly cache-sized chunks. Second, the partitioning is then used to reorder the da existing tiling techniques [73] and devised techniques to tile 3D scientific computations. Researchers have examined irregular computations mostly in the context of parallel computing, using run-time =-=[26, 27, 28]-=- and compiler [74] support to support accesses on message-passing multiprocessors. A few have also looked at techniques for improving locality [22, 23, 24, 25]. Metrics such as reference affinity have   </text>
<query_num> 20727 </query_num>
<text>   rations), tiling allows reuse to be exploited on all the tiled dimensions so that data in cache can be accessed multiple times before it is flushed. Tiling is very effective with linear algebra codes =-=[14, 15, 16, 17, 18]-=-, and has been been extended to handle stencil codes used in iterative PDE solvers as well [19, 20, 13]. A major problem with tiling is that limited cache associativity may cause data in a tile to be   computation in indexed array computations. �������������������������������������������� Previous research found tile size selection and array padding can be applied to avoid conflict misses in tiles =-=[14, 17, 18]-=-. Tile-size-selection algorithms carefully select tile dimensions tailored to individual array dimensions so that no conflicts occur. For 2D arrays, the Euclidean remainder algorithm may be used to qu zes [21]. Padding increases the size of leading array dimensions, increasing the range of non-conflicting tile shapes. It has proven to be very useful for improving tiling for 2D linear algebra codes =-=[18]-=-. To combine padding with tile size selection for 2D arrays, we can test a small set of pads and choose the best choice. For 3D tiles, we would need to evaluate a much larger space of possible pads, s nhanced algorithm is identical to the one described in Section 4.1 except we select a larger innermost tile dimension size (e.g. TI in Figure 6). Our tiling heuristic uses the Euclidean GCD algorithm =-=[14, 18]-=- to generate a series of non-conflicting tile sizes. Although tiles with a square aspect ratio typically achieve the best cache utilization, we can bias the selection towards taller tiles with greater .” Padding is introduced incrementally to the leading array dimension until the Euclidean GCD algorithm gives a conflictfree tile size whose leading dimension is at least equal to or larger than ”PD” =-=[18, 19]-=-. Figure 16 illustrates this algorithm, showing how the leading dimension of a tile ”H” is related to the prefetch distance ”PD,” how padding is introduced along the leading array dimension, and how a   </text>
<query_num> 20728 </query_num>
<text>   rations), tiling allows reuse to be exploited on all the tiled dimensions so that data in cache can be accessed multiple times before it is flushed. Tiling is very effective with linear algebra codes =-=[14, 15, 16, 17, 18]-=-, and has been been extended to handle stencil codes used in iterative PDE solvers as well [19, 20, 13]. A major problem with tiling is that limited cache associativity may cause data in a tile to be   dimensions through a simple recurrence [14, 18]. An alternative algorithm finds non-conflicting 2D tile using an greedy algorithm which expands tile dimensions while checking that no conflicts occur =-=[16]-=-. We can adapt this algorithm for finding non-conflicting 3D tiles by iteratively attempting to increase each tile dimension until none may be increased without introducing conflicts [19]. Tile size s improving spatial locality [21]. Several cache miss estimation techniques have been proposed to help guide data locality optimizations [72, 13]. Tiling has been proven useful for linear algebra codes =-=[14, 16, 13]-=- and multiple loop nests across time-step loops [20]. In comparison we apply tiling to 3D stencil codes which cannot be tiled with existing methods. Rivera in [18, 19] studied existing tiling techniqu   </text>
<query_num> 20729 </query_num>
<text>   rmally exhibit striding. 8. Related Work Our work is most similar to Saavedra et al [43], which evaluated unimodular transformations, tiling, and software prefetching for matrix multiply. Mowry et al =-=[44]-=- also evaluated software prefetching and tiling for two scientific applications. In comparison, this paper focuses on memory trends and quantifies their impact on software prefetching and locality opt   </text>
<query_num> 20730 </query_num>
<text>   rt, how will our software techniques interact with hardware prefetching? To address these questions, we augmented our simulator to model a stride-based hardware prefetcher, similar to the one used in =-=[42]-=-. Our stride prefetcher consists of a 256-entry stride table and 8 stream buffers. The stride table observes the post-L1 miss stream to detect strides on a per-load instruction basis. When a striding  early as possible to maximize memory latency overlap. A third approach uses a correlation-based predictor in hardware, also known as a Markov predictor, to predict link node addresses for prefetching =-=[54, 55, 42]-=-. A fourth approach uses a special allocation technique to allocate nodes contiguously in memory which enables indexed access to the link nodes. This approach was first proposed in [8] and is called d   </text>
<query_num> 20731 </query_num>
<text>   s in isolation. Software prefetching for affine array accesses has been studied in [4, 5, 6]. Hardware prefetching techniques 28sEfficacy of Software Prefetching and Locality Optimizations studied in =-=[40, 45, 41, 46, 47]-=- are similarly limited to affine array accesses, but use hardware to identify the access pattern automatically. Prefetch engines for affine array accesses [48, 49, 50, 51] provide hardware support for   </text>
<query_num> 20732 </query_num>
<text>   s, called jump pointers, into dynamic data structures to connect non-consecutive link elements [52, 7, 8, 11], as described in Section 3.3. Another approach uses only natural pointers for prefetching =-=[53, 8, 9, 10]-=-. These techniques prefetch pointer chains sequentially, but schedule each prefetch as early as possible to maximize memory latency overlap. A third approach uses a correlation-based predictor in hard   </text>
<query_num> 20733 </query_num>
<text>   tures tolerate latency by means of context switching between different processes or threads of control. Context switches can be triggered by a particular event like a cache miss [66] (the MIT Alewife =-=[67]-=- is a machine that makes use of such an idea), or they can be made on every cycle [68]. More recently, researchers have investigated Simultaneous Multi-Threading [69] in which instructions from multip   </text>
<query_num> 20734 </query_num>
<text>   ware prefetching. Specifically, EM3D simulates alternating electric and magnetic fields characteristic of electro-magnetic wave propagation via two arrays, one for “e-nodes” and another for “h-nodes” =-=[35]-=-. “Edge pointers” connect e-nodes to h-nodes and vice versa to form a bipartite graph, representing a mesh over which the electro-magnetic fields are simulated [36]. Like MST, the traversal code for e   </text>
<query_num> 20735 </query_num>
<text>   “e-nodes” and another for “h-nodes” [35]. “Edge pointers” connect e-nodes to h-nodes and vice versa to form a bipartite graph, representing a mesh over which the electro-magnetic fields are simulated =-=[36]-=-. Like MST, the traversal code for each edge pointer is short, potentially limiting prefetching. However, because the edge pointers are rooted inside e-node and h-node array elements (and because thes   </text>
</top>
<top>
<paper_num> 208 </paper_num>
<paper_title>   Back to the roots: a probabilistic framework for query-performance prediction.  </paper_title>
<paper_abstract>   The query-performance prediction task is estimating the effectiveness of a search performed in response to a query when no relevance judgments are available. Although there exist many effective prediction methods, these differ substantially in their basic principles, and rely on diverse hypotheses about the characteristics of effective retrieval. We present a novel fundamental probabilistic prediction framework. Using the framework, we derive and explain various previously proposed prediction methods that might seem completely different, but turn out to share the same formal basis. The derivations provide new perspectives on several predictors (e.g., Clarity). The framework is also used to devise new prediction approaches that outperform the state-of-the-art.  </paper_abstract>
<query_num> 20801 </query_num>
<text>   12 ACM 978-1-4503-1156-4/12/10 ...$15.00. ods assume that effective retrieval should be manifested in a result list that is robust with respect to query perturbations [37, 42], document perturbations =-=[36, 41]-=-, and retrieval method perturbations [2]. Another class of prediction methods is based on various analyses of retrieval scores in the result list [35, 14, 42, 30, 12, 13]. Given the large variety of p el perspective about the actual property of the result list that Clarity quantifies. Furthermore, we use the framework to show that some predictors that are based on the result-list robustness notion =-=[37, 36, 41, 42]-=- and on analysis of the retrieval-scores distribution [14], implicitly share the same formal basis for prediction. Our framework also provides a formal ground to using, for prediction, measures of que  29, 25, 16, 39, 28] analyze the query expression, often using corpusbased information. Post-retrieval predictors also use information induced from the result list of the most highly ranked documents =-=[10, 1, 35, 37, 11, 5, 36, 2, 14, 42, 18, 30, 31, 7, 12, 13]-=-. As noted above, the framework we present sets formal probabilistic grounds to the integration of preretrieval and post-retrieval prediction. We empirically show that the integration yields predictio ormal grounds to several aspects of this framework; and, helps explain several prediction methods and techniques that do not naturally fit in this framework; e.g, those using reference document lists =-=[37, 36, 41, 14, 42]-=-. A prediction framework [31], based on statistical decision theory, uses multiple re-rankings of the result list and their estimated effectiveness. We show that our framework provides probabilistic g ., Clarity [10] and WIG [42]. In addition, the integration of pre-retrieval and post-retrieval prediction that emerges in our framework was not addressed. Notions of result list cohesion (dispersion) =-=[37, 36]-=- were argued to be correlated to some extent with retrieval effectiveness. We show that the formal prediction aspect targeted by measures of this query-independent result list property is complementar   </text>
<query_num> 20802 </query_num>
<text>   29, 25, 16, 39, 28] analyze the query expression, often using corpusbased information. Post-retrieval predictors also use information induced from the result list of the most highly ranked documents =-=[10, 1, 35, 37, 11, 5, 36, 2, 14, 42, 18, 30, 31, 7, 12, 13]-=-. As noted above, the framework we present sets formal probabilistic grounds to the integration of preretrieval and post-retrieval prediction. We empirically show that the integration yields predictio   </text>
<query_num> 20803 </query_num>
<text>   29, 25, 16, 39, 28] analyze the query expression, often using corpusbased information. Post-retrieval predictors also use information induced from the result list of the most highly ranked documents =-=[10, 1, 35, 37, 11, 5, 36, 2, 14, 42, 18, 30, 31, 7, 12, 13]-=-. As noted above, the framework we present sets formal probabilistic grounds to the integration of preretrieval and post-retrieval prediction. We empirically show that the integration yields predictio ework; and, helps explain several prediction methods and techniques that do not naturally fit in this framework; e.g, those using reference document lists [37, 36, 41, 14, 42]. A prediction framework =-=[31]-=-, based on statistical decision theory, uses multiple re-rankings of the result list and their estimated effectiveness. We show that our framework provides probabilistic grounds for the underlying pre ntegration of these two types of measures. Integrating predictors using a linear interpolation of prediction values was employed with either pre-retrieval predictors [15] or post-retrieval predictors =-=[36, 14, 42, 31]-=-. In contrast, our framework provides formal grounds to the integration of different types of predictors which target different formal aspects of prediction; namely, pre-retrieval and post-retrieval p uation 4. The underlying idea was to use documents in Dres as its proxies for estimating the result list likelihood, p(Dres|q, r). An alternative type of a proxy for Dres is a reference document list =-=[37, 41, 14, 42, 31]-=-. Formally, a similar formulation to that in Equation 3 that uses a set Sref of reference document lists (denoted Dref) is: Pref(Dres; q) def = X ˆp(Dres|Dref,r)ˆp(Dref|q, r). (6) D ref ∈S ref Thus, t  for. Thus, it is not surprising that QF was found to be a highly effective predictor for the effectiveness of the reference result list (Dref) and not only to that of the original result list (Dres) =-=[31]-=-. Indeed, this finding can be formally explained using Equation 6. That is, the effectiveness (relevance) of the reference list Dref can be estimated using Dres as its proxy, by switching their roles   </text>
<query_num> 20804 </query_num>
<text>   e effectiveness of a search performed in response to a query when there is a lack of relevance judgments. The prediction can be performed before retrieval using the query and corpus-based information =-=[19, 16]-=-. Post-retrieval prediction, on the other hand, also uses information induced from the result list of the most highly ranked documents [4]. Although there exists abundance of effective prediction meth nnections between them. The second contribution is using the framework to devise new prediction approaches that outperform state-of-the-art. 2. RELATED WORK Pre-retrieval query-performance predictors =-=[19, 29, 25, 16, 39, 28]-=- analyze the query expression, often using corpusbased information. Post-retrieval predictors also use information induced from the result list of the most highly ranked documents [10, 1, 35, 37, 11,  d post- retrieval prediction Estimating the prior probability of relevance to q in Equation 2, p(r|q), is (in spirit) the goal of pre-retrieval prediction methods that operate prior to retrieval time =-=[19, 16]-=-. Indeed, pre-retrieval predictors quantify this query difficulty notion using information induced only from q and the corpus. The probability p(Dres|q, r) is the likelihood of the result list Dres gi  QF and WIG methods described above. For pre-retrieval prediction we use two sets of methods, each based on a different type of statistics computed for a query term. The first set, referred to as IDF =-=[10, 19, 16]-=-, uses the inverse document frequency (IDF) value of a query term. The second set of pre-retrieval predictors, denoted VarTF.IDF [39], uses the variance of the TF.IDF value of a query term in document   </text>
<query_num> 20805 </query_num>
<text>   elations) are determined at the 95% confidence level [34]. As in many previous reports of work on predicting query performance [10, 11, 41, 14, 16, 30, 17, 31], we use the query likelihood (QL) model =-=[32]-=- for the retrieval method. The goal of the predictors we study is to estimate the retrieval effectiveness of this standard language-model-based retrieval approach. The QL retrieval score assigned to d   </text>
<query_num> 20806 </query_num>
<text>   erturbations [37, 42], document perturbations [36, 41], and retrieval method perturbations [2]. Another class of prediction methods is based on various analyses of retrieval scores in the result list =-=[35, 14, 42, 30, 12, 13]-=-. Given the large variety of prediction approaches, and the underlying hypotheses on which they are based, a few questions arise. The most fundamental one is whether there is a unified formal basis (f  29, 25, 16, 39, 28] analyze the query expression, often using corpusbased information. Post-retrieval predictors also use information induced from the result list of the most highly ranked documents =-=[10, 1, 35, 37, 11, 5, 36, 2, 14, 42, 18, 30, 31, 7, 12, 13]-=-. As noted above, the framework we present sets formal probabilistic grounds to the integration of preretrieval and post-retrieval prediction. We empirically show that the integration yields predictio cally significant differences of prediction quality (between Pearson correlations) are determined at the 95% confidence level [34]. As in many previous reports of work on predicting query performance =-=[10, 11, 41, 14, 16, 30, 17, 31]-=-, we use the query likelihood (QL) model [32] for the retrieval method. The goal of the predictors we study is to estimate the retrieval effectiveness of this standard language-model-based retrieval a   </text>
<query_num> 20807 </query_num>
<text>   erturbations [37, 42], document perturbations [36, 41], and retrieval method perturbations [2]. Another class of prediction methods is based on various analyses of retrieval scores in the result list =-=[35, 14, 42, 30, 12, 13]-=-. Given the large variety of prediction approaches, and the underlying hypotheses on which they are based, a few questions arise. The most fundamental one is whether there is a unified formal basis (f ty quantifies. Furthermore, we use the framework to show that some predictors that are based on the result-list robustness notion [37, 36, 41, 42] and on analysis of the retrieval-scores distribution =-=[14]-=-, implicitly share the same formal basis for prediction. Our framework also provides a formal ground to using, for prediction, measures of query-independent properties of the result list (e.g., cohesi  29, 25, 16, 39, 28] analyze the query expression, often using corpusbased information. Post-retrieval predictors also use information induced from the result list of the most highly ranked documents =-=[10, 1, 35, 37, 11, 5, 36, 2, 14, 42, 18, 30, 31, 7, 12, 13]-=-. As noted above, the framework we present sets formal probabilistic grounds to the integration of preretrieval and post-retrieval prediction. We empirically show that the integration yields predictio ormal grounds to several aspects of this framework; and, helps explain several prediction methods and techniques that do not naturally fit in this framework; e.g, those using reference document lists =-=[37, 36, 41, 14, 42]-=-. A prediction framework [31], based on statistical decision theory, uses multiple re-rankings of the result list and their estimated effectiveness. We show that our framework provides probabilistic g ntegration of these two types of measures. Integrating predictors using a linear interpolation of prediction values was employed with either pre-retrieval predictors [15] or post-retrieval predictors =-=[36, 14, 42, 31]-=-. In contrast, our framework provides formal grounds to the integration of different types of predictors which target different formal aspects of prediction; namely, pre-retrieval and post-retrieval p   </text>
<query_num> 20808 </query_num>
<text>   in effective retrieval. For example, some post-retrieval prediction methods rely on the premise that a result list that exhibits high clarity with respect to the corpus indicates effective retrieval =-=[10, 1, 11, 5, 18]-=-. Other prediction methPermission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for p  29, 25, 16, 39, 28] analyze the query expression, often using corpusbased information. Post-retrieval predictors also use information induced from the result list of the most highly ranked documents =-=[10, 1, 35, 37, 11, 5, 36, 2, 14, 42, 18, 30, 31, 7, 12, 13]-=-. As noted above, the framework we present sets formal probabilistic grounds to the integration of preretrieval and post-retrieval prediction. We empirically show that the integration yields predictio   </text>
<query_num> 20809 </query_num>
<text>   in effective retrieval. For example, some post-retrieval prediction methods rely on the premise that a result list that exhibits high clarity with respect to the corpus indicates effective retrieval =-=[10, 1, 11, 5, 18]-=-. Other prediction methPermission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for p derive various postretrieval predictors that might seem at first glance to rely on completely different hypotheses and use different principles. For example, we derive (explain) the Clarity predictor =-=[10]-=-. This derivation provides a novel perspective about the actual property of the result list that Clarity quantifies. Furthermore, we use the framework to show that some predictors that are based on th  29, 25, 16, 39, 28] analyze the query expression, often using corpusbased information. Post-retrieval predictors also use information induced from the result list of the most highly ranked documents =-=[10, 1, 35, 37, 11, 5, 36, 2, 14, 42, 18, 30, 31, 7, 12, 13]-=-. As noted above, the framework we present sets formal probabilistic grounds to the integration of preretrieval and post-retrieval prediction. We empirically show that the integration yields predictio re, our framework accounts for prediction aspects not accounted for by this framework [21]. (See Section 3.1.3 for a discussion.) Our framework also provides novel views of predictors such as Clarity =-=[10]-=- and WIG [42], which substantially depart from those previously proposed [21]. A conceptual framework for predicting query difficulty [5] is based on measuring similarities between the result list, th istic grounds for the underlying prediction principle of this framework. Furthermore, we use our framework to explain predictors that cannot be explained in terms of this approach [31]; e.g., Clarity =-=[10]-=- and WIG [42]. In addition, the integration of pre-retrieval and post-retrieval prediction that emerges in our framework was not addressed. Notions of result list cohesion (dispersion) [37, 36] were a   </text>
<query_num> 20810 </query_num>
<text>   k on prediction: “What is the probability that this result list is relevant to this query?”. This question is a generalization to the documentlist case of the core question of probabilistic retrieval =-=[33, 22, 27]-=-: “What is the probability that this document is relevant to this query?”. The framework we present sets the first formal grounds for integrating pre-retrieval and post-retrieval prediction methods. W  measures (e.g., average precision, precision at top ranks, and NDCG). Furthermore, the question just stated is a generalization to the document list case of the question posed by Sparck Jones et al. =-=[33]-=- with respect to a single document: “What is the probability that this document is relevant to this query”? This question directly connects to the probability ranking principle [26] that states that m s a single document, then the prediction task modeled in Equation 2 is based on estimating the likelihood of the document given the query. This is the basis of the probabilistic approach to retrieval =-=[33]-=-. 3.1.1 Integrating pre- and post- retrieval prediction Estimating the prior probability of relevance to q in Equation 2, p(r|q), is (in spirit) the goal of pre-retrieval prediction methods that opera or q; i.e., the probability that Dres is the list that provides information pertaining to q. This is the list-based generalization of the document likelihood principle used in probabilistic retrieval =-=[33]-=-. Estimating the list likelihood is the implicit goal of post-retrieval predictors. Hence, our second key observation is that while by design, post-retrieval predictors use result-list-based informati r) canbeestimated by using documents in Dres as Dres’s proxies. We use ˆp(·) to denote an estimate for p(·). Let p(d|q, r) bethe probability that d is the document relevant to q; i.e., d’s likelihood =-=[33]-=-. If we set ˆp(d|q, r) def = 0 for d ̸∈ Dres, i.e., a document not in the result list is not considered relevant; and, P ˆp(di|q, r) = 1 holds, that is, ˆp(di|q, r) isa di∈Dres probability distributio   </text>
<query_num> 20811 </query_num>
<text>   nnections between them. The second contribution is using the framework to devise new prediction approaches that outperform state-of-the-art. 2. RELATED WORK Pre-retrieval query-performance predictors =-=[19, 29, 25, 16, 39, 28]-=- analyze the query expression, often using corpusbased information. Post-retrieval predictors also use information induced from the result list of the most highly ranked documents [10, 1, 35, 37, 11,   </text>
<query_num> 20812 </query_num>
<text>   nnections between them. The second contribution is using the framework to devise new prediction approaches that outperform state-of-the-art. 2. RELATED WORK Pre-retrieval query-performance predictors =-=[19, 29, 25, 16, 39, 28]-=- analyze the query expression, often using corpusbased information. Post-retrieval predictors also use information induced from the result list of the most highly ranked documents [10, 1, 35, 37, 11,  mputed for a query term. The first set, referred to as IDF [10, 19, 16], uses the inverse document frequency (IDF) value of a query term. The second set of pre-retrieval predictors, denoted VarTF.IDF =-=[39]-=-, uses the variance of the TF.IDF value of a query term in documents across the corpus in which it appears. Predictors based on VarTF.IDF were shown to substantially outperform other pre-retrieval pre   </text>
<query_num> 20813 </query_num>
<text>   t hand, whose effectiveness we want to predict, uses the language-model-based surface-level similarity between documents and the query. This is the case for the query likelihood [32] and KL retrieval =-=[23]-=- methods that are commonly used in work on using Clarity for prediction. Hence, the induced ranking is essentially based on the document likelihood scores defined in Appendix A (i.e., normalized query   </text>
<query_num> 20814 </query_num>
<text>   t list attests to reduced query drift [36], and hence, to improved retrieval. The second is that list dispersion might indicate increased cover of query aspects, and thereby imply effective retrieval =-=[5]-=-. We employ two measures that quantify list cohesion and dispersion: the diameter of the list, adapted in spirit from [5], and the list entropy, a variant of a measure proposed in [20]. List diameter.   </text>
<query_num> 20815 </query_num>
<text>   the framework provides a unified formal basis that can be used to explain (derive), and provide new perspectives for, many previously proposed post-retrieval predictors. A recently proposed framework =-=[21]-=- explains several postretrieval predictors. The idea is that an effective result list is that which is similar to some pseudo effective list and dissimilar to a pseudo ineffective list. Our framework  ramework =-=[21]-=-. (See Section 3.1.3 for a discussion.) Our framework also provides novel views of predictors such as Clarity [10] and WIG [42], which substantially depart from those previously proposed [21]. A conceptual framework for predicting query difficulty [5] is based on measuring similarities between the result list, the query, and the corpus. Our framework provides formal grounds to several asp e list is estimated using previously proposed predictors. Thus, while UEF is based on statistical decision theory, it can be directly explained by Equation 6. A prediction framework recently proposed =-=[21]-=- relies on the premise that the result list Dres is relevant to the extent that it is similar to a pseudo relevant result list and dissimilar to a pseudo non-relevant result list. The framework was us ures of result list properties. Yet, in Appendix B we show that Equation 1, which served as the basis for deriving our framework, can be used to provide formal probabilistic grounds to this framework =-=[21]-=-. 4. EVALUATION We present an empirical exploration of three formal aspects that emerged in the development of our prediction framework. These give rise to new prediction methods and shed new light on   </text>
<query_num> 20816 </query_num>
<text>   uation 4. The underlying idea was to use documents in Dres as its proxies for estimating the result list likelihood, p(Dres|q, r). An alternative type of a proxy for Dres is a reference document list =-=[37, 41, 14, 42, 31]-=-. Formally, a similar formulation to that in Equation 3 that uses a set Sref of reference document lists (denoted Dref) is: Pref(Dres; q) def = X ˆp(Dres|Dref,r)ˆp(Dref|q, r). (6) D ref ∈S ref Thus, t iation measure. Explaining QF and autocorrelation. Devising the estimate ˆp(Dref|q, r) is a prediction problem in its own right, which is not addressed by most predictors that utilize reference lists =-=[37, 41, 2, 14, 42]-=-. For example, the query feedback (QF) predictor [42] uses the overlap at top ranks between Dres and a (single) list Dref for ˆp(Dres|Dref,r); Dref is retrieved from the corpus using a (relevance) lan  further support to the prediction principle presented in Equation 6. Other predictors that use reference lists without accounting for their presumed relevance include those using query perturbations =-=[37]-=-, document perturbations [41], and retrieval method perturbations [2, 14], to induce reference lists. Explaining additional prediction approaches. The UEF predictor [31] does estimate the presumed rel th reference lists Dref (ˆp(Dres|Dref,r)), where each Dref is weighted by its presumed relevance to q (ˆp(Dref|q, r)). As discussed in Section 3, this prediction paradigm underlies several predictors =-=[37, 41, 14, 42, 31]-=-. However, many of RefList(Uni) RefList(Clarity) Clarity TREC5 0.414 0.459 0.431 ROBUST 0.520 0.562 c 0.522 WT10G 0.415 0.418 0.432 GOV2 0.377 0.486 u 0.456 Clue09 −0.06 0.019 0.105 Clue09+SpamRm 0.12 ally significant differences with Clarity and RefList(Uni), respectively. Best result in a row is boldfaced. these predictors do not estimate, and consequently utilize, the presumed relevance of Dref =-=[37, 41, 14, 42]-=-. To further explore the merits, or lack thereof, of the referencelists-based prediction paradigm, when using both the association between Dres and Dref and the estimated relevance of Dref to q, we st   </text>
</top>
<top>
<paper_num> 209 </paper_num>
<paper_title>   Semi-Latent Dirichlet Allocation: A Hierarchical Model for Human Action Recognition.  </paper_title>
<paper_abstract>   Abstract. We propose a new method for human action recognition from video sequences using latent topic models. Video sequences are represented by a novel “bag-of-words ” representation, where each frame corresponds to a “word”. The major difference between our model and previous latent topic models for recognition problems in computer vision is that, our model is trained in a “semi-supervised ” way. Our model has several advantages over other similar models. First of all, the training is much easier due to the decoupling of the model parameters. Secondly, it naturally solves the problem of how to choose the appropriate number of latent topics. Thirdly, it achieves much better performance by utilizing the information provided by the class labels in the training set. We present action classification and irregularity detection results, and show improvement over previous methods. 1  </paper_abstract>
<query_num> 20901 </query_num>
<text>   , (d) ignore the ordering of words and represent the image sequences of a tracked person as a histogram over “motion words”. In particular, our model is based on the latent Dirichlet allocation (LDA) =-=[2]-=- model. LDA, the probabilistic Latent Semantic Indexing (pLSI) [13] model, and their variants have been applied to various computer vision applications, such as scene recognition [5, 10], object recog  patterns of motion and appearance. Our approach is closely related to a body of work on recognition using “bagof-words”. The “bag-of-words” model was originally proposed for analyzing text documents =-=[2,13]-=-. Recently, researchers in the computer vision community have used “bag-of-words” models for various recognition problems. Fei-Fei &amp; Perona [10] use a variant of LDA for natural scene categorization.  nces are converted to the “bag-of-words” representation by replacing each frame by its corresponding codeword. 3.2 Latent Dirichlet Allocation Our model is based the Latent Dirichlet Allocation (LDA) =-=[2]-=-. In the following, we briefly introduce LDA model using the terminology in our context. Suppose we are given a collection D of video sequences {w1,w2, ...,wM}. Each video sequence w is a collection o hoose a motion word wn from wn ∼ p(wn|zn, β), a multinomial probability conditioned on zn. α θ β z w N M (a) (b) γ θ φ z N M Fig.3. (a) Graphical representation of LDA model, adopted from Blei et al. =-=[2]-=-; (b) Graphical representation of the variational distribution. The parameter θ indicates the mixing proportion of different actions labels in a particular video sequence. α is the parameter of a Diri  finding α and β that maximize the log likelihood of the data l(α, β) = � M d=1 logP(wd|α, β). This parameter estimation problem can be solved by the variational EM algorithm developed in Blei et al. =-=[2]-=-. 3.3 Semi-Latent Dirichlet Allocation In the original LDA, we are only given the word (w1, w2, ..., wN) in each video sequence, but we do not know the topic zi for the word wi, nor the mixing proport   </text>
<query_num> 20902 </query_num>
<text>   I) [13] model, and their variants have been applied to various computer vision applications, such as scene recognition [5, 10], object recognition [11,22,25], action recognition [19], human detection =-=[1]-=-, etc. Despite the great success achieved, there are some unsolved, important issues remaining in this line of research. First of all, it is not clear how to choose the right number of latent topics i tures used in these approaches are usually SIFT-like local features computed at locations found by interest-point detectors. The only exceptions are histogram of oriented gradients in Bissacco et al. =-=[1]-=- and multiple segmentations in Russell et al. [22]. Features based on local patches may be appropriate for certain recognition problems, such as scene recognition or object recognition. But for humans ] and Russell et al. [22] use pLSI for unsupervised object class recognition and segmentation. Niebles et al. [19] use pLSI for action recognitionsusing spatial-temporal visual words. Bissacco et al. =-=[1]-=- use LDA for human pose classification from vector-quantized words from histograms of oriented gradients. 3 Our Approach Similar to Niebles et al. [19], we represent a video sequence as a “bag of word   </text>
<query_num> 20903 </query_num>
<text>   LDA, the probabilistic Latent Semantic Indexing (pLSI) [13] model, and their variants have been applied to various computer vision applications, such as scene recognition [5, 10], object recognition =-=[11,22,25]-=-, action recognition [19], human detection [1], etc. Despite the great success achieved, there are some unsolved, important issues remaining in this line of research. First of all, it is not clear how ike local features computed at locations found by interest-point detectors. The only exceptions are histogram of oriented gradients in Bissacco et al. [1] and multiple segmentations in Russell et al. =-=[22]-=-. Features based on local patches may be appropriate for certain recognition problems, such as scene recognition or object recognition. But for humansaction recognition, it is not clear that they can  have used “bag-of-words” models for various recognition problems. Fei-Fei &amp; Perona [10] use a variant of LDA for natural scene categorization. Sivic et al. [25], Fergus et al. [11] and Russell et al. =-=[22]-=- use pLSI for unsupervised object class recognition and segmentation. Niebles et al. [19] use pLSI for action recognitionsusing spatial-temporal visual words. Bissacco et al. [1] use LDA for human pos   </text>
<query_num> 20904 </query_num>
<text>   Semantic Indexing (pLSI) [13] model, and their variants have been applied to various computer vision applications, such as scene recognition [5, 10], object recognition [11,22,25], action recognition =-=[19]-=-, human detection [1], etc. Despite the great success achieved, there are some unsolved, important issues remaining in this line of research. First of all, it is not clear how to choose the right numb tioned issues in two aspects. First of all, we introduce a new “bag-of-words” representation for image sequences. Our representation is dramatically different from previous ones (e.g., Niebles et al. =-=[19]-=-) in that we represent a frame in an image sequence as a “single word”, rather than a “collection of words” computed at some spatialtemporal interest points. Our main motivation for this new represent se a variant of LDA for natural scene categorization. Sivic et al. [25], Fergus et al. [11] and Russell et al. [22] use pLSI for unsupervised object class recognition and segmentation. Niebles et al. =-=[19]-=- use pLSI for action recognitionsusing spatial-temporal visual words. Bissacco et al. [1] use LDA for human pose classification from vector-quantized words from histograms of oriented gradients. 3 Our set), the recognition accuracy is similar. Table 1. Comparison of different methods in terms of recognition accuracy on the KTH dataset methods recognition accuracy(%) Our method 92.43 Niebles et al. =-=[19]-=- 81.50 Dollár et al. [7] 81.17 Schuldt et al. [24] 71.72 Ke et al. [14] 62.96 4.2 Action Classification on Soccer Dataset The soccer dataset we use is from Efros et al. [8] 1 . This dataset contains s   </text>
<query_num> 20905 </query_num>
<text>   allocation (LDA) [2] model. LDA, the probabilistic Latent Semantic Indexing (pLSI) [13] model, and their variants have been applied to various computer vision applications, such as scene recognition =-=[5, 10]-=-, object recognition [11,22,25], action recognition [19], human detection [1], etc. Despite the great success achieved, there are some unsolved, important issues remaining in this line of research. Fi  was originally proposed for analyzing text documents [2,13]. Recently, researchers in the computer vision community have used “bag-of-words” models for various recognition problems. Fei-Fei &amp; Perona =-=[10]-=- use a variant of LDA for natural scene categorization. Sivic et al. [25], Fergus et al. [11] and Russell et al. [22] use pLSI for unsupervised object class recognition and segmentation. Niebles et al   </text>
<query_num> 20906 </query_num>
<text>   analysis, human-computer interaction, surveillance and security, environmental control and monitoring, sport and entertainment analysis, etc. Various visual cues (e.g., motion [6, 8,16,20] and shape =-=[26]-=-) can be used for recognizing actions. In this paper, we focus on recognizing the action of a person in an image sequence based on motion cues. We develop a novel model of human actions based on the “ s for gait recognition. Rao et al. [21] describe a view-invariant representation for 2D trajectories of tracked skin blobs. Others consider the shape of human figure. For example, Sullivan &amp; Carlsson =-=[26]-=- use “order structure” to compare the shape of extracted edges for the purpose of action recognition. There is also work using both motion and shape cues. For example, Bobick &amp; Davis [3] use a represe   </text>
<query_num> 20907 </query_num>
<text>   cal analysis, ergonomic analysis, human-computer interaction, surveillance and security, environmental control and monitoring, sport and entertainment analysis, etc. Various visual cues (e.g., motion =-=[6, 8,16,20]-=- and shape [26]) can be used for recognizing actions. In this paper, we focus on recognizing the action of a person in an image sequence based on motion cues. We develop a novel model of human actions  Previous Work A lot of work has been done in recognizing actions from both still images and video sequences. Much of this work is focused on analyzing patterns of motion. For example, Cutler &amp; Davis =-=[6]-=-, and Polana &amp; Nelson [20] detect and classify periodic motions. Little &amp; Boyd [16] analyze the periodic structure of optical flow patterns for gait recognition. Rao et al. [21] describe a view-invari   </text>
<query_num> 20908 </query_num>
<text>   cal analysis, ergonomic analysis, human-computer interaction, surveillance and security, environmental control and monitoring, sport and entertainment analysis, etc. Various visual cues (e.g., motion =-=[6, 8,16,20]-=- and shape [26]) can be used for recognizing actions. In this paper, we focus on recognizing the action of a person in an image sequence based on motion cues. We develop a novel model of human actions ork has been done in recognizing actions from both still images and video sequences. Much of this work is focused on analyzing patterns of motion. For example, Cutler &amp; Davis [6], and Polana &amp; Nelson =-=[20]-=- detect and classify periodic motions. Little &amp; Boyd [16] analyze the periodic structure of optical flow patterns for gait recognition. Rao et al. [21] describe a view-invariant representation for 2D   </text>
<query_num> 20909 </query_num>
<text>   cal analysis, ergonomic analysis, human-computer interaction, surveillance and security, environmental control and monitoring, sport and entertainment analysis, etc. Various visual cues (e.g., motion =-=[6, 8,16,20]-=- and shape [26]) can be used for recognizing actions. In this paper, we focus on recognizing the action of a person in an image sequence based on motion cues. We develop a novel model of human actions utilizing the class labels, we can greatly simplify the training algorithm, and achieve much better recognition accuracy. 3.1 Motion Features and Codebook We use the motion descriptor in Efros et al. =-=[8]-=- to represent the video sequences. This motion descriptor has been shown to perform reliably with noisy image sequences, and has been applied in various tasks, such as action classification, motion sy ame B are b1, b2, b3 and b4, then the similarity between frame A and frame B is: S(A, B) = 4� � c=1 x,y∈I ac(x, y)bc(x, y) (1) where I is the spatial extent of the motion descriptors. In Efros et al. =-=[8]-=-, a temporal smoothing is also used, but we found the simplified version without temporal smoothing works good enough for our application.soriginal image optical flow F Fx, Fy F + x , F − x , F + y ,   video sequence is an indicator of “irregularity”. Lower likelihood means being more “irregular”. 4 Experiments We test our algorithm on two datasets: KTH human motion dataset [24] and soccer dataset =-=[8]-=-.sFig.5. Representative frames in KTH dataset 4.1 Action Classification on KTH Dataset The KTH human motion dataset is one of the largest video datasets of human actions. It contains six types of huma ethod 92.43 Niebles et al. [19] 81.50 Dollár et al. [7] 81.17 Schuldt et al. [24] 71.72 Ke et al. [14] 62.96 4.2 Action Classification on Soccer Dataset The soccer dataset we use is from Efros et al. =-=[8]-=- 1 . This dataset contains several minutes of digitized World Cup football game from an NTSC video tape. A preprocessing step is taken to track and stabilize each human figure. In the end, we obtain 3   </text>
<query_num> 20910 </query_num>
<text>   es a representation of a document in the topic simplex. Also notice that Dir(γ ∗ (w)) is the distribution from which the mixing proportion θ for the document w is drawn. We can imagine that if we j=1 =-=(3)-=-sdraw a sample θ ∼ Dir(γ ∗ (w)), θ will tend to peak towards the true mixing proportion θ ∗ of topics for the document w. So the true mixing proportion θ ∗ can be approximated by the empirical mean of   </text>
<query_num> 20911 </query_num>
<text>   es of a tracked person as a histogram over “motion words”. In particular, our model is based on the latent Dirichlet allocation (LDA) [2] model. LDA, the probabilistic Latent Semantic Indexing (pLSI) =-=[13]-=- model, and their variants have been applied to various computer vision applications, such as scene recognition [5, 10], object recognition [11,22,25], action recognition [19], human detection [1], et  patterns of motion and appearance. Our approach is closely related to a body of work on recognition using “bagof-words”. The “bag-of-words” model was originally proposed for analyzing text documents =-=[2,13]-=-. Recently, researchers in the computer vision community have used “bag-of-words” models for various recognition problems. Fei-Fei &amp; Perona [10] use a variant of LDA for natural scene categorization.   </text>
<query_num> 20912 </query_num>
<text>   example, Cutler &amp; Davis [6], and Polana &amp; Nelson [20] detect and classify periodic motions. Little &amp; Boyd [16] analyze the periodic structure of optical flow patterns for gait recognition. Rao et al. =-=[21]-=- describe a view-invariant representation for 2D trajectories of tracked skin blobs. Others consider the shape of human figure. For example, Sullivan &amp; Carlsson [26] use “order structure” to compare t   </text>
<query_num> 20913 </query_num>
<text>   has a lot of potential applications in surveillance and monitoring. Previous approaches to irregularity detection can be broadly classified into two classes: rule-based method and statistical methods =-=[4]-=-. Our method falls into the statistical methods, which try to learn a model of regularity from data, and infer about irregularity using the model. There are various notions of “irregularity”. For exam nusual and suspicious behavior. This irregularity is characterized by the unique combination of regular actions. Other irregularity detection algorithms using only low-level cues (e.g. Boiman &amp; Irani =-=[4]-=-) would not be able to identify it. The application of our method to irregularity detection is quite straightforward. We first build our S-LDA model from a collection of training video sequences that   </text>
<query_num> 20914 </query_num>
<text>   hood of this new testing video sequence is an indicator of “irregularity”. Lower likelihood means being more “irregular”. 4 Experiments We test our algorithm on two datasets: KTH human motion dataset =-=[24]-=- and soccer dataset [8].sFig.5. Representative frames in KTH dataset 4.1 Action Classification on KTH Dataset The KTH human motion dataset is one of the largest video datasets of human actions. It con . Comparison of different methods in terms of recognition accuracy on the KTH dataset methods recognition accuracy(%) Our method 92.43 Niebles et al. [19] 81.50 Dollár et al. [7] 81.17 Schuldt et al. =-=[24]-=- 71.72 Ke et al. [14] 62.96 4.2 Action Classification on Soccer Dataset The soccer dataset we use is from Efros et al. [8] 1 . This dataset contains several minutes of digitized World Cup football gam   </text>
<query_num> 20915 </query_num>
<text>   lassification, motion synthesis, etc. To calculate the motion descriptor, we first need to track and stabilize the persons in a video sequence. We use the human detection method in Sabzmeydani &amp; Mori =-=[23]-=- in some of our experiments. But any tracking or human detection methods can be used, since the motion descriptor we use is very robust to jitters introduced by the tracking. Given a stabilized video  nd indoors. Representative frames of this dataset are shown in Fig. 5. We first run an automatic preprocessing step to track and stabilize the video sequences using the algorithm in Sabzmeydani &amp;Mori =-=[23]-=-, so that all the figures appear in the center of the field of view. We perform leave-one-out crossvalidation on this dataset. For each run, we choose the video sequences of one subject as the test se   </text>
<query_num> 20916 </query_num>
<text>   roduced by the tracking. Given a stabilized video sequence in which the person of interest appears in the center of the field of view, we compute the optical flow at each frame using the Lucas-Kanade =-=[17]-=- algorithm. The optical flow vector field F is then split into two scalar fields Fx and Fy, corresponding to the x and y components of F. Fx and Fy are further half-wave rectified into four non-negati   </text>
<query_num> 20917 </query_num>
<text>   rts are certainly not “correct” ones, for example they only model a few parts of objects and often ignore much structure, they have been demonstrated to be quite effective in object recognition tasks =-=[9,12,15]-=-.sIn this paper we explore the use of a similar model, for recognizing human actions. Figure 1 shows an overview of our “bag-of-words” representation. In our model, each frame in an image sequence is   </text>
<query_num> 20918 </query_num>
<text>   sing both motion and shape cues. For example, Bobick &amp; Davis [3] use a representation known as “temporal templates” to capture both motion and shape, represented as evolving silhouettes. Zhong et al. =-=[27]-=- cluster segments of long video sequences by looking at co-occurrences of patterns of motion and appearance. Our approach is closely related to a body of work on recognition using “bagof-words”. The “   </text>
</top>
<top>
<paper_num> 210 </paper_num>
<paper_title>   Issues in Reasoning about Interaction Networks in Cells: Necessity of Event Ordering Knowledge.  </paper_title>
<paper_abstract>   In this paper we discuss several representation issues that we came across while modelling molecular interactions in cells of living organisms. One of the issues was that the triggering of events inside cells, an important modelling component, are not necessarily immediate, leading to multiple evolution models in the absence of additional information. Second, often an action or a trigger at one level of granularity of representation can be elaborated and refined. We show the problem that existing representation and modelling formalisms have in dealing with the above issues. We then present an action language which builds up on a previous language, and has the ability to express event ordering knowledge. We show that our language is able to adequately address the above-mentioned issues. Motivation  </paper_abstract>
<query_num> 21001 </query_num>
<text>   =-=, &amp; Shapiro 2001; Giordano, Martelli, &amp; Schwind 2001; Khan et al. 2003; Talcott et al. 2004; Chabrier et al. 2004; Giunchiglia et al. 2004; Calvanese, de Giacomo, &amp; Vardi. 2002; Fukuda &amp; Takagi 2001;agi 2001; Reiter 1996-=-). ¬ f, ¬ g a’ ¬ g a’ a f ¬ g, f’ f’ ¬ f, ¬ f’ (A) (B) g b f’ ¬ f, ¬ f’ Figure 1: Different interpretations of the same process. (A) Original interpretation, regardless the details of ¬f, ¬g triggers  inement, and do not allow for specification of event ordering. There exist action formalisms for representing triggered events and concurrent actions, such as situation calculus with natural actions (=-=Reiter 1996-=-), the language E (=-=Kakas &amp; Miller 1998-=-), the language C+ (=-=Giunchiglia et al. 2004-=-), and (=-=Thielscher 2000-=-), but none of them address the issues studied in this paper, in particular the issue of elabora   </text>
<query_num> 21002 </query_num>
<text>   =-=ral 2004-=-) and in almost all other approaches (=-=Reddy, Liebman, &amp; Mavrovouniotis 1996; Peleg, Yeh, &amp; Altman 2002; Regev, Silverman, &amp; Shapiro 2001; Giordano, Martelli, &amp; Schwind 2001; Khan et al. 2003; Talcott et al. 2004; Chabrier et al. 2004; Giunchiglia et al. 2004; Calvanese, de Giacomo, &amp; Vardi. 2002; Fukuda &amp; Takagi 2001; Reiter 1996-=-). ¬ f, ¬ g a’ ¬ g a’ a f ¬ g, f’ f’ ¬ f, ¬ f’ (A) (B) g b f’ ¬ f, ¬ f’ Figure 1   </text>
<query_num> 21003 </query_num>
<text>   ntailment Let τ be a trajectory model of a theory (D, E, I). A query Q is entailed by τ iff Q is entailed by the sequence of states 〈s0(τ), s1(τ), . . . , si(τ), . . .〉 by the standard LTL semantics (=-=Emerson 1990-=-). A theory (D, E, I) is called consistent if it has at least one model. A query Q is weakly entailed by a consistent theory (D, E, I) if it is entailed by a model of (D, E, I). The weak entailment is   </text>
<query_num> 21004 </query_num>
<text>   s (=-=Reddy, Liebman, &amp; Mavrovouniotis 1996; Peleg, Yeh, &amp; Altman 2002; Regev, Silverman, &amp; Shapiro 2001; Giordano, Martelli, &amp; Schwind 2001; Khan et al. 2003; Talcott et al. 2004; Chabrier et al. 2004; Giunchiglia et al. 2004; Calvanese, de Giacomo, &amp; Vardi. 2002; Fukuda &amp; Takagi 2001; Reiter 1996-=-). ¬ f, ¬ g a’ ¬ g a’ a f ¬ g, f’ f’ ¬ f, ¬ f’ (A) (B) g b f’ ¬ f, ¬ f’ Figure 1: Different interpretations of the same process exist action formalisms for representing triggered events and concurrent actions, such as situation calculus with natural actions (=-=Reiter 1996-=-), the language E (=-=Kakas &amp; Miller 1998-=-), the language C+ (=-=Giunchiglia et al. 2004-=-), and (=-=Thielscher 2000-=-), but none of them address the issues studied in this paper, in particular the issue of elaboration of actions and triggers at one granularity via more detailed description at   </text>
<query_num> 21005 </query_num>
<text>   specification of event ordering. There exist action formalisms for representing triggered events and concurrent actions, such as situation calculus with natural actions (=-=Reiter 1996-=-), the language E (=-=Kakas &amp; Miller 1998-=-), the language C+ (=-=Giunchiglia et al. 2004-=-), and (=-=Thielscher 2000-=-), but none of them address the issues studied in this paper, in particular the issue of elaboration of actions and triggers at one gr   </text>
<query_num> 21006 </query_num>
<text>   st all other approaches (=-=Reddy, Liebman, &amp; Mavrovouniotis 1996; Peleg, Yeh, &amp; Altman 2002; Regev, Silverman, &amp; Shapiro 2001; Giordano, Martelli, &amp; Schwind 2001; Khan et al. 2003; Talcott et al. 2004;Chabrier et al. 2004; Giunchiglia et al. 2004; Calvanese, de Giacomo, &amp; Vardi. 2002; Fukuda &amp; Takagi 2001; Reiter 1996-=-). ¬ f, ¬ g a’ ¬ g a’ a f ¬ g, f’ f’ ¬ f, ¬ f’ (A) (B) g b f’ ¬ f, ¬ f’ Figure 1: Different interpreta   </text>
</top>
<top>
<paper_num> 211 </paper_num>
<paper_title>   Approximation Techniques for Spatial Data.  </paper_title>
<paper_abstract>   Spatial Database Management Systems (SDBMS), e.g., Geographical Information Systems, that manage spatial objects such as points, lines, and hyper-rectangles, often have very high query processing costs. Accurate selectivity estimation during query optimization therefore is crucially important for finding good query plans, especially when spatial joins are involved. Selectivity estimation has been studied for relational database systems, but to date has only received little attention in SDBMS. In this paper, we introduce novel methods that permit high-quality selectivity estimation for spatial joins and range queries. Our techniques can be constructed in a single scan over the input, handle inserts and deletes to the database incrementally, and hence they can also be used for processing of streaming spatial data. In contrast to previous approaches, our techniques return approximate results that come with provable probabilistic quality guarantees. We present a detailed analysis and experimentally demonstrate the e#cacy of the proposed techniques.  </paper_abstract>
<query_num> 21101 </query_num>
<text>   . Let r and s be rectangles from R and S, respectively. The counting procedure adds the following values: number of corners of 2 The spatial relationships are slightly different from the ones used in =-=[24, 25]-=- since we selected them to correspond to the sketches we use.s(3, 3) overlap (3, 4) overlap (2, 3) no overlap (4, 5) overlap Figure 4: Spatial relationships between rectangles r which are covered by s   </text>
<query_num> 21102 </query_num>
<text>   . Let r and s be rectangles from R and S, respectively. The counting procedure adds the following values: number of corners of 2 The spatial relationships are slightly different from the ones used in =-=[24, 25]-=- since we selected them to correspond to the sketches we use.s(3, 3) overlap (3, 4) overlap (2, 3) no overlap (4, 5) overlap Figure 4: Spatial relationships between rectangles r which are covered by s  0.6 0.5 0.4 0.3 0.2 0.1 Relative Error for the LANDO+SOIL spatial join SKETCH EH GH 0 0 5 10 15 20 25 30 35 40 Space Allocated (K words) Figure 11: LANDO ⊲⊳o SOIL framework based on Euler histograms =-=[25]-=-. They also use a regular grid partitioning of the space, but the Euler histograms together with adaptively selected per-cell estimation techniques provide more accurate estimates for spatial joins wi   </text>
<query_num> 21103 </query_num>
<text>   2D spatial datasets. We compare our algorithms with recently proposed techniques based on generalized Euler Histograms [26] (henceforth referred to as EH) and with the Geometric Histograms technique =-=[5]-=- (henceforth referred to as GH). These are currently the best known techniques for spatial join selectivity estimation. 3 The EH approach partitions the data space using a grid of a given level L. A g chniques with provable error bounds. 8. RELATED WORK The best known general techniques for selectivity estimation for joins involving spatial objects are based on sampling or on histograms. An et al. =-=[5]-=- examine different sampling techniques and propose new histogram-based approaches. Their results indicate that for achieving a comparable accuracy, histograms will require less storage and estimation   </text>
<query_num> 21104 </query_num>
<text>   6] with their corresponding error guarantees suitably adapted to our spatial framework. In future work, we plan to extend our techniques to other spatial queries, including non-rectangular selections =-=[1]-=-; and we would like to incorporate quality boosting techniques such as sketch partitioning into our framework. One especially promising candidate for future work are complex spatial queries that conta   </text>
<query_num> 21105 </query_num>
<text>   ches [4, 3] which are dynamically maintainable and provide approximation quality guarantees. Different types of sketches so far have been used for efficiently maintaining aggregates over data streams =-=[6]-=-. Example applications are complex aggregates [14], quantiles [17], frequent items [10, 11], multidimensional histograms [28], and graph statistics [7]. An upcoming paper by Tao et al. proposes to use   </text>
<query_num> 21106 </query_num>
<text>   d provide approximation quality guarantees. Different types of sketches so far have been used for efficiently maintaining aggregates over data streams [6]. Example applications are complex aggregates =-=[14]-=-, quantiles [17], frequent items [10, 11], multidimensional histograms [28], and graph statistics [7]. An upcoming paper by Tao et al. proposes to use sketches to improve the result quality for aggreg   </text>
<query_num> 21107 </query_num>
<text>   e dyadic intervals is at a different level. Lemma 4. A point c ∈ N is contained in an interval [a, b] iff there is exactly one dyadic interval δ ∈ D such that δ ∈ D([a, b]) and δ ∈ D([c]). Proof. See =-=[13]-=-. Instead of having a ξ-variable for each coordinate in N we will use a ξ-variable for each dyadic interval over N. For data set R, the corresponding atomic sketches for intervals and endpoints then a we define a random variable Z = (XIYE + XEYI)/2. This random variable is an unbiased estimator for the join cardinality. Lemma 5. Random variable Z has the expected value E[Z] = |R ⊲⊳o S|. Proof. See =-=[13]-=-. For our example in Figure 2 we have XI = ξ2 + ξ6, XE = 2ξ1 + ξ2 + ξ3 + ξ4 + ξ6, YI = ξ3 + ξ5, and YE = 2ξ1 + ξ2 + ξ3 + ξ5 + ξ7. For Z we therefore obtain Z = ((ξ2 + ξ6)(2ξ1 + ξ2 + ξ3 + ξ5 + ξ7) +(2ξ ith the above inequality we have: Var[Z] ≤ 1/16 · 8 · ( � � 2 SJ(Xw)SJ(Y ¯w)) w∈IE 2 Using Cauchy-Schwarz and the fact that SJ(R) = � w∈IE 2 SJ(Xw) and SJ(S) = � w∈IE 2 SJ(Yw) yields: For details see =-=[13]-=-. Var[Z] ≤ 1/16 (8SJ(R)SJ(S)) . Here SJ(R) = SJ(XII) + SJ(XIE) + SJ(XEI) + SJ(XEE), similarly for SJ(S). 4.2.2 The Overall Technique As for interval sketches, we boost the accuracy of the atomic recta ation to each dimension of the data space. Instead of using the endpoint transformation, we can alternatively adapt our original simple counting procedure to explicitly keep track of common endpoints =-=[13]-=-. 6. EXTENSIONS 6.1 Join of Hyper-Rectangles The one-and two-dimensional spatial join estimators generalize naturally to higher dimensionality d as the following theorem shows. Theorem 3. For d-dimens  for i �= j are independent of each other. Set Z = 2 −d � w∈IE d XwY ¯w. Then it holds that E[Z] = |R ⊲⊳o S| and Var[Z] ≤ 3d −1 4 d SJ(R)SJ(S). Proof. The proof is fairly involved and can be found=-= in [13]-=-. The atomic sketches are as defined in Section 3.2. As before, for w ∈ IE d we define ¯w as the string that is obtained from w by replacing I with E and vice versa (the “complement” of w). At a first   </text>
<query_num> 21108 </query_num>
<text>   estimation are approaches for estimating the cost (in terms of CPU or I/O) of specific operator implementations. Examples are cost models for range queries and index-supported joins for spatial data =-=[9, 20, 21, 22, 29, 30]-=-. None of the previous approaches provides any result quality guarantees. Samples are difficult to maintain in the presence of updates, especially deletes which could remove objects from the sample. S   </text>
<query_num> 21109 </query_num>
<text>   g of the space, but the Euler histograms together with adaptively selected per-cell estimation techniques provide more accurate estimates for spatial joins with geometric selections. Faloutsos et al. =-=[15]-=- and Belussi and Faloutsos [8] propose parametric methods for estimating the selectivity of ε-(self) joins of point-sets. For self-similar data sets they achieve good approximations using power laws a   </text>
<query_num> 21110 </query_num>
<text>   ge Z(1) Average Z(2) Average Z(k2) Median of the Z Figure 1: Boosting the accuracy to estimate! This is a common problem shared by previous sketching techniques (see Section 8) and Online Aggregation =-=[19, 18]-=-, in fact also by any experiment in the natural sciences where we want to estimate the error in measuring an unknown quantity. We can generally use simple approximation techniques that provide a lower   </text>
<query_num> 21111 </query_num>
<text>   guaranteed error bounds (not only AMS sketches) could be incorporated into our framework. Thus for example one could potentially use some of the sketching techniques proposed in upcoming publications =-=[12, 16]-=- with their corresponding error guarantees suitably adapted to our spatial framework. In future work, we plan to extend our techniques to other spatial queries, including non-rectangular selections [1   </text>
<query_num> 21112 </query_num>
<text>   kets and hencespoor approximation. Histogram-based techniques therefore typically quantize the data space and make assumptions about the distribution within a bucket in order to obtain good estimates =-=[23, 26]-=-. 5.2 Spatial Join with Common Endpoints The counting algorithms used so far relied on Assumption 1 for correctness. We can make this assumption hold for any data set as follows. Recall that the inter  propose new histogram-based approaches. Their results indicate that for achieving a comparable accuracy, histograms will require less storage and estimation time than sampling. Mamoulis and Papadias =-=[23]-=- take a more general approach of analytically estimating the selectivity of complex spatial queries, i.e., queries that combine selection and join operators. Their formulas are based on uniformity ass   </text>
<query_num> 21113 </query_num>
<text>   multidimensional histograms [28], and graph statistics [7]. An upcoming paper by Tao et al. proposes to use sketches to improve the result quality for aggregate queries over spatiotemporal point sets =-=[27]-=-. 9. CONCLUSION AND FUTURE WORK In this paper, we proposed a new framework for using spatial sketch techniques for approximately answering spatial queries. To the best of our knowledge, our technique   </text>
<query_num> 21114 </query_num>
<text>   ntees. Different types of sketches so far have been used for efficiently maintaining aggregates over data streams [6]. Example applications are complex aggregates [14], quantiles [17], frequent items =-=[10, 11]-=-, multidimensional histograms [28], and graph statistics [7]. An upcoming paper by Tao et al. proposes to use sketches to improve the result quality for aggregate queries over spatiotemporal point set   </text>
<query_num> 21115 </query_num>
<text>   rdinate domains, which are not uncommon in spatial applications, this quickly becomes costly and also results in high variance of the estimates. Hence we introduce dyadic spatial sketches. Similar to =-=[11, 17], -=-we partition the domain N into intervals of size 2 i . For simplicity let n = |N | be a power of 2, say 2 h for some positive integer h. 1 For each level 0 ≤ i ≤ h we partition N into 2 h−i interval imation quality guarantees. Different types of sketches so far have been used for efficiently maintaining aggregates over data streams [6]. Example applications are complex aggregates [14], quantiles =-=[17]-=-, frequent items [10, 11], multidimensional histograms [28], and graph statistics [7]. An upcoming paper by Tao et al. proposes to use sketches to improve the result quality for aggregate queries over   </text>
<query_num> 21116 </query_num>
<text>   rdinate domains, which are not uncommon in spatial applications, this quickly becomes costly and also results in high variance of the estimates. Hence we introduce dyadic spatial sketches. Similar to =-=[11, 17], -=-we partition the domain N into intervals of size 2 i . For simplicity let n = |N | be a power of 2, say 2 h for some positive integer h. 1 For each level 0 ≤ i ≤ h we partition N into 2 h−i interval ntees. Different types of sketches so far have been used for efficiently maintaining aggregates over data streams [6]. Example applications are complex aggregates [14], quantiles [17], frequent items =-=[10, 11]-=-, multidimensional histograms [28], and graph statistics [7]. An upcoming paper by Tao et al. proposes to use sketches to improve the result quality for aggregate queries over spatiotemporal point set   </text>
<query_num> 21117 </query_num>
<text>   rrent state of the art in estimating the sizes of spatial joins uses either sampling or histograms, where recent work has shown that histograms are superior for a wide range of possible query classes =-=[5, 26]-=-. However, existing histogram-based techniques still have significant drawbacks. First, they either give no or only very conservative worstcase error guarantees that are usually overly pessimistic in  kets and hencespoor approximation. Histogram-based techniques therefore typically quantize the data space and make assumptions about the distribution within a bucket in order to obtain good estimates =-=[23, 26]-=-. 5.2 Spatial Join with Common Endpoints The counting algorithms used so far relied on Assumption 1 for correctness. We can make this assumption hold for any data set as follows. Recall that the inter niques (henceforth referred to as SKETCH) on both synthetic and real life 1D and 2D spatial datasets. We compare our algorithms with recently proposed techniques based on generalized Euler Histograms =-=[26]-=- (henceforth referred to as EH) and with the Geometric Histograms technique [5] (henceforth referred to as GH). These are currently the best known techniques for spatial join selectivity estimation. 3 2 L equi-width cells. An Euler histogram allocates buckets not only for grid cells, but also for grid edges and grid vertices. Besides storing object counts in a cell, the generalized Euler histogram =-=[26]-=- also stores information such as the average height, width and area of the intersection regions between the objects and the cell. In terms of storage space, a generalized Euler histogram of level L us For uniform data (z = 0), the SKETCH and GH techniques perform similarly, with an average rel3 The authors would like to thank Chengyu Sun for providing the original EH code and the data sets used in =-=[26]-=-.sRelative Error 0.6 0.5 0.4 0.3 0.2 0.1 Relative Error Vs Dataset size [ zipf=0 (uniform)] SK EH GH 0 0 100 200 300 400 500 Dataset Size (K) Figure 5: Zipf = 0 ative error which is much lower than th   </text>
<query_num> 21118 </query_num>
<text>   sciences where we want to estimate the error in measuring an unknown quantity. We can generally use simple approximation techniques that provide a lower bound on E[Z] (“sanity bounds” as discussed in =-=[3]-=-), or use historic data, e.g., previously computed exact answers, to predict future values of E[Z]. The tradeoff here is that the tighter these bounds are, the stronger the quality guarantees returned  definition of the atomic sketches in the form of Equation 5. For this type of sketches (tug-of-war sketch) it can be shown that their variance is bounded by twice the product of their self-join size =-=[3]-=-: Var[XIYE] ≤ 2SJ(XI)SJ(YE) and Var[XEYI] ≤ 2SJ(XE)SJ(YI) . Together with Equation 7 we have Var[Z] ≤ 1/2( � SJ(XI)SJ(YE) + � SJ(XE)SJ(YI)) 2 . Since XI and XE together account for all dyadic interval are more appropriate for skewed data cannot be maintained efficiently and introduce an additional error when the buckets of the joined data sets do not align. Our techniques are based on AMS sketches =-=[4, 3]-=- which are dynamically maintainable and provide approximation quality guarantees. Different types of sketches so far have been used for efficiently maintaining aggregates over data streams [6]. Exampl   </text>
<query_num> 21119 </query_num>
<text>   small, pseudo-random sketches of the spatial dataset. The basic sketching technique was originally introduced for on-line self-join size estimation by Alon, Matias, and Szegedy in their seminal paper =-=[4]-=- and, as we demonstrate in our work, can be generalized to provide approximate answers to spatial join queries with explicit and tunable performance guarantees on the approximation error. The main cha ata for the exact computation of query results may be prohibitively expensive. The rest of the paper is organized as follows. Section 2 defines spatial queries and discusses the sketching approach of =-=[4]-=-. In Section 3 we introduce the basic atomic sketches which will be used to construct selectivity estimators. Selectivity estimation for joins of interval sets and rectangle sets are discussed in Sect sing f(i) to denote the frequency of attribute value i in R.A, we can rewrite query Q as Q = SJ(A) = � i∈dom(A) f(i)2 (i.e., the second moment of A). In their seminal paper, Alon, Matias, and Szegedy =-=[4]-=- prove that any deterministic algorithm that produces a tight approximation to SJ(A) requires at least Ω(|dom(A)|) bits of storage, rendering such solutions impractical for a data-stream setting. Inst l arrays) for the explicit construction of small sample spaces supporting four-wise independent random variables, such families can be efficiently constructed on-line using only O(log |dom(A)|) space =-=[4]-=-. More precisely, we do not explicitly store the ξi. Instead we store a single seed (for ξi with i of length k bits, the seed has length 2k+1 bits) for the whole ξ-family. Whenever a ξi is needed for  s value is generated on-the-fly from the seed in time linear in the seed size. 2.3 Boosting Accuracy To improve the quality of the estimation guarantees one can use a standard boosting technique (see =-=[4]-=-) that maintains several independent identically-distributed (i.i.d.) instantiations of a random variable and uses averaging and median-selection operators to boost accuracy and probabilistic confiden   </text>
</top>
<top>
<paper_num> 212 </paper_num>
<paper_title>   Mobile Resource Guarantees (project evaluation paper).  </paper_title>
<paper_abstract>   This paper summarises the main outcomes of the Mobile Resource Guarantees (MRG) project, a three year project funded by the EC under the FET proactive initiative on Global Computing, and discusses follow-up work in new projects that will build on these results. 1  </paper_abstract>
<query_num> 21201 </query_num>
<text>   a fully formalised infrastructure for reasoning about resource consumption. We now elaborate our approach on a simple static analysis of heap-space consumption based on the work by Cachera and Jensen =-=[6]-=-. The idea is to prove a constant upper bound on heap allocation, by proving that no function allocates memory in a loop. The goal is to detect such non-loop allocating cases and sepa10srate them from   </text>
<query_num> 21202 </query_num>
<text>   d mathematical proofs of resourcerelated properties which are by their very nature self-evident, unforgeable, and independent of trust networks. This is the “proof-carrying-code” approach to security =-=[12]-=-, which has become increasingly popular in recent years [7, 1, 14]. Typical application scenarios for such an infrastructure include the following. • A provider of a distributed computational power, f erived logic w.r.t. the core program logic.s3.2 Software infrastructure The overall structure of our software infrastructure is depicted in Figure 2 and is an instance of a general PCC infrastructure =-=[12]-=- with a code producer (left hand side) and a code consumer (right hand side). The main components on the producer side are a certifying compiler, which translates high-level Camelot programs into the   </text>
<query_num> 21203 </query_num>
<text>   e by their very nature self-evident, unforgeable, and independent of trust networks. This is the “proof-carrying-code” approach to security [12], which has become increasingly popular in recent years =-=[7, 1, 14]-=-. Typical application scenarios for such an infrastructure include the following. • A provider of a distributed computational power, for example a node in a computational Grid, may only be willing to   </text>
<query_num> 21204 </query_num>
<text>   e by their very nature self-evident, unforgeable, and independent of trust networks. This is the “proof-carrying-code” approach to security [12], which has become increasingly popular in recent years =-=[7, 1, 14]-=-. Typical application scenarios for such an infrastructure include the following. • A provider of a distributed computational power, for example a node in a computational Grid, may only be willing to  s have the form E ⊢ h,e ⇓ h ′ ,v, p, where E maps variables to values, h represents the pre-heap and h ′ the post-heap, and v is the result value, consuming p resources. The Foundational PCC approach =-=[1]-=- performs proofs directly on this level and thereby reduces the size of the trusted code base (TCB). Note that, since we formalise the entire hierarchy of logics in a theorem prover, we do not need to   </text>
<query_num> 21205 </query_num>
<text>   e goal is to detect such non-loop allocating cases and sepa10srate them from the rest, for which no guarantees are given. We use a fragment of a simple first-order, strict language similar to Camelot =-=[11]-=-, with lists as only composed data-type and expressions in let-normal-form, or ANF meaning arguments to functions must be variables (k are constants, x variables, f function name): e ∈ expr ::= k | x  sumption [4] that is built on top of the program logic, using our novel multi-layered logics approach. • A certifying compiler for the strict, first-order functional, object-oriented language Camelot =-=[11]-=-, integrated into a prototype proof-carrying-code infrastructure, which is available on-line [13]. • Advanced reasoning principles [9, 10] for resources, based on high-level type systems. New projects   </text>
<query_num> 21206 </query_num>
<text>   in the termination logic have the form ⊲ T {P} e ↓, meaning an expressioneterminates under the precondition P. On top of the general-purpose logic, a specialised logic (for example the heap logic of =-=[4]-=-) is defined that captures the specifics of a particular security policy. This logic uses a restricted format of assertions, called derived assertions, which reflects the information of the high-level es high-level Camelot programs into the Grail intermediate code and additionally generates a certificate of its heap consumption. The latter is formalised as a lemma in the heap space logic for Grail =-=[4]-=-, an abstract fragment of Java Virtual Machine (JVM) code. The Grail code is processed by an assembler, the Grail de-functionaliser (gdf), to generate JVM bytecode. This bytecode is transmitted togeth or this formalisation and for encoding the logics on top of it. • A resource aware program logic [2] for the bytecode language of the above virtual machine. • A specialised logic for heap consumption =-=[4]-=- that is built on top of the program logic, using our novel multi-layered logics approach. • A certifying compiler for the strict, first-order functional, object-oriented language Camelot [11], integr   </text>
<query_num> 21207 </query_num>
<text>   is also possible to write the top level program in Camelot, and call other JVM code from Camelot. This is particularly useful for accessing Java library functions, e.g. for GUI parts of the code. In =-=[15]-=- an extension of Camelot with object-oriented features is described. These extensions have been used in implementing a directory lookup application to be executed on a PDA, based on the MIDP standard   </text>
<query_num> 21208 </query_num>
<text>   l is of immediate relevance for the programming languages area, and many type-based inferences have been suggested. The case we have worked out in [2] is the Hofmann &amp; Jost type system for heap usage =-=[9]-=- and a simpler instance is given in the rest of this section. In our work, however, we give a general framework for tying such analyses into a fully formalised infrastructure for reasoning about resou strict, first-order functional, object-oriented language Camelot [11], integrated into a prototype proof-carrying-code infrastructure, which is available on-line [13]. • Advanced reasoning principles =-=[9, 10]-=- for resources, based on high-level type systems. New projects that build on the MRG infrastructure are: • EmBounded, an FET-Open STREP project funded by the EC, that aims to provide resource bounded   </text>
<query_num> 21209 </query_num>
<text>   n order to reduce the size of the certificate, albeit at the cost of increasing the TCB size. More specifically we have produced the following • A completely formalised virtual machine and cost model =-=[5]-=- for a JVM-like language. We have used Isabelle/HOL as theorem proving platform for this formalisation and for encoding the logics on top of it. • A resource aware program logic [2] for the bytecode l   </text>
<query_num> 21210 </query_num>
<text>   strict, first-order functional, object-oriented language Camelot [11], integrated into a prototype proof-carrying-code infrastructure, which is available on-line [13]. • Advanced reasoning principles =-=[9, 10]-=- for resources, based on high-level type systems. New projects that build on the MRG infrastructure are: • EmBounded, an FET-Open STREP project funded by the EC, that aims to provide resource bounded   </text>
<query_num> 21211 </query_num>
<text>   totype that covers the entire path of mobile code sent in a distributed system, a software infrastructure. A general overview of the project, developed about half-way through the project, is given in =-=[3]-=-. Objective 1 is the development of a framework in which certificates of resource consumption make formal sense. This consists of a cost model and a program logic for an appropriate virtual machine an   </text>
</top>
<top>
<paper_num> 213 </paper_num>
<paper_title>   PRIMME: preconditioned iterative multimethod eigensolver - methods and software description.  </paper_title>
<paper_abstract>   This paper describes the PRIMME software package for the solving large, sparse Hermitian and real symmetric eigenvalue problems. The difficulty and importance of these problems have increased over the years, necessitating the use of preconditioning and near optimally converging iterative methods. On the other hand, the complexity of tuning or even using such methods has kept them outside the reach of many users. Responding to this problem, our goal was to develop a general purpose software that requires minimal or no tuning, yet it provides the best possible robustness and efficiency. PRIMME is a comprehensive package that brings state-of-the-art methods from “bleeding edge ” to production, with a flexible, yet highly usable interface. We review the theory that gives rise to the near optimal methods GD+k and JDQMR, and present the various algorithms that constitute the basis of PRIMME. We also describe the software implementation, interface, and provide some sample experimental results.  </paper_abstract>
<query_num> 21301 </query_num>
<text>   RPACK is compiled with the g77 compiler. We link with the Apple vecLib library that includes optimized versions of BLAS/LAPACK libraries. We use 10 matrix problems, six from the University of Florida =-=[16]-=- and the FEAP [3] collections, one from vibrational analysis of molecular structures [75] and three standard 7-point 3D Laplacian matrices generated by SPARSKIT [58] with zero Dirichlet boundary condi   </text>
<query_num> 21302 </query_num>
<text>   X, is sufficient. Intuitively, projecting out u (m) from a very accurate preconditioner helps avoid the classical Davidson problems where the correction is almost completely in the direction of u (m) =-=[53]-=-. However, projecting out X does not serve the same purpose. Instead, one would hope that it produces a better conditioned correction equation. Our analysis in [68] showed that usually there is no sig   </text>
<query_num> 21303 </query_num>
<text>   if only matrix-vector multiplication is available, the software ARPACK by Lehoucq, Sorensen, and Yang has set the standard for good quality code that is easy to use with very little parameter tuning =-=[44]-=-. Yet, the implicitly restarted Lanczos method [64], on which ARPACK is based, does not converge optimally, and it cannot directly use preconditioning, which is required for very difficult problems. T for symmetric problems. The second is the BLOPEXs28 A. STATHOPOULOS and J. R. McCOMBS implementation of LOBPCG [38]. The third is ARPACK’s function dsaupd, which implements IRL for symmetric matrices =-=[44]-=-. Although ARPACK does not use preconditioning, it is included as the default benchmark for unpreconditioned cases. We have not compared with SLEPc methods as they do not allow for preconditioning. Al   </text>
<query_num> 21304 </query_num>
<text>   mmetric and Hermitian eigenvalue problems. The reason for block methods is twofold: First, as an increased robustness measure, since Davidson methods have no problems identifying multiple eigenvalues =-=[47]-=-. Second, to take advantage of the increased cache locality in block matrix-vector, preconditioning, and BLAS operations. Although the total number of matrix-vector multiplications increases with larg y cannot converge to full accuracy because of locking, they are practically converged (see [65]), an algorithm that repeats steps (4) to (20) until convergence is verified for all required pairs (see =-=[47]-=-), the handling of multiple user-defined shifts, and many others. Section 3.3 outlines some of them. Algorithms 3.2–3.4 describe the PRIMME implementation of three important components of GD+k: conveg ors are kept in the search space and improve with time. Locking usually provides a better mechanism than non-locking for identifying eigenvalues that are highly clustered or of very high multiplicity =-=[47]-=-. However, locking introduces a subtle numerical, but not floating point, problem. Specifically, a large number of locked, approximate eigenvectors, that have converged to tol residual accuracy, may i tioners of A, or other problem specific preconditioners can be readily used. 4. The current distribution of PRIMME implements only a subset of the Iterative Validation of Eigensolvers (IVE) algorithm =-=[47]-=-. We have a fully functional IVE working with an older version of PRIMME that will be ported to the current distribution. This will also be coordinated with the final Rayleigh-Ritz procedure over all   </text>
<query_num> 21305 </query_num>
<text>   ore the iteration vectors for computing eigenvector approximations. With slow convergence, the storage demands of these applications can be staggering. Recently, iterative methods have been developed =-=[62, 69, 41, 61, 52, 66, 68]-=-, that can use effectively the large arsenal of preconditioners for linear systems, and converge near optimally to an eigenpair under limited memory requirements. 2.1. In search of (near) optimal meth irs, it is an open question whether optimality can be achieved under limited memory. If one eigenvalue is known exactly, the corresponding eigenvector can be obtained optimally through a CG iteration =-=[41, 66]-=-. If numEvals eigenvalues are known, one may think that the analogue optimality is to run numEvals separate CG iterations. This is the approach taken by most limited memory, preconditioned eigensolver problems” [7]. Since then, the consensus on the relative merits of methods may not have changed for symmetric linear systems, but the area of symmetric eigenproblems has seen some remarkable progress =-=[66, 68, 53, 1, 52, 26, 46, 30, 41, 6, 74]-=-. This recent progress is reflected in the large number of codes for symmetric eigenproblems. In their most recent survey of eigenvalue codes, [33], Hernandez et al. list 20 eigenvalue codes that have e their own matrix-multivector and preconditioner-multivector operations. ANASAZI is still under development, but currently it does not include the near optimal GD+k or JDQMR methods. As we showed in =-=[66, 68, 69]-=- the difference in convergence over LOBPCG and (Generalized) Davidson can be substantial, if the preconditioner is not powerful enough to result in convergence in only a few iterations. Also, despite  y minimizing x T Ax/x T x [19]. For many eigenpairs, the same formulation applies for minimizing the trace of a block of vectors [59], working with numEvals-dimensional spaces [2]. As we discussed in =-=[66, 68]-=-, most eigenmethods can be interpreted through the inexact Newton viewpoint or the quasi-Newton viewpoint. The following includes excerpts from these two papers. 3.1.1. The inexact Newton approach. Th   </text>
<query_num> 21306 </query_num>
<text>   ore the iteration vectors for computing eigenvector approximations. With slow convergence, the storage demands of these applications can be staggering. Recently, iterative methods have been developed =-=[62, 69, 41, 61, 52, 66, 68]-=-, that can use effectively the large arsenal of preconditioners for linear systems, and converge near optimally to an eigenpair under limited memory requirements. 2.1. In search of (near) optimal meth problems” [7]. Since then, the consensus on the relative merits of methods may not have changed for symmetric linear systems, but the area of symmetric eigenproblems has seen some remarkable progress =-=[66, 68, 53, 1, 52, 26, 46, 30, 41, 6, 74]-=-. This recent progress is reflected in the large number of codes for symmetric eigenproblems. In their most recent survey of eigenvalue codes, [33], Hernandez et al. list 20 eigenvalue codes that have e their own matrix-multivector and preconditioner-multivector operations. ANASAZI is still under development, but currently it does not include the near optimal GD+k or JDQMR methods. As we showed in =-=[66, 68, 69]-=- the difference in convergence over LOBPCG and (Generalized) Davidson can be substantial, if the preconditioner is not powerful enough to result in convergence in only a few iterations. Also, despite   to implement locking as a wrapper to BLOPEX. Finally, a robustness issue may arise in this particular implementation of LOBPCG, because it does not maintain an orthonormal basis for the search space =-=[35, 68]-=-. JDBSYM is a stand-alone software written in C that implements a block version of the Jacobi-Davidson (JD) method for solving standard and generalized real symmetric eigenproblems. Hermitian eigenpro y minimizing x T Ax/x T x [19]. For many eigenpairs, the same formulation applies for minimizing the trace of a block of vectors [59], working with numEvals-dimensional spaces [2]. As we discussed in =-=[66, 68]-=-, most eigenmethods can be interpreted through the inexact Newton viewpoint or the quasi-Newton viewpoint. The following includes excerpts from these two papers. 3.1.1. The inexact Newton approach. Th   </text>
<query_num> 21307 </query_num>
<text>   problems” [7]. Since then, the consensus on the relative merits of methods may not have changed for symmetric linear systems, but the area of symmetric eigenproblems has seen some remarkable progress =-=[66, 68, 53, 1, 52, 26, 46, 30, 41, 6, 74]-=-. This recent progress is reflected in the large number of codes for symmetric eigenproblems. In their most recent survey of eigenvalue codes, [33], Hernandez et al. list 20 eigenvalue codes that have   </text>
</top>
<top>
<paper_num> 214 </paper_num>
<paper_title>   Minimizing interference in ad hoc and sensor networks.  </paper_title>
<paper_abstract>   Reducing interference is one of the main challenges in wireless communication, and particularly in ad hoc networks. The amount of interference experienced by a node v corresponds to the number of other nodes whose transmission range covers v. At the cost of communication links being dropped, interference can be reduced by decreasing the node’s transmission power. In this paper, we study the problem of minimizing the average interference while still maintaining desired network properties, such as connectivity, point-to-point connections, or multicast trees. In particular, we present a greedy algorithm that computes an O(log n) approximation to the interference problem with connectivity requirement, where n is the number of nodes in the network. We then show the algorithm to be asymptotically optimal by proving a corresponding Ω(log n) lower bound that holds even in a more restricted interference model. Finally, we show how the algorithm can be generalized towards solving the interference problem for network properties that can be formulated as a 0-1 proper function.  </paper_abstract>
<query_num> 21401 </query_num>
<text>   -1 proper functions [8]. In fact, the upper bounds of some of the interference problems studied in this paper have been obtained earlier in an entirely different context. Particularly, the authors of =-=[3, 4]-=- study the problem of reducing the total amount of energy spent by the nodes in wireless ad hoc networks, while keep-ing connectivity properties. By transforming the network instances of [3, 4] prope otically optimal solutions for the interference problem. In this paper, however, we formulate and analyze the algorithms directly in the interference model which yields a proof that is different from =-=[3, 4]-=-. Additionally, our result on 0-1 proper functions is strictly more general than the connectivity requirements studied in [3, 4] and hence, is a true generalization of their result. The remainder of t ssume points s ∈ S to be located in a metric space (S, µ). We slightly change the problem definition in the way that the power assignment r maps every active node to a real value, r : A → R + . As in =-=[3, 4]-=-, this real value represents a node’s transmission range. In other words, a node a ∈ A interferes with all nodes b ∈ B whose distance to a is at most r(a). It is clear that at the cost of a less intui a)|. 3. APPROXIMATION ALGORITHM In this section, we present a greedy algorithm that achieves an O(log n) approximation ratio, the same guarantee that can be achieved by the greedy algorithms given in =-=[3, 4]-=-. Our solution differs in the sense that it is directly formulated and analyzed in the interference model. Many of the algorithmic challenges faced in designing approximation algorithms for interferen   </text>
<query_num> 21402 </query_num>
<text>   7, 14, 12, 13, 9]. Most protocols adopt structures from the field of computational geometry, focusing on preserving energy efficient paths or computing planar subgraphs in order to facilitate routing =-=[1]-=-. Yet, in spite of its practical importance, there has been virtually no algorithmic work dealing with interference reduction explicitly. Instead, it has been assumed that the above mentioned traditio   </text>
<query_num> 21403 </query_num>
<text>   d that holds even in the restricted ad hoc metric case. Finally, we also show how various other network properties can be solved. These include properties that can be modelled as 0-1 proper functions =-=[8]-=-. In fact, the upper bounds of some of the interference problems studied in this paper have been obtained earlier in an entirely different context. Particularly, the authors of [3, 4] study the proble lems. In particular, we consider problems that can be formulated as special cut-covering problems. A function f : 2 A → {0, 1} is called a 0-1 proper function if it satisfies the following properties =-=[8]-=-: • f(A) = 0 • f(C) = f(A \ C) for all C ⊆ A • if C and C ′ are disjoint, then f(C) = f(C ′ ) = 0 ⇒ f(C ∪ C ′ ) = 0 As explained in [8], the class of network design problems which can be formulated by   </text>
<query_num> 21404 </query_num>
<text>   induced interference is minimized. In the ad hoc network community, much effort has been invested in finding network structures having desirable properties such as sparseness or low node-degree, e.g. =-=[17, 14, 12, 13, 9]-=-. Most protocols adopt structures from the field of computational geometry, focusing on preserving energy efficient paths or computing planar subgraphs in order to facilitate routing [1]. Yet, in spit   </text>
<query_num> 21405 </query_num>
<text>   oned traditional network structures (which guarantee sparseness and low node-degree) also help in confining interference. By disproving this long-standing conjecture, a recent paper by Burkhart et al =-=[2]-=- has changed the focus in the wireless networking community towards an explicit study of interference in ad hoc radio networks [7, 16, 18]. The work of [2] is of practical nature and verifies the effi   </text>
<query_num> 21406 </query_num>
<text>   oving this long-standing conjecture, a recent paper by Burkhart et al [2] has changed the focus in the wireless networking community towards an explicit study of interference in ad hoc radio networks =-=[7, 16, 18]-=-. The work of [2] is of practical nature and verifies the efficiency of the proposed solution by means of simulation. A recent algorithmic work is [18] in which the problem of minimizing the maximum i  and challenging from a theoretical point of view is their somewhat counter-intuitive behavior. Consider for instance the simple network shown in Figure 1 consisting of nodes a1, . . . , an on a line =-=[18]-=-. 1 2 4 8 16 01 01 01 01 01 01 000000000000000000000000000000000000 111111111111111111111111111111111111 0000000000 1111111111 01 01 a a 1 2 01 a 3 01 a 4 Figure 1: Example with nodes ai being located   </text>
<query_num> 21407 </query_num>
<text>   oving this long-standing conjecture, a recent paper by Burkhart et al [2] has changed the focus in the wireless networking community towards an explicit study of interference in ad hoc radio networks =-=[7, 16, 18]-=-. The work of [2] is of practical nature and verifies the efficiency of the proposed solution by means of simulation. A recent algorithmic work is [18] in which the problem of minimizing the maximum i r the special case when the nodes are located on the Euclidean line, [18] presents an O( 4√ n) approximation algorithm. For a different, less robust link-based definition of interference, the work of =-=[16]-=- presents optimal or approximate solutions of interference reduction. As for the node-based definition studied in this paper, [16] proposes a heuristic, but without providing analytic approximation gu   </text>
<query_num> 21408 </query_num>
<text>   ˜ B in iteration i. This concludes the proof of Lemma 3.2. Given Lemma 3.2, it is now straightforward to derive the main theorem using a proof technique used for instance in [11] and subsequently in =-=[10]-=-. Theorem 3.3. Let rALG and rOP T be the power assignments obtained by Algorithm 1 and by an optimal algorithm, respectively. Further, χ(rALG) and χ(rOP T ) denote their respective average interferenc   </text>
</top>
<top>
<paper_num> 215 </paper_num>
<paper_title>   On Modeling Round-Trip Time Dynamics of the Internet Using System Identification.  </paper_title>
<paper_abstract>   Understanding the end-to-end packet delay dynamics of the  Internet is of crucial importance since it directly affects the QoS (Quality  of Services) of realtime services, and it enables us to design an efficient  congestion control mechanism. In this paper, we measure the round-trip  time, and build a mathematical model representing its dynamics using  system identification. We first measure, as the input and output data for  system identification, the packet inter-departure time from a source host  and the corresponding round-trip time measured by the source host.  </paper_abstract>
<query_num> 21501 </query_num>
<text>   11], the authors have proposed a traffic model for wide-area TCP traffic by characterizing several distributions of, for example, the packet inter-arrival time and the number of bytes transferred. In =-=[12]-=-, the authors have proposed a fast algorithm to construct a CMRP (Circulant Modulated Rate Process) for traffic modeling. In [13], CMRP and ARMA (Auto-Regressive Moving Average) have been discussed as   </text>
<query_num> 21502 </query_num>
<text>   e has been set to 0.2 ms, resulting in the average packet transmission rate of 43.2 Mbps and the average round-trip time of 0.8 ms. Shown in Fig. 4 are results in the network configuration N1 for i = =-=[8; 8; 1]-=-. This figure shows: (a) the packet inter-departure time u(k), (b) the measured round-trip time, (c) the measured round-trip time variation y(k), and (d) comparison between the measured output data an  in the network configuration N1, the round-trip time is almost independent of the packet inter-departure time. Network Configuration N2 Figure 5 shows results in the network configuration N2 for i = =-=[8; 8; 1]-=-. In this case, the mean packet inter-departure time has been set to 0.6 ms, resulting the average packet transmission rate of 18.0 Mbps and the average round-trip time of 1.8 ms. Figure 5(c) shows th ed round-trip time variation is disturbed by other traffic, which is unknown so that not included in the model output y 3 (k). Network Configuration N3 Results in the network configuration N3 for i = =-=[8; 8; 1]-=- are shown in Fig. 6. In this case, the mean packet inter-departure time has been set to 12.0 ms, resulting the average packet transmission rate of 967 Kbps and the average round-trip time of 16.7 ms.  can be modeled by the ARX model when the network is moderately congested. Choice of Model Orders and Number of Samples In the above results, the orders and the delay of the ARX model is fixed at i = =-=[8; 8; 1]-=-. In general, the accuracy of the ARX model is dependent on the On Modeling Round-Trip Time Dynamics of the Internet 11 40 50 60 70 80 90 100 0.04 0.045 0.05 0.055 0.06 0.065 0.07 0.075 number of samp tween the loss function JN (`) and the number of samples used from the input and output data in the network configuration N2. In this figure, the orders and the delay of the ARX model is fixed at i = =-=[8; 8; 1]-=-, while the number of samples is changed from 40 to 100. This figure shows a tendency that, as the number of samples increases, the loss function first decreases and then gradually increases. The simi   </text>
<query_num> 21503 </query_num>
<text>   is paper, has not been investigated. Aside from analyses of the end-to-end packet delay, another area of measurementbased studies is regarding a black-box modeling of the network traffic [11--15]. In =-=[11]-=-, the authors have proposed a traffic model for wide-area TCP traffic by characterizing several distributions of, for example, the packet inter-arrival time and the number of bytes transferred. In [12   </text>
<query_num> 21504 </query_num>
<text>   n. 1 Introduction Understanding the end-to-end packet delay dynamics of the Internet is of crucial importance since (1) it directly affects the QoS (Quality of Services) of realtime applications, and =-=(2)-=- it enables us to design an efficient congestion control mechanism for both realtime and non-realtime applications. For non-realtime applications, a delay-based approach for congestion control mechani ) j u(k 0 1). The numbers n a and n b are the orders of polynomials. The number n d corresponds to delays from the input to the output. For compact notation, i is introduced as i = [n a ; n b ; n d ] =-=(2)-=- In our case, u(k) and y(k) correspond to k-th packet inter-departure time and k-th round-trip time variation. All coefficients of the polynomials, an and b n , are parameters of the ARX model, and ar   </text>
<query_num> 21505 </query_num>
<text>   nclusion of this paper. 2 Related Works In the literature, there have been several measurement-based studies regarding the end-to-end packet delay [3, 4, 8, 9] and the end-to-end path characteristics =-=[5, 10]-=-. In [3], the authors have examined the end-to-end packet delay and loss behavior in the Internet using small UDP probe packets. In [4], the authors have examined the correlation between packet delay  eristics of a transmission link based on end-to-end multicast measurements. In [5], the packet dynamics of the Internet have been analyzed based on measurements of about 20,000 TCP data transfers. In =-=[10]-=-, the routing behavior of the Internet has been analyzed based on measurements of about 40,000 end-to-end traceroute results. However, those studies are limited to a statistical behavior of the endto   </text>
<query_num> 21506 </query_num>
<text>   nclusion of this paper. 2 Related Works In the literature, there have been several measurement-based studies regarding the end-to-end packet delay [3, 4, 8, 9] and the end-to-end path characteristics =-=[5, 10]-=-. In [3], the authors have examined the end-to-end packet delay and loss behavior in the Internet using small UDP probe packets. In [4], the authors have examined the correlation between packet delay  odeling round-trip time dynamics as SISO system authors have presented an approach to characterize loss and delay characteristics of a transmission link based on end-to-end multicast measurements. In =-=[5]-=-, the packet dynamics of the Internet have been analyzed based on measurements of about 20,000 TCP data transfers. In [10], the routing behavior of the Internet has been analyzed based on measurements   </text>
<query_num> 21507 </query_num>
<text>   ns. For non-realtime applications, a delay-based approach for congestion control mechanisms, rather than a loss-based approach as used in TCP (Transmission Control Protocol), has been proposed (e.g., =-=[1, 2]-=-). The main advantage of such a delay-based approach is, if it is properly designed, packet losses can be prevented by anticipating impending congestion from increasing packet delays. For a long time, e has been set to 0.2 ms, resulting in the average packet transmission rate of 43.2 Mbps and the average round-trip time of 0.8 ms. Shown in Fig. 4 are results in the network configuration N1 for i = =-=[8; 8; 1]-=-. This figure shows: (a) the packet inter-departure time u(k), (b) the measured round-trip time, (c) the measured round-trip time variation y(k), and (d) comparison between the measured output data an  in the network configuration N1, the round-trip time is almost independent of the packet inter-departure time. Network Configuration N2 Figure 5 shows results in the network configuration N2 for i = =-=[8; 8; 1]-=-. In this case, the mean packet inter-departure time has been set to 0.6 ms, resulting the average packet transmission rate of 18.0 Mbps and the average round-trip time of 1.8 ms. Figure 5(c) shows th ed round-trip time variation is disturbed by other traffic, which is unknown so that not included in the model output y 3 (k). Network Configuration N3 Results in the network configuration N3 for i = =-=[8; 8; 1]-=- are shown in Fig. 6. In this case, the mean packet inter-departure time has been set to 12.0 ms, resulting the average packet transmission rate of 967 Kbps and the average round-trip time of 16.7 ms.  can be modeled by the ARX model when the network is moderately congested. Choice of Model Orders and Number of Samples In the above results, the orders and the delay of the ARX model is fixed at i = =-=[8; 8; 1]-=-. In general, the accuracy of the ARX model is dependent on the On Modeling Round-Trip Time Dynamics of the Internet 11 40 50 60 70 80 90 100 0.04 0.045 0.05 0.055 0.06 0.065 0.07 0.075 number of samp   </text>
<query_num> 21508 </query_num>
<text>   other black-box approaches, which model network traffic using the On Modeling Round-Trip Time Dynamics of the Internet 5 AR (Auto-Regressive) model or the ARMA (Auto-Regressive Moving Average) model =-=[13, 17, 18]-=-. Figure 3 illustrates a typical usage of the AR model or the ARMA model for modeling network traffic. Comparing Figs. 2 and 3, the ARX model has the input whereas either the AR model or the ARMA mode   </text>
<query_num> 21509 </query_num>
<text>   our modeling approach for a through set of input and output data obtained from various network configurations. We are currentry measuring the input and output data in working LAN and WAN environments =-=[23]-=-. Acknowledgement This work was supported in part by Research for the Future Program of Japan Society for the Promotion of Science under the Project &amp;quot;Integrated Network Architecture for Advanced Multi   </text>
<query_num> 21510 </query_num>
<text>   possible applications of our approach, followed by conclusion of this paper. 2 Related Works In the literature, there have been several measurement-based studies regarding the end-to-end packet delay =-=[3, 4, 8, 9]-=- and the end-to-end path characteristics [5, 10]. In [3], the authors have examined the end-to-end packet delay and loss behavior in the Internet using small UDP probe packets. In [4], the authors hav   </text>
</top>
<top>
<paper_num> 216 </paper_num>
<paper_title>   Eclat: Automatic Generation and Classification of Test Inputs.  </paper_title>
<paper_abstract>   Abstract. This paper describes a technique that selects, from a large set of test inputs, a small subset likely to reveal faults in the software under test. The technique takes a program or software component, plus a set of correct executions— say, from observations of the software running properly, or from an existing test suite that a user wishes to enhance. The technique first infers an operational model of the software’s operation. Then, inputs whose operational pattern of execution differs from the model in specific ways are suggestive of faults. These inputs are further reduced by selecting only one input per operational pattern. The result is a small portion of the original inputs, deemed by the technique as most likely to reveal faults. Thus, the technique can also be seen as an error-detection technique. The paper describes two additional techniques that complement test input selection. One is a technique for automatically producing an oracle (a set of assertions) for a test input from the operational model, thus transforming the test input into a test case. The other is a classification-guided test input generation technique that also makes use of operational models and patterns. When generating inputs, it filters out code sequences that are unlikely to contribute to legal inputs, improving the efficiency of its search for fault-revealing inputs. We have implemented these techniques in the Eclat tool, which generates unit tests for Java classes. Eclat’s input is a set of classes to test and an example program execution—say, a passing test suite. Eclat’s output is a set of JUnit test cases, each containing a potentially fault-revealing input and a set of assertions at least one of which fails. In our experiments, Eclat successfully generated inputs that exposed fault-revealing behavior; we have used Eclat to reveal real errors in programs. The inputs it selects as fault-revealing are an order of magnitude as likely to reveal a fault as all generated inputs. 1  </paper_abstract>
<query_num> 21601 </query_num>
<text>   BoundedStack We illustrate the test generation and selection technique by describing the operation of the Eclat tool, when applied to a bounded stack implementation used previously in the literature =-=[22, 30, 9]-=-. The bounded stack implementation (Figure 1) and testing code were written in Java by two students, an “author” and a “tester.” The tester wrote a set of axioms on which the author based the implemen  throws an exception. Eclat classifies the input as fault-revealing. The equals method (Figure 1) incorrectly handles a null argument. This fault went undetected in all previous analyses of the class =-=[22, 30, 9]-=-.s508 C. Pacheco and M.D. Ernst public void test_3_pop() throws Exception { } ubs.BoundedStack var8 = new ubs.BoundedStack(); // Check preconditions. checkPreconditions_pop(var8); checkObjectInvariant t-revealing. Our techniques are intended for use when formal specifications are not available, as was the case for most of the programs. Comparison with other tools. JCrasher [9], Jtest [19], and Jov =-=[30]-=- have the same goals as Eclat: to generate random candidate inputs and select potentially fault-revealing ones. We report results from running JCrasher. We tried the other tools, but Jov and Jtest wer uture work, we plan to investigate exhaustive generation combined with techniques for avoiding generation of duplicate inputs [28, 29]. 6 Related Work The most closely related work to ours is the Jov =-=[30]-=- and JCrasher [9] tools, which share the goal of selecting, from a randomly-generated set of candidate inputs, a set most likely to be useful. This reduces the number of test inputs a human must exami g many illegal ones), not just inputs similar to the ones in the original test suite. However, the user gets no help in recognizing such illegal inputs. In fact, the majority of errors that Jov finds =-=[30]-=- are illegal inputs and precondition violations, not true errors [27]. Our work extends that of Xie and Notkin in several ways. Our technique explicitly addresses the imperfect nature of a derived ope   </text>
<query_num> 21602 </query_num>
<text>   BoundedStack We illustrate the test generation and selection technique by describing the operation of the Eclat tool, when applied to a bounded stack implementation used previously in the literature =-=[22, 30, 9]-=-. The bounded stack implementation (Figure 1) and testing code were written in Java by two students, an “author” and a “tester.” The tester wrote a set of axioms on which the author based the implemen s.getNumberOfElements() != numElems) return false; int[] sElems = s.getArray(); for (int j=0; j&amp;lt;numElems; j++) { if (elems[j] != sElems[j]) return false; } return true; } } Fig. 1. Class BoundedStack =-=[22]-=- (abbreviated). Methods pop and equals contain errorssEclat: Automatic Generation and Classification of Test Inputs 507 Eclat Report Input 1 BoundedStack var8 = new BoundedStack(); var8.push(2); int v  throws an exception. Eclat classifies the input as fault-revealing. The equals method (Figure 1) incorrectly handles a null argument. This fault went undetected in all previous analyses of the class =-=[22, 30, 9]-=-.s508 C. Pacheco and M.D. Ernst public void test_3_pop() throws Exception { } ubs.BoundedStack var8 = new ubs.BoundedStack(); // Check preconditions. checkPreconditions_pop(var8); checkObjectInvariant   </text>
<query_num> 21603 </query_num>
<text>   domly-generated set of candidate inputs, a set most likely to be useful. This reduces the number of test inputs a human must examine. Our research was inspired by Jov [30]. Jov builds on earlier work =-=[15]-=- that identified a test as a potentially valuable addition to a test suite if the test violates an operational abstraction built from the suite: the test represents some combination of values that dif s the operational abstraction not just to select tests, but also to guide test generation, by iterated use of the Jtest tool [19]. Jov also differs from the previous, automated work on test selection =-=[15]-=- by placing it in a loop with human interaction and iterating as many times as desired: 1. Create an operational model (invariants) from a test suite. 2. Generate test inputs that violate the invarian   </text>
<query_num> 21604 </query_num>
<text>   n step clusters test inputs in order to reduce their number, and JCrasher has a similar step. Several researchers have used machine learning to classify program executions as either correct or faulty =-=[20, 5, 3]-=-. It would be interesting to apply such techniques in order to further improve Eclat.s7 Conclusion Eclat: Automatic Generation and Classification of Test Inputs 525 We have presented an input selectio   </text>
<query_num> 21605 </query_num>
<text>   ore fault-revealing inputs than exhaustive generation (bottom plots). In future work, we plan to investigate exhaustive generation combined with techniques for avoiding generation of duplicate inputs =-=[28, 29]-=-. 6 Related Work The most closely related work to ours is the Jov [30] and JCrasher [9] tools, which share the goal of selecting, from a randomly-generated set of candidate inputs, a set most likely t   </text>
<query_num> 21606 </query_num>
<text>   plementation uses operational abstractions generated by the Daikon invariant detector [11]. There are other techniques for generating models of program behavior based on an example use of the program =-=[14, 26, 1, 16]-=-. The models that these techniques generate vary in the kinds of properties they express, from legal sequences of method calls [26] to algebraic specifications of method behavior [16]. Figure 5 shows   </text>
<query_num> 21607 </query_num>
<text>   plementation uses operational abstractions generated by the Daikon invariant detector [11]. There are other techniques for generating models of program behavior based on an example use of the program =-=[14, 26, 1, 16]-=-. The models that these techniques generate vary in the kinds of properties they express, from legal sequences of method calls [26] to algebraic specifications of method behavior [16]. Figure 5 shows   a test suite if the test violates an operational abstraction built from the suite: the test represents some combination of values that differs from all tests currently in the suite. (The DIDUCE tool =-=[14]-=- takes a similar approach, though with the goal of identifying bugs at run time rather than improving test suites: a property that has held for part of a run, but is later violated, is suggestive of a   </text>
<query_num> 21608 </query_num>
<text>   r experiments that use the same programs with different test suites because JCrasher does not make use of the test suite. We also executed all the inputs against the formal specifications (using jmlc =-=[6]-=-). Wes518 C. Pacheco and M.D. Ernst Generated inputs Selected inputs JCrasher inputs inputs reveal preci- inputs reveal preci- inputs reveal preciProgram generated faults sion selected faults sion sel   </text>
<query_num> 21609 </query_num>
<text>   ut to the program or module, and an oracle, a procedure that determines whether the program behaves as expected on the input. Many techniques can automatically generate candidate inputs for a program =-=[10, 18, 17, 23, 8, 4, 19, 9, 12]-=-, but constructing an oracle for each input remains a largely manual task (unless a formal specification of A.P. Black (Ed.): ECOOP 2005, LNCS 3586, pp. 504–527, 2005. c○ Springer-Verlag Berlin Heidel   </text>
<query_num> 21610 </query_num>
<text>   ut to the program or module, and an oracle, a procedure that determines whether the program behaves as expected on the input. Many techniques can automatically generate candidate inputs for a program =-=[10, 18, 17, 23, 8, 4, 19, 9, 12]-=-, but constructing an oracle for each input remains a largely manual task (unless a formal specification of A.P. Black (Ed.): ECOOP 2005, LNCS 3586, pp. 504–527, 2005. c○ Springer-Verlag Berlin Heidel  BoundedStack We illustrate the test generation and selection technique by describing the operation of the Eclat tool, when applied to a bounded stack implementation used previously in the literature =-=[22, 30, 9]-=-. The bounded stack implementation (Figure 1) and testing code were written in Java by two students, an “author” and a “tester.” The tester wrote a set of axioms on which the author based the implemen  throws an exception. Eclat classifies the input as fault-revealing. The equals method (Figure 1) incorrectly handles a null argument. This fault went undetected in all previous analyses of the class =-=[22, 30, 9]-=-.s508 C. Pacheco and M.D. Ernst public void test_3_pop() throws Exception { } ubs.BoundedStack var8 = new ubs.BoundedStack(); // Check preconditions. checkPreconditions_pop(var8); checkObjectInvariant  illegal, normal, or fault-revealing. Our techniques are intended for use when formal specifications are not available, as was the case for most of the programs. Comparison with other tools. JCrasher =-=[9]-=-, Jtest [19], and Jov [30] have the same goals as Eclat: to generate random candidate inputs and select potentially fault-revealing ones. We report results from running JCrasher. We tried the other to n to investigate exhaustive generation combined with techniques for avoiding generation of duplicate inputs [28, 29]. 6 Related Work The most closely related work to ours is the Jov [30] and JCrasher =-=[9]-=- tools, which share the goal of selecting, from a randomly-generated set of candidate inputs, a set most likely to be useful. This reduces the number of test inputs a human must examine. Our research   </text>
<query_num> 21611 </query_num>
<text>   ut to the program or module, and an oracle, a procedure that determines whether the program behaves as expected on the input. Many techniques can automatically generate candidate inputs for a program =-=[10, 18, 17, 23, 8, 4, 19, 9, 12]-=-, but constructing an oracle for each input remains a largely manual task (unless a formal specification of A.P. Black (Ed.): ECOOP 2005, LNCS 3586, pp. 504–527, 2005. c○ Springer-Verlag Berlin Heidel hen a specification of valid inputs is available. Therefore, techniques that make it more effective are valuable contributions. Our technique could be combined with any technique for generating tests =-=[8, 4]-=-, in order to filter the tests before being presented to a user. Our technique is attractive because it does not require a human-written formal specification; when one is present, much more powerful t   </text>
</top>
<top>
<paper_num> 217 </paper_num>
<paper_title>   Harmony in Motion.  </paper_title>
<paper_abstract>   Cross-modal analysis offers information beyond that extracted from individual modalities. Consider a camcorder having a single microphone in a cocktail-party: it captures several moving visual objects which emit sounds. A task for audio-visual analysis is to identify the number of independent audio-associated visual objects (AVOs), pinpoint the AVOs ’ spatial locations in the video and isolate each corresponding audio component. Part of these problems were considered by prior studies, which were limited to simple cases, e.g., a single AVO or stationary sounds. We describe an approach that seeks to overcome these challenges. It acknowledges the importance of temporal features that are based on significant changes in each modality. A probabilistic formalism identifies temporal coincidences between these features, yielding cross-modal association and visual localization. This association is of particular benefit in harmonic sounds, as it enables subsequent isolation of each audio source. We demonstrate this in challenging experiments, having multiple, simultaneous highly nonstationary AVOs. 1. Cross-Modal Analysis Cross modal analysis is gaining interest in computer vision. Such analysis seeks associations between sources of input data, which have very different natures. Examples of this include registration of images acquired using sensors of different kinds [16], or association of images to text [13], such as in web pages and multimedia subtitles. It also includes audio-visual analysis [24, 26, 30], which has seen a growing expansion of research directions, including lip-reading [8, 14], tracking [25], and spatial localization [7, 10, 18, 19, 23]. This follows evidence of audiovisual cross-modal processing in biology [12]. This work deals with complex scenarios that are sometimes referred to in the literature as a cocktail party [10, 14, 27]: multiple sources exist simultaneously in all modalities.  </paper_abstract>
<query_num> 21701 </query_num>
<text>   , 18, 19, 23]. This follows evidence of audiovisual cross-modal processing in biology [12]. This work deals with complex scenarios that are sometimes referred to in the literature as a cocktail party =-=[10, 14, 27]-=-: multiple sources exist simultaneously in all modalities. Harmony in Motion Zohar Barzelay and Yoav Y. Schechner Department of Electrical Engineering Technion - Israel Inst. Technology Haifa 32000, I The spectrogram is |F (t, f)| 2 . See for example the spectrograms in Fig. 2. As seen in Fig. 2, the energy of each distinct sound lies in a set Γ of time-frequency bins {(t, f)}. A common assumption =-=[1, 27, 32]-=- is that if there are other sound sources, then the energy distribution in {(t, f)} of these disturbances has only little overlap with the bins in Γ. This assumption is based on the sparsity of typica ones, in spectrograms. Consequently, a sound of interest can be enhanced by maintaining the values of F (t, f) in Γ, while nulling the other bins. This binary masking forms the basis for many methods =-=[1, 27, 32]-=-. The masked F (t, f) is then transformed back [27] to a sound signal ˜s(n). How is the set Γ of a sound characterized? In an harmonic sound, the acoustic energy lies in a pitch frequency f0 and in in ments we compound separately-recorded movies (e.g., a violin sequence and a guitar sequence) into a single video. 7 Such a procedure is a common practice in single-micrhopone audio-separation studies =-=[1, 14, 27]-=-, since it provides access to the audio ground-truth data. This allows quantitative assessment of the quality of audio isolation, as we describe below. The cross-modal method has several parameters, s   </text>
<query_num> 21702 </query_num>
<text>   . Specifically, the relative temporal 5 We do not know the number of sound sources in the scene: in addition to the visual AVOs there can be audio sources of objects out of view. Hence, we cannot use =-=[1, 21, 31]-=-. derivative A(t, f) − A(t − 1,f) D(t, f) = (12) A(t − 1,f) emphasizes an increase of amplitude in frequency bins that have been quiet (no sound) just before t. As a practical criterion, however, Eq.   </text>
<query_num> 21703 </query_num>
<text>   The spectrogram is |F (t, f)| 2 . See for example the spectrograms in Fig. 2. As seen in Fig. 2, the energy of each distinct sound lies in a set Γ of time-frequency bins {(t, f)}. A common assumption =-=[1, 27, 32]-=- is that if there are other sound sources, then the energy distribution in {(t, f)} of these disturbances has only little overlap with the bins in Γ. This assumption is based on the sparsity of typica ones, in spectrograms. Consequently, a sound of interest can be enhanced by maintaining the values of F (t, f) in Γ, while nulling the other bins. This binary masking forms the basis for many methods =-=[1, 27, 32]-=-. The masked F (t, f) is then transformed back [27] to a sound signal ˜s(n). How is the set Γ of a sound characterized? In an harmonic sound, the acoustic energy lies in a pitch frequency f0 and in in . Specifically, the relative temporal 5 We do not know the number of sound sources in the scene: in addition to the visual AVOs there can be audio sources of objects out of view. Hence, we cannot use =-=[1, 21, 31]-=-. derivative A(t, f) − A(t − 1,f) D(t, f) = (12) A(t − 1,f) emphasizes an increase of amplitude in frequency bins that have been quiet (no sound) just before t. As a practical criterion, however, Eq.  ments we compound separately-recorded movies (e.g., a violin sequence and a guitar sequence) into a single video. 7 Such a procedure is a common practice in single-micrhopone audio-separation studies =-=[1, 14, 27]-=-, since it provides access to the audio ground-truth data. This allows quantitative assessment of the quality of audio isolation, as we describe below. The cross-modal method has several parameters, s iciently wide baseline. Other methods, which use a single microphone, generally separate audio based on training on specific classes of sources, particularly speech and typical potential disturbances =-=[1]-=-. Such methods may succeed in enhancing continuous sounds, but may fail to group discontinuous sounds correctly to a single stream. This is the case when the audio-characteristics of the different sou   </text>
<query_num> 21704 </query_num>
<text>   The spectrogram is |F (t, f)| 2 . See for example the spectrograms in Fig. 2. As seen in Fig. 2, the energy of each distinct sound lies in a set Γ of time-frequency bins {(t, f)}. A common assumption =-=[1, 27, 32]-=- is that if there are other sound sources, then the energy distribution in {(t, f)} of these disturbances has only little overlap with the bins in Γ. This assumption is based on the sparsity of typica ones, in spectrograms. Consequently, a sound of interest can be enhanced by maintaining the values of F (t, f) in Γ, while nulling the other bins. This binary masking forms the basis for many methods =-=[1, 27, 32]-=-. The masked F (t, f) is then transformed back [27] to a sound signal ˜s(n). How is the set Γ of a sound characterized? In an harmonic sound, the acoustic energy lies in a pitch frequency f0 and in in d the separation results are available through the above-mentioned link. Quantitative Recovery Criteria We quantify the quality of the audio-isolation in the experiments by criteria described in Ref. =-=[32]-=-. These measures utilize our access to the ground-truth audio data. The first measure evaluates the improvement of the signal-tointerference-ratio (SIR). The second measure calculates the preserved-si sual data for audio isolation. This raises the question of how audio-only (unrelated to vision) methods can benefit from such a framework. Some audio-separation methods are based on microphone arrays =-=[32]-=- having a sufficiently wide baseline. Other methods, which use a single microphone, generally separate audio based on training on specific classes of sources, particularly speech and typical potential   </text>
<query_num> 21705 </query_num>
<text>   ages to text [13], such as in web pages and multimedia subtitles. It also includes audio-visual analysis [24, 26, 30], which has seen a growing expansion of research directions, including lip-reading =-=[8, 14]-=-, tracking [25], and spatial localization [7, 10, 18, 19, 23]. This follows evidence of audiovisual cross-modal processing in biology [12]. This work deals with complex scenarios that are sometimes re   </text>
<query_num> 21706 </query_num>
<text>   er vision. Such analysis seeks associations between sources of input data, which have very different natures. Examples of this include registration of images acquired using sensors of different kinds =-=[16]-=-, or association of images to text [13], such as in web pages and multimedia subtitles. It also includes audio-visual analysis [24, 26, 30], which has seen a growing expansion of research directions,  ng problem: that of images. Feature-based image registration focuses on sharp spatial changes (edges and corners) [6], rather than the smooth regions between them. In crosssensor image matching, Ref. =-=[16]-=- highlighted sharp spatial 1 Some studies used an approach motivated by computer-vision in order to make only-audio analysis [17, 28]. changes by high-pass filtering. Analogously, in our audiovisual m   </text>
<query_num> 21707 </query_num>
<text>   ions between sources of input data, which have very different natures. Examples of this include registration of images acquired using sensors of different kinds [16], or association of images to text =-=[13]-=-, such as in web pages and multimedia subtitles. It also includes audio-visual analysis [24, 26, 30], which has seen a growing expansion of research directions, including lip-reading [8, 14], tracking   </text>
<query_num> 21708 </query_num>
<text>   is based on a probabilistic argument and enables imperfect matching. It favors coincidences, and penalizes for mismatches. This criterion is then used in a fast iterative algorithm, in the spirit of =-=[22]-=-. 3.1. Matching Algorithm We now describe both the matching criterion, and the iterative algorithm. Define 1 as a column vector, all of whose elements equal 1. The criterion we use is �L(i) = 2[(a on   </text>
<query_num> 21709 </query_num>
<text>   media subtitles. It also includes audio-visual analysis [24, 26, 30], which has seen a growing expansion of research directions, including lip-reading [8, 14], tracking [25], and spatial localization =-=[7, 10, 18, 19, 23]-=-. This follows evidence of audiovisual cross-modal processing in biology [12]. This work deals with complex scenarios that are sometimes referred to in the literature as a cocktail party [10, 14, 27]: but it cannot, on its own, provide audio spatial localization. Hence, locating audio sources using a camera and a single microphone poses a significant computational challenge. In this context, Refs. =-=[18, 23]-=- spatially localize a single audio-associated visual object (AVO). Ref. [7] localizes multiple AVOs if their sounds are repetitive and non-simultaneous. Neither of these studies attempted audio separa  prior methods use continuous valued variables to represent each modality, e.g., a weighted sum of pixel values. Maximal canonical correlation or mutual information was sought between these variables =-=[10, 15, 18]-=-. That approach is analogous to intensity-based image matching. It implicitly assumes some correlation (possibly nonlinear) between the raw data values in each modality. We do not look at the raw data   </text>
<query_num> 21710 </query_num>
<text>   media subtitles. It also includes audio-visual analysis [24, 26, 30], which has seen a growing expansion of research directions, including lip-reading [8, 14], tracking [25], and spatial localization =-=[7, 10, 18, 19, 23]-=-. This follows evidence of audiovisual cross-modal processing in biology [12]. This work deals with complex scenarios that are sometimes referred to in the literature as a cocktail party [10, 14, 27]: l object (AVO). Ref. [7] localizes multiple AVOs if their sounds are repetitive and non-simultaneous. Neither of these studies attempted audio separation. A pioneering exploration of audio separation =-=[10]-=- used complex optimization of mutual information based on Parzen windows. It can automatically localize an AVO if no other sound is present. Results demonstrated in Ref. [30] were mainly of repetitive  prior methods use continuous valued variables to represent each modality, e.g., a weighted sum of pixel values. Maximal canonical correlation or mutual information was sought between these variables =-=[10, 15, 18]-=-. That approach is analogous to intensity-based image matching. It implicitly assumes some correlation (possibly nonlinear) between the raw data values in each modality. We do not look at the raw data   </text>
<query_num> 21711 </query_num>
<text>   n the smooth regions between them. In crosssensor image matching, Ref. [16] highlighted sharp spatial 1 Some studies used an approach motivated by computer-vision in order to make only-audio analysis =-=[17, 28]-=-. changes by high-pass filtering. Analogously, in our audiovisual matching problem, we use features having strong temporal variations in each of the modalities. As a pre-processing step, image feature is indeed much less sensitive to drift, and is responsive to true onsets (Fig 3). The map � D+(t, f) = max{0, � D(t, f)} 6Treating the spectrogram as a two-dimensional signal (image) was suggested in =-=[17]-=-.smaintains the onset response, while ignoring amplitude decrease caused by fade-outs. We may now use � D+(t on ,f) as input to the algorithm of Ref. [9]. This yields the pitch f0 at t on . Following   </text>
<query_num> 21712 </query_num>
<text>   tion based on salient features [11]. Which features are good? Recall a familiar matching problem: that of images. Feature-based image registration focuses on sharp spatial changes (edges and corners) =-=[6]-=-, rather than the smooth regions between them. In crosssensor image matching, Ref. [16] highlighted sharp spatial 1 Some studies used an approach motivated by computer-vision in order to make only-aud   </text>
</top>
<top>
<paper_num> 218 </paper_num>
<paper_title>   Quantifying Isolation Anomalies.  </paper_title>
<paper_abstract>   Choosing a weak isolation level such as Read Committed is understood as a trade-off, where less isolation means that higher performance is gained but there is an increased possibility that data integrity will be lost. Previously, one side of this trade-off has been carefully studied quantitatively – there are well-known metrics for performance such as transactions per minute, standardized benchmarks that measure these in a controlled way, and analytic models that can predict how performance is influenced by system parameters like multiprogramming level. This paper contributes to quantifying the other aspect of the trade-off. We define a novel microbenchmark that measures how rapidly integrity violations are produced at different isolation levels, for a simple set of transactions. We explore how this rate is impacted by configuration factors such as multiprogramming level, or contention frequency. For the isolation levels in multi-version platforms (Snapshot Isolation and the multiversion variant of Read Committed), we offer a simple probabilistic model that predicts the rate of integrity violations in our microbenchmark from configuration parameters. We validate the predictive model against measurements from the microbenchmark. The model identifies a region of the configuration space where a surprising inversion occurs: for these parameter settings, more integrity violations happen with Snapshot Isolation than with multi-version Read Committed, even though the latter is considered a lower isolation level. 1.  </paper_abstract>
<query_num> 21801 </query_num>
<text>   . More recently, several prominent platforms including Oracle and PostgreSQL have used multiversion concurrency control algorithms without any locks for reading. The Snapshot Isolation (SI) mechanism =-=[3]-=- makes each read see the version of the data that had committed most recently before the transaction started (that is, a read skips over versions that committed after the reading transaction started);  allowing the application developer to take advantage of domain knowledge which means that their particular code will work correctly with low isolation; however, there 1 Some Oracle documentation and =-=[3]-=- use the term Oracle Read Consistency for the RC MV algorithm. 2 Unlike what happens with SI, under RC MV, a write does not force the transaction to abort in case a concurrent transaction has already  viors which the lower levels allow. The standard does not however mandate particular implementation approaches, and there are some well-known flaws in the way the behaviors of each level are specified=-=[3]-=-. If we assume that the well-known locking-based concurrency control algorithms are used, then the hierarchy is valid. Each algorithm holds all the locks of the lower levels, and also 3 As a check on  implement multiversion algorithms for concurrency control. Snapshot Isolation is treated as a strong algorithm (in particular, it prevents all the anomalies described in the ANSI standard as shown by =-=[3]-=-, although it does allow an anomaly called Write-Skew). It is generally assumed that the multiversion Read Committed algorithm is weaker than SI, because it does allow anomalies like Inconsistent Read everal platforms use SI when an application is declared to be Serializable, and they use the multiversion form of Read Committed when the application requests the lower isolation level. Despite this, =-=[3]-=- shows that SI is not strictly less permissive than RC MV. Consider the interleaving shown in Eq (1), where rX(A) denotes transaction X reading item A, cY is the commit of transaction Y, etc. rX(A)rY   </text>
<query_num> 21802 </query_num>
<text>   MS engines is presented very clearly in [15]. For locking approaches, the widely-used techniques include an efficient lock manager [13], with phantoms prevented by some form of next-key or index lock =-=[19, 20, 18]-=-. The alternative multiversion approach to concurrency control usually provides Snapshot Isolation; indeed several products run with SI even if the application declares its isolation level to be Seria   </text>
<query_num> 21803 </query_num>
<text>   ed on queueing theory and the focus is on the impact of different distributions for the random choices, such as the variance in the work needed for different requests, or in the inter-arrival spacing =-=[22]-=-. A different style aims for much simpler formulas, usually with grossly simplified assumptions and estimates of the probability of various situations arising. The seminal example is the prediction of   </text>
<query_num> 21804 </query_num>
<text>   ferent isolation levels. Another way to investigate performance is by detailed simulations, where a computer program representation of a whole system is evolved through many steps. An early paper was =-=[6]-=-. [1] compares the throughput obtained at different MPL, under different assumptions about the level of disk and CPU resources. A similar simulation study has focused on multiversion concurrency contr   </text>
<query_num> 21805 </query_num>
<text>   on, in terms of the parameters that define the configuration. Our model deals with both the multiversion isolation levels: SI and RC MV. We were inspired by Gray et al’s predictions of deadlock rates =-=[12, 11]-=-, and by the Elnikety et al model predicting performance of replication algorithms [7]. We illustrate our approach on a particular configuration (where MPL=10 and with SI) from the experiment of Figur ssly simplified assumptions and estimates of the probability of various situations arising. The seminal example is the prediction of waiting and deadlock probabilities in [12]; later examples include =-=[11, 7]-=-. 7. CONCLUSIONS AND FUTURE WORK We have described a way to quantify the extent to which different, weak, isolation levels allow data corruption. Our microbenchmark allows us to explore carefully the   </text>
<query_num> 21806 </query_num>
<text>   t force the transaction to abort in case a concurrent transaction has already committed changes to the same item.is no set of principles or guidelines to determine whether this is so for given code. =-=[17]-=- reported examples of deployed applications using SI which produced non-serializable executions. Thus in practice, what most developers actually see is a trade-off between performance and correctness. on was shown in [3]. A graph-based condition can determine whether particular application programs give rise to non-serializable execution under SI [9], and checks for this condition can be automated =-=[17]-=-. A variant of the SI algorithm provides true serializability [4]; however, this is not yet available in commercial platforms. There are several different ways in which the database community has quan   </text>
<query_num> 21807 </query_num>
<text>   t isolation levels. Another way to investigate performance is by detailed simulations, where a computer program representation of a whole system is evolved through many steps. An early paper was [6]. =-=[1]-=- compares the throughput obtained at different MPL, under different assumptions about the level of disk and CPU resources. A similar simulation study has focused on multiversion concurrency control al   </text>
<query_num> 21808 </query_num>
<text>   tandard. The phantom problem, and the distinction between Repeatable Read and Serializable is described in [8]. The state-of-practice concurrency control for DBMS engines is presented very clearly in =-=[15]-=-. For locking approaches, the widely-used techniques include an efficient lock manager [13], with phantoms prevented by some form of next-key or index lock [19, 20, 18]. The alternative multiversion a   </text>
<query_num> 21809 </query_num>
<text>   the stronger level. Finding this also demonstrates the usefulness of our predictive model. 6. RELATED WORK The proposal to provide application developers with a choice of isolation levels comes from =-=[14]-=-. In this paper the proposal gives specific locking algorithms that provide each level; the levels are named “Degree 0”, “Degree 1”, etc. Degree 2 is what was later called Read Committed in the SQL st   </text>
<query_num> 21810 </query_num>
<text>   the throughput obtained at different MPL, under different assumptions about the level of disk and CPU resources. A similar simulation study has focused on multiversion concurrency control algorithms =-=[5]-=-. Instead of measuring performance in a real system or a computer model, another approach tries to find a formula which relates performance to configuration parameters. In the analytical style, the fo   </text>
</top>
<top>
<paper_num> 219 </paper_num>
<paper_title>   Scoped Types for Real-Time Java.  </paper_title>
<paper_abstract>   Abstract — A memory model based on scoped areas is one of the distinctive features of the Real-Time Specification for Java (RTSJ). Scoped Types ensure timely reclamation of memory and predictable performance. The price to pay for these benefits is an unfamiliar programming model that, at the same time, is complex, requires checking all memory accesses, and rewards design-time errors with run-time crashes. We investigate an alternative approach, referred to as Scoped Types, that simplifies the task of managing memory in real-time codes. The key feature of our proposal is that the run-time partition of memory imposed by scoped areas is straightforwardly mirrored in the program text. Thus cursory inspection of a program reveals which objects will inhabit the different scopes, significantly simplifying the task of understanding real-time Java programs. Moreover, we introduce a type system which ensures that no run-time errors due to memory access checks will occur. Thus a RTSJ-compliant virtual machine does not require memory access checks. The contributions of this paper are the concept of Scoped Types, and a proof soundness of the type system. Experimental results will be described in future work. I.  </paper_abstract>
<query_num> 21901 </query_num>
<text>   In Ravenscar memory areas cannot be nested and are single threaded. Scoped Types are intended to relax some of the restrictions of Ravenscar while remaining easy to understand and to verify. Boyapati =-=[13]-=- combine region-based memory management with ownership types to statically guarantee that real-time threads do not interfere with GC. While more flexible than Scoped Types, this approach is more invas   </text>
<query_num> 21902 </query_num>
<text>   THE SJ CALCULUS To gain confidence in the programming model underlying our proposal, we introduce the SJ calculus, a sparse imperative and concurrent object calculus, modeled after Featherweight Java =-=[11]-=-, in which scopes are first-class values. SJ formalizes the type confinement rules of Scoped Type in terms of a type system. Our proof of type soundness gives us the guarantee that confinement cannot   </text>
<query_num> 21903 </query_num>
<text>   e type cannot be widened to other types. S3 The methods invoked on a gate type must be defined in the same class.sstatic types. These rules are similar in spirit to the confinement rules presented in =-=[9]-=-. The type system presented in the next section formalizes these intuitive rules. Restrictions. Scoped Types do restrict the set of valid Java programs. Even though they do not require changes to the   memory such as the size and the type of the memory area (linear or variable allocation time) are also omitted. A. Syntax and Types The syntax of the SJ calculus, Figure 7, draws on our previous work =-=[9]-=-. The formalism and syntax is based on the Featherweight Java (FJ) system which has been widely adopted as a vehicle of language research. SJ has two kinds of class declarations, scoped classes and ga  of the variable this in the method call and consequently, objects in the outer scope may hold references to the receiver object. To prevent such problems, we may require such methods to be anonymous =-=[9]-=- so that the variable this can only be used for field access and calls to other anonymous methods. Another limitation of Scoped Types is the lack of reuse of library classes. This problem may be addre   </text>
<query_num> 21904 </query_num>
<text>   guarantee that real-time threads do not interfere with GC. While more flexible than Scoped Types, this approach is more invasive, requiring more program annotations, and more complex overall. Cyclone =-=[14]-=- is a type-safe language derived from C and it supports region-based memory management. Cyclone includes dynamic regions with lexically scoped lifetimes, stack and a heap region. To prevent dereferenc   </text>
<query_num> 21905 </query_num>
<text>   h real-time Java, since the Java’s type safety requirement does not allow objects to hold invalid references even if never used. Grossman extended Cyclone with a type system for preventing data races =-=[15]-=-. The MLKit is an implementation of ML which uses regions and region-inference [16], [17]. One of the main difference with the model presented here is that ML is a functional language without built-in   </text>
<query_num> 21906 </query_num>
<text>   invalid references even if never used. Grossman extended Cyclone with a type system for preventing data races [15]. The MLKit is an implementation of ML which uses regions and region-inference [16], =-=[17]-=-. One of the main difference with the model presented here is that ML is a functional language without built-in support for concurrency. There are two other open source virtual machines that implement   </text>
<query_num> 21907 </query_num>
<text>   nce with the model presented here is that ML is a functional language without built-in support for concurrency. There are two other open source virtual machines that implement parts of the RTSJ: Flex =-=[18]-=- and JRate [19], [5], as well as a number of commercial products and alternative proposals [20], [21], [22], [23], [24].sVI. CONCLUSION In this paper we have introduced Scoped Types, a static programm   </text>
<query_num> 21908 </query_num>
<text>   o hold invalid references even if never used. Grossman extended Cyclone with a type system for preventing data races [15]. The MLKit is an implementation of ML which uses regions and region-inference =-=[16]-=-, [17]. One of the main difference with the model presented here is that ML is a functional language without built-in support for concurrency. There are two other open source virtual machines that imp   </text>
<query_num> 21909 </query_num>
<text>   rogram called Zen to use Scoped Types. Zen is a CORBA object request broker designed to support distributed, real-time and embedded (DRE) applications, written in the Real-time Specification for Java =-=[10]-=-. Zen has been designed for memory-constrained DRE applications. For our experiment we have selected a minimal configuration (about 20K LOC of Java code) that provides sufficient functionality for a n   </text>
<query_num> 21910 </query_num>
<text>   te σ(ℓ0) are not reachable in σ ′ , P ′ . V. RELATED WORK The dangers involved in the RTSJ programming model have motivated Kwon et.al. to propose a restricted programming model called Ravenscar-Java =-=[12]-=-. In Ravenscar memory areas cannot be nested and are single threaded. Scoped Types are intended to relax some of the restrictions of Ravenscar while remaining easy to understand and to verify. Boyapat   </text>
</top>
<top>
<paper_num> 220 </paper_num>
<paper_title>   Towards efficient and scalable mediation: the AURORA approach.  </paper_title>
<paper_abstract>   We develop a 2-tier, plug-and-play mediation model for accessing a large number of heterogeneous data sources. This model defines a divide-and-conquer approach towards information integration. It is more suitable for applications such as electronic commerce than existing models. We also develop algebras that manipulate heterogeneous data, the mediation enabling algebras, that provide new techniques for efficient query processing in large-scale middleware. This paper presents the mediation model, architecture and techniques studied in the AURORA project. 1 Introduction  Today, a vast amount of digital information is provided by sources ranging from database systems, WWW pages, file systems and spreadsheets, to special-purpose data repositories. With the advent of the Internet, the way people use information is changing rapidly. In general, software that facilitates access to multiple heterogeneous information sources is needed. Such software is commonly known as a middleware.  Accessing...  </paper_abstract>
<query_num> 22001 </query_num>
<text>   al with this situation =-=[LRO96b, TRV96, LPL96]-=-. DISCO =-=[TRV96]-=- extends the ODMG ODL to allow a bag of extents for a single interface type. Adding a new source means adding an extent into this bag. DIOM =-=[LPL96]-=- intends to identify relevant data sources and bind to them at runtime. The Information Manifold project =-=[LRO96b, Lev96, LRO96a]-=- considers large number of data sources with varying query capabilities.   </text>
<query_num> 22002 </query_num>
<text>   amic Sources With the Internet and WWW, more data sources are open with diverse availability and query processing capabilities. A few approaches deal with this situation =-=[LRO96b, TRV96, LPL96]-=-. DISCO =-=[TRV96]-=- extends the ODMG ODL to allow a bag of extents for a single interface type. Adding a new source means adding an extent into this bag. DIOM =-=[LPL96]-=- intends to identify relevant data sources and bind t ed. Data sources contribute to and withdraw from this view without affecting other participating sources or the applications that access this view. To our knowledge, two systems use this model: DISCO =-=[TRV96]-=- and Information Manifold (IM) =-=[LRO96b]-=-. Apparently, derivation-based models do not favor scalability since including or excluding a source is difficult when the number of sources involved is large. T   </text>
<query_num> 22003 </query_num>
<text>   lized view. The problem of answering a query against the worldview is transformed to that of answering a query with existing materialized views, with additional constraints. This problem is solved in =-=[LRO96b]-=-. Handling sources with limited query capabilities is an especially useful technique for accessing sources such as those on the Web. 4 The AURORA Project The AURORA project builds mediators that can b hdraw from this view without affecting other participating sources or the applications that access this view. To our knowledge, two systems use this model: DISCO =-=[TRV96]-=- and Information Manifold (IM) =-=[LRO96b]-=-. Apparently, derivation-based models do not favor scalability since including or excluding a source is difficult when the number of sources involved is large. The plug-and-play model avoids this prob   </text>
<query_num> 22004 </query_num>
<text>   query optimization strategy; and (4) developing techniques for evaluating expensive MEOs efficiently. A specific suite for AURORA-RH mediator has already developed. It has been described in detail in =-=[YOL97a]-=- and is reviewed briefly in Section 5.1. This suite indeed sets the paradigm of all suites. Other suites can be developed in a similar fashion. Work in integration mediator suites are in early stage.  allows specification of domain value functions; these can be user-defined functions or stored mapping tables. Detailed description of these environments and examples of their application are given in =-=[YOL97a]-=-. 5.1.3 AURORA-RH Query Processor Query processing in AURORA-RH is based on MEA-RH. Homogenizing views are defined using these operators via MAT-RH. A mediator query is processed in three steps. First writing algorithm has been developed and is given in =-=[YOL97a]-=-. Second, we transform this formula into an &amp;quot;optimal&amp;quot; form using transformation rules. A complete set of transformation rules are given in =-=[YOL97a]-=-. These rules allow exchanging the usual relational operators with MEA-RH operators. Query optimization in AURORA aims at maximizing queries sent to the source so as to leverage the query optimization Third, we assemble query results returned from the data source to produce an answer to the mediator query. This assembly uses MEA-RH operators. A complete walk-through of query processing is given in =-=[YOL97a]-=- by an example. 5.2 Integration Mediators AURORA integration mediators, AURORA-RI and AURORA-OI (Figure 5), are responsible for integrating a large number of homogenized sources. Since the sources are   </text>
<query_num> 22005 </query_num>
<text>   the access scope. Scalability requires that both steps be performed rapidly. For 1, this requires rapid, if not automatic, wrapper generation. Various enabling techniques have already been developed =-=[PGGMU95]-=-. Much is known about how 2 can be done, the scalability aspect of it has not been well addressed (ref. Section 4.1.3). The scalability issue in AURORA is reduced to the issues of rapid homogenization   </text>
<query_num> 22006 </query_num>
<text>   ting Technologies A few paradigms exist for facilitating integrated access to heterogeneous and autonomous data sources, notably the federated database systems (FDBs) =-=[SL90]-=-, and the mediator systems =-=[Wie92]-=-. More recent approaches handle dynamic sources and sources with diverse query processing capabilities. 3.1 Federated Database Systems Federated databases (FDBs) =-=[Chu90, LR82, THMB95, Ahm91, ACHK93,-=- C pancies at query time. 3.2 Mediator Systems A mediator is a software module that exploits encoded knowledge about some sets or subsets of data to create information for a higher layer of applications =-=[Wie92]-=-. A mediator system has the form illustrated in Figure 3. A wrapper is a special mediator that handles idiosyncrasies of individual data sources. Today, many mediator systems are being built =-=[PGMW95, -=-  </text>
<query_num> 22007 </query_num>
<text>   to assist in construction. 3 Existing Technologies A few paradigms exist for facilitating integrated access to heterogeneous and autonomous data sources, notably the federated database systems (FDBs) =-=[SL90]-=-, and the mediator systems =-=[Wie92]-=-. More recent approaches handle dynamic sources and sources with diverse query processing capabilities. 3.1 Federated Database Systems Federated databases (FDBs) =-=[Chu-=-   </text>
</top>
<top>
<paper_num> 221 </paper_num>
<paper_title>   Darwin phones: the evolution of sensing and inference on mobile phones.  </paper_title>
<paper_abstract>   We present Darwin, an enabling technology for mobile phone sensing that combines collaborative sensing and classification techniques to reason about human behavior and context on mobile phones. Darwin advances mobile phone sensing through the deployment of efficient but sophisticated machine learning techniques specifically designed to run directly on sensor-enabled mobile phones (i.e., smartphones). Darwin tackles three key sensing and inference challenges that are barriers to mass-scale adoption of mobile phone sensing applications: (i) the human-burden of training classifiers, (ii) the ability to perform reliably in different environments (e.g., indoor, outdoor) and (iii) the ability to scale to a large number of phones without jeopardizing the “phone experience ” (e.g., usability and battery lifetime). Darwin is a collaborative reasoning framework built on three concepts: classifier/model evolution, model pooling, and collaborative inference. To the best of our knowledge Darwin is the first system that applies distributed machine learning techniques and collaborative inference concepts to mobile phones. We implement the Darwin system on the Nokia N97 and Apple iPhone. While Darwin represents a general framework applicable to a wide variety of emerging mobile sensing applications, we implement a speaker recognition application and an augmented reality application to evaluate the benefits of Darwin. We show experimental results from eight individuals carrying Nokia N97s and demonstrate that Darwin improves the reliability and scalability of the proof-ofconcept speaker recognition application without additional burden to users.  </paper_abstract>
<query_num> 22101 </query_num>
<text>   ery lifetime (∼27 hours) for a periodic sampling interval of 10 seconds (this sampling interval guarantees the highest inference responsiveness). However, if smart duty-cycling techniques are adopted =-=[48]-=-, the phone could operate in a low sensing duty-cycle mode, e.g., with a sampling rate of 60 seconds, when Darwin is not running. As an event is detected, such as voice in case of speaker recognition, urce consumption and the development of low-energy duty cycling for Darwin are important future work. We believe however that new duty cycling techniques discussed in the literature for mobile phones =-=[48, 32]-=- could boost the phone’s battery lifetime of Darwin phones. 5. DEMO APPLICATIONS Darwin can be used by other emerging mobile phone sensing applications in addition to the speaker recognition applicati   </text>
<query_num> 22102 </query_num>
<text>   for example, tagging the people that the local inferences returned with highest confidence. 6. RELATED WORK Work on applications and systems for sensing enabled mobile phones is growing in importance =-=[31, 9, 35, 8, 19, 21, 13, 27, 10, 30, 29]-=-. Most of the work in the literature, however, propose local sensing operations running on individual devices and do not exploit in-field mobile phones interactions. An exception to this is the work i   </text>
<query_num> 22103 </query_num>
<text>   ince we propose collaborative inference techniques that combine with classifier evolution and model pooling. Semi-supervised machine learning techniques are investigated for word sense disambiguation =-=[50]-=-, to identify subjective nouns [41], or to classify emotional and non emotional dialogues [28]. However, no work studies semi-supervised learning techniques in the context of mobile sensing applicatio   </text>
<query_num> 22104 </query_num>
<text>   n [19], which considers context driven sampling and calibration techniques for mobile sensor networks. Sensor node co-operation is studied mainly in the context of static sensor networks where fusion =-=[24, 38, 49]-=- and aggregation [52, 45, 33] techniques are applied. The benefit of sensor nodes cooperation in the context of object tracking using distributed Kalman Filters is discussed in [36, 37]. In [24] the a   </text>
<query_num> 22105 </query_num>
<text>   n are used by eight people. The reason we select speaker recognition is not because we intend to design a new speaker recognition algorithm (there is a considerable amount of literature on this topic =-=[44, 43, 26, 18, 51]-=-), but to show how Darwin improves a mobile sensing application inference quality. Darwin is founded on an opportunistic sensing paradigm [12], where the user is not an active participant in the sensi r, no work studies semi-supervised learning techniques in the context of mobile sensing applications or frameworks. Audio analysis for speaker identification is a well explored area in the literature =-=[17, 42, 16, 44, 43, 26, 18, 40, 23]-=-. Although we do not propose new speaker recognition techniques, we show how to build a lightweight speaker identification application capable of running on mobile phones. 7. CONCLUSION In this paper   </text>
<query_num> 22106 </query_num>
<text>   ve task performance, and to aide in the assessment of health and wellness. There is a growing research effort in using mobile phones to infer information about people’s behavior and their environment =-=[46, 21, 20, 30, 29, 10, 27]-=-. These systems typically rely on pre-trained models or classifiers, where the training data from events of interest are acquired in advance. It is often exceedingly hard to obtain a representative tr  that run on the phone to process sensor data should be implemented in an efficient lightweight manner. Darwin is designed to reduce on-the-phone computation based on a split-level computation design =-=[29]-=-, offloading some of the work to backend servers (as discussed in Section 2) while trading off the cost for local computation and wireless communication with backend servers. Users carrying mobile pho r WiFi scan could potentially return in the area where the person is. We have developed Virtual Square, an application that exploits augmented reality to present a person’s sensing status information =-=[29]-=- including whom the person is in proximity/chatting with at a certain moment and location. Our Virtual Square prototype is built for the Nokia N97 writing a combination of QT code, for the body of the for example, tagging the people that the local inferences returned with highest confidence. 6. RELATED WORK Work on applications and systems for sensing enabled mobile phones is growing in importance =-=[31, 9, 35, 8, 19, 21, 13, 27, 10, 30, 29]-=-. Most of the work in the literature, however, propose local sensing operations running on individual devices and do not exploit in-field mobile phones interactions. An exception to this is the work i   </text>
<query_num> 22107 </query_num>
<text>   ve task performance, and to aide in the assessment of health and wellness. There is a growing research effort in using mobile phones to infer information about people’s behavior and their environment =-=[46, 21, 20, 30, 29, 10, 27]-=-. These systems typically rely on pre-trained models or classifiers, where the training data from events of interest are acquired in advance. It is often exceedingly hard to obtain a representative tr for example, tagging the people that the local inferences returned with highest confidence. 6. RELATED WORK Work on applications and systems for sensing enabled mobile phones is growing in importance =-=[31, 9, 35, 8, 19, 21, 13, 27, 10, 30, 29]-=-. Most of the work in the literature, however, propose local sensing operations running on individual devices and do not exploit in-field mobile phones interactions. An exception to this is the work i   </text>
<query_num> 22108 </query_num>
<text>   xt driven sampling and calibration techniques for mobile sensor networks. Sensor node co-operation is studied mainly in the context of static sensor networks where fusion [24, 38, 49] and aggregation =-=[52, 45, 33]-=- techniques are applied. The benefit of sensor nodes cooperation in the context of object tracking using distributed Kalman Filters is discussed in [36, 37]. In [24] the authors propose distributed en   </text>
</top>
<top>
<paper_num> 222 </paper_num>
<paper_title>   A Comparative Study of Methods for Transductive Transfer Learning.  </paper_title>
<paper_abstract>   The problem of transfer learning, where information gained in one learning task is used to improve performance in another related task, is an important new area of research. While previous work has studied the supervised version of this problem, we study the more challenging case of unsupervised transductive transfer learning, where no labeled data from the target domain are available at training. We describe some current state-of-the-art inductive and transductive approaches and then adapt these models to the problem of transfer learning for protein name extraction. In the process, we introduce a novel maximum entropy based technique, Iterative Feature Transformation (IFT), and show that it achieves comparable performance with state-of-the-art transductive SVMs. We also show how simple relaxations, such as providing additional information like the proportion of positive examples in the test data, can significantly improve the performance of some of the transductive transfer learners. 1  </paper_abstract>
<query_num> 22201 </query_num>
<text>   , and thus serves as a nice compromise between the two extremes of transduction and supervision. 3 Methods considered 3.1 Maximum entropy models 3.1.1 Inductive learning Entropy maximization (MaxEnt) =-=[2, 14]-=- is a way of modeling the conditional distribution of labels given examples. Given a set of training examples Xtrain ≡{x1,...,xN }, their labels Ytrain ≡ {y1,...,yN}, and the set of features F ≡ {f1,.   </text>
<query_num> 22202 </query_num>
<text>   R) 160 48,417 15.1% Yapex-test (YTT) 40 12,113 14.5% 4.2 Data and evaluation Our corpora are abstracts from biological journals coming from two sources: University of Texas, Austin (UT) [3] and Yapex =-=[8]-=-. Each abstract was tokenized and each token was hand-labeled as either being part of a protein name or not. We used a standard natural language toolkit [5] to compute tens of thousands of binary feat   </text>
<query_num> 22203 </query_num>
<text>   apex-train (YTR) 160 48,417 15.1% Yapex-test (YTT) 40 12,113 14.5% 4.2 Data and evaluation Our corpora are abstracts from biological journals coming from two sources: University of Texas, Austin (UT) =-=[3]-=- and Yapex [8]. Each abstract was tokenized and each token was hand-labeled as either being part of a protein name or not. We used a standard natural language toolkit [5] to compute tens of thousands   </text>
<query_num> 22204 </query_num>
<text>   been noted that such auxiliary data typically helps boost the performance of the classifier significantly. Another setting that is closely related to semi-supervised learning is transductive learning =-=[18, 11, 13]-=-, in which Xtest (but, importantly, not Ytest), is known at training time. That is, the learning algorithm knows exactly which examples it will be evaluated on after training. This can be a great asse   </text>
<query_num> 22205 </query_num>
<text>   been noted that such auxiliary data typically helps boost the performance of the classifier significantly. Another setting that is closely related to semi-supervised learning is transductive learning =-=[18, 11, 13]-=-, in which Xtest (but, importantly, not Ytest), is known at training time. That is, the learning algorithm knows exactly which examples it will be evaluated on after training. This can be a great asse eparate. Since we do not know the labels of the testing data, however, we cannot perform a straight forward margin maximization, as in the supervised case. Instead, one can use an iterative algorithm =-=[11]-=- similar in flavor to the MaxEnt iterative feature transformation (IFT) algorithm of section 3.1.3. Specifically, a hyperplane is trained on the labeled source data and then used to classify the unlab   </text>
<query_num> 22206 </query_num>
<text>   e labels) is the same for both D source and D target , while D source and D target themselves are allowed to vary between domains. This is in contrast to the related subproblem of multi-task learning =-=[1, 17]-=- in which the marginal distribution of the data is assumed not to change, while the task (and therefore the labels) is allowed to vary from source to target. In this paper we choose to focus on extens   </text>
<query_num> 22207 </query_num>
<text>   ion problem. In the paradigm of inductive learning, (Xtrain,Ytrain) are known, while both Xtest and Ytest are completely hidden during training time. In the case of semi-supervised inductive learning =-=[20, 16, 9]-=-, the learner is also provided with auxiliary unlabeled data Xauxiliary, that is not part of the test set. It has been noted that such auxiliary data typically helps boost the performance of the class   </text>
</top>
<top>
<paper_num> 223 </paper_num>
<paper_title>   On the Semantics of a Semantic Network.  </paper_title>
<paper_abstract>   We elaborate on the semantics of an enhanced object-oriented semantic network, where multiple instantiation, multiple specialization, and meta-classes are supported for both kinds of objects: entities and properties. By semantics of a semantic network, we mean the information (both explicit  and derived) that the semantic network carries. Several data models use semantic networks to organize information. However, many of these models do not have a formalism defining what the semantics of the semantic network is. In our data model, in addition to the Isa relation, we consider a stronger form of specialization for properties, that we call restriction isa, or Risa for short. The Risa relation expresses property value refinement. A distinctive feature of our data model is that it supports the interaction between  Isa and Risa relations. The combination of Isa and Risa provides a powerful conceptual modeling mechanism.    Research conducted while this author was visiting with the Institute ...  </paper_abstract>
<query_num> 22301 </query_num>
<text>   Token Level p p&amp;apos; p&amp;quot; c c&amp;apos; Figure 7. Example of typing properties having same label and same semantics We will compare our approach to typing property inheritance and typechecking with that of F-logic =-=[26]-=-, terminological languages [7, 33, 28, 4], DOT [41, 42], QUIXOTE [45], and LOGIN [1]. First, we present common limitations of these data models: (i) These data models support Risa only implicitly, bas ar; Boat), which is a known class. We now present the approaches followed by F-logic, terminological languages, DOT, and QUIXOTE, with respect to typing property inheritance and typechecking. F-LOGIC =-=[26]-=- F-logic is a powerful deductive object-oriented language that supports inheritance of typing properties. Let c be a class having a typing property with label l. Then, the statement c[l ) fd 1 ; :::;   </text>
<query_num> 22302 </query_num>
<text>   a data model, we feel that it imposes a discipline that protects against the declaration of erroneous information. Many object-oriented data models that define their semantics based on logic, such as =-=[29, 8, 6, 31, 18]-=-, do not consider inheritance of typing properties 6 . Yet, in many applications, reasoning on the structural definitions of the data is a necessity [32]. A property inherited by a class provides info   </text>
<query_num> 22303 </query_num>
<text>   cs and, in particular, property inheritance, in a procedural way. A detailed comparison between our approach to inheritance and that of several systems, such as ORION [5], O 2 [17], ODE [2], POSTGRES =-=[35, 37]-=-, EXODUS [10], is given in [3]. In this section, we review systems that define their semantics based on logic. We first establish a common framework and vocabulary for comparing our data model with re   </text>
<query_num> 22304 </query_num>
<text>   ng properties having same label and same semantics We will compare our approach to typing property inheritance and typechecking with that of F-logic [26], terminological languages [7, 33, 28, 4], DOT =-=[41, 42]-=-, QUIXOTE [45], and LOGIN [1]. First, we present common limitations of these data models: (i) These data models support Risa only implicitly, based on property labels. Specifically, if two classes c,  g at instance and schema level. We consider this to be a severe limitation, as adding meaning to the data, should be accompanied by convenient ways of querying the schema (through a meta-schema). DOT =-=[41, 42]-=- The knowledge representation model DOT describes property values using the Isa relation and supports typing property inheritance. In this model, the In relation is not distinguished from Isa (for thi   </text>
<query_num> 22305 </query_num>
<text>   sting information. We call the information (both explicit and derived) that a semantic network carries, the semantics of the network. Several data models use semantic networks to organize information =-=[12, 21, 39, 43, 7, 23, 34, 30, 19, 22, 9]-=- and their usefulness to conceptual modeling is unquestionable. However, many of these models do not provide a formalism defining what the semantics of the semantic network is. This can lead to incons   </text>
<query_num> 22306 </query_num>
<text>   t-oriented semantic network, where multiple instantiation, multiple specialization, and meta-classes are supported for both nodes and links. The context of our work is the Semantic Index System (SIS) =-=[14, 15, 16, 40]-=-. In fact, the data model presented in this paper, is a (self-contained) part of the SIS data model. The SIS is targeted at supporting large descriptive knowledge structures in real applications. Typi   </text>
<query_num> 22307 </query_num>
<text>   their semantics and, in particular, property inheritance, in a procedural way. A detailed comparison between our approach to inheritance and that of several systems, such as ORION [5], O 2 [17], ODE =-=[2]-=-, POSTGRES [35, 37], EXODUS [10], is given in [3]. In this section, we review systems that define their semantics based on logic. We first establish a common framework and vocabulary for comparing our   </text>
</top>
<top>
<paper_num> 224 </paper_num>
<paper_title>   Randomized mutual exclusion in O(log N / log log N) RMRs.  </paper_title>
<paper_abstract>   Mutual exclusion is a fundamental distributed coordination problem. Shared-memory mutual exclusion research focuses on local-spin algorithms and uses the remote memory references (RMRs) metric. A recent proof [9] established an Ω(log N) lower bound on the number of RMRs incurred by processes as they enter and exit the critical section, matching an upper bound by Yang and Anderson [18]. Both these bounds apply for algorithms that only use read and write operations. The lower bound of [9] only holds for deterministic algorithms, however; the question of whether randomized mutual exclusion algorithms, using reads and writes only, can achieve sub-logarithmic expected RMR complexity remained open. This paper answers this question in the affirmative. We present two strong-adversary [8] randomized local-spin mutual exclusion algorithms. In both algorithms, processes incur O(log N / log log N) expected RMRs per passage in every execution. Our first algorithm has sub-optimal worstcase RMR complexity of O ( (log N / log log N) 2). Our second algorithm is a variant of the first that can be combined with a deterministic algorithm, such as [18], to obtain O(log N) worst-case RMR complexity. The combined algorithm thus achieves sub-logarithmic expected RMR complexity while maintaining optimal worst-case RMR complexity. Our upper bounds apply for both the cache coherent (CC) and the distributed shared memory (DSM) models.  </paper_abstract>
<query_num> 22401 </query_num>
<text>   esses’ future steps. The question of whether randomization can help break the logarithmic barrier of [9] thus remained open. In this paper, we provide a positive answer to this question. Golab et al. =-=[12]-=- presented a constant-RMRs read/write implementation of compare-and-swap. Moreover, they proved that any shared-memory algorithm using reads, writes, and conditional operations (such as compare-and-sw  a read/write algorithm, with only a constant multiplicative increase in RMR complexity. The algorithms we present use variables that support the compare-and-swap and read operations. It follows from =-=[12]-=-, that these variables can be implemented from reads and writes only while maintaining asymptotic RMR complexities. 1 Our Contributions. We establish a separation in terms of RMR complexity between ra he compare-and-swap operation (abbreviated CAS) is defined as follows: CAS(v,expected,new) changes the value of variable v to new only if its value just before CAS is ap1 In the CAS implementation of =-=[12]-=-, the CAS operation returns the previous object value. Although a read operation was not implemented in [12], it can be easily implemented by calling CAS with an expected value outside the object’s va nd we say it is successful. Otherwise, CAS does not change the value of v and returns false; in this case, we say that the CAS was unsuccessful. A stronger version of CAS, such as that implemented in =-=[12]-=-, returns the value of v just before CAS is applied, instead of returning true or false. 2. THE RANDOMIZED ALGORITHM In this section, we describe our first randomized mutualexclusion algorithm. In the w stores the number of the lowest level whose lock was not captured by p. If the CAS operation of either line 13 or line 14 fails, 2 This is done by a CAS operation, since the CAS object presented in =-=[12]-=- does not support a write operation.p busy-waits in line 20 until it either identifies that it was promoted (this occurs immediately if p previously failed at line 14) or it reads ⊥ from n’s lock. If   </text>
<query_num> 22402 </query_num>
<text>   le shared variables, achieve bounded RMR complexity and have practical performance benefits [7]. Indeed, recent mutual exclusion research investigates the RMR complexity of local-spin algorithms (see =-=[2, 5, 6, 10, 14, 15, 16]-=- for some examples). Anderson and Kim presented a simple randomized variant [16] of their (deterministic) read/write adaptive mutual exclusion algorithm [5]. Their randomized variant has expected O(lo   </text>
<query_num> 22403 </query_num>
<text>   le shared variables, achieve bounded RMR complexity and have practical performance benefits [7]. Indeed, recent mutual exclusion research investigates the RMR complexity of local-spin algorithms (see =-=[2, 5, 6, 10, 14, 15, 16]-=- for some examples). Anderson and Kim presented a simple randomized variant [16] of their (deterministic) read/write adaptive mutual exclusion algorithm [5]. Their randomized variant has expected O(lo  deterministic algorithms with O(log k) RMR complexity, this established a separation in terms of RMR complexity between randomized and deterministic adaptive algorithms. With the single exception of =-=[16]-=-, prior art local-spin mutual exclusion research dealt exclusively with deterministic algorithms. Yang and Anderson presented the first O(log N) RMRs read/write mutual exclusion algorithm [18]. Anders   </text>
<query_num> 22404 </query_num>
<text>   local-spin mutual exclusion research dealt exclusively with deterministic algorithms. Yang and Anderson presented the first O(log N) RMRs read/write mutual exclusion algorithm [18]. Anderson and Kim =-=[4]-=- conjectured that this was best possible. This conjecture was recently proved by Attiya, Hendler and Woelfel [9]. The lower bound of [9] holds only for deterministic algorithms, however, since it assu   </text>
<query_num> 22405 </query_num>
<text>   lusion algorithms, using reads and writes only, can achieve sub-logarithmic expected RMR complexity remained open. This paper answers this question in the affirmative. We present two strong-adversary =-=[8]-=- randomized local-spin mutual exclusion algorithms. In both algorithms, processes incur O(log N/ log log N) expected RMRs per passage in every execution. Our first algorithm has sub-optimal worstcase  ent starvation-free randomized mutual exclusion algorithms for the CC model, in which processes incur O(log N/ log log N) expected RMRs per passage in every execution, even against a strong-adversary =-=[8]-=- that can schedule processes according to their execution history. Our first algorithm has sub-optimal worst-case RMR complexity of O ( (log N/ log log N) 2) . Our second algorithm is a variant of the   </text>
<query_num> 22406 </query_num>
<text>   s) metric. A recent proof [9] established an Ω(log N) lower bound on the number of RMRs incurred by processes as they enter and exit the critical section, matching an upper bound by Yang and Anderson =-=[18]-=-. Both these bounds apply for algorithms that only use read and write operations. The lower bound of [9] only holds for deterministic algorithms, however; the question of whether randomized mutual exc Our first algorithm has sub-optimal worstcase RMR complexity of O ( (log N/ log log N) 2) . Our second algorithm is a variant of the first that can be combined with a deterministic algorithm, such as =-=[18]-=-, to obtain O(log N) worst-case RMR complexity. The combined algorithm thus achieves sub-logarithmic expected RMR complexity while maintaining optimal worst-case RMR complexity. Our upper bounds apply tion of [16], prior art local-spin mutual exclusion research dealt exclusively with deterministic algorithms. Yang and Anderson presented the first O(log N) RMRs read/write mutual exclusion algorithm =-=[18]-=-. Anderson and Kim [4] conjectured that this was best possible. This conjecture was recently proved by Attiya, Hendler and Woelfel [9]. The lower bound of [9] holds only for deterministic algorithms,  ur first algorithm has sub-optimal worst-case RMR complexity of O ( (log N/ log log N) 2) . Our second algorithm is a variant of the first that can be combined with a deterministic algorithm, such as =-=[18]-=-, to obtain O(log N) worst-case RMR complexity. The combined algorithm thus achieves sublogarithmic expected RMR complexity while maintaining optimal worst-case RMR complexity. We then describe (in Se se RMR complexity, which is sub-optimal. In this section, we present a variant of that algorithm, henceforth called the quitting algorithm, that can be combined with a deterministic algorithm such as =-=[18]-=-. The resulting Combined Randomized Deterministic (henceforth CRD) algorithm achieves optimal worst-case RMR complexity of O(log N) while maintaining expected RMR complexity of O(log N/ log N log N).   </text>
<query_num> 22407 </query_num>
<text>   t a single process is inside the critical section. Introduced by Dijkstra in 1965 [11], the mutual exclusion problem is at the core of Distributed Computing and is still the focus of intense research =-=[3, 17]-=-. In this paper, we consider mutual exclusion in the asynchronous shared-memory model. A natural way to measure the time complexity of algorithms in this model is to count the number of sharedmemory a   </text>
</top>
<top>
<paper_num> 225 </paper_num>
<paper_title>   Dynamic Black-Box Performance Model Estimation for Self-Tuning Regulators.  </paper_title>
<paper_abstract>   Methods for automatically managing the performance of computing services must estimate a performance model of that service. This paper explores properties that are necessary for performance model estimation of black-box computer systems when used together with adaptive feedback loops. It shows that the standard method of least-squares estimation often gives rise to models that make the control loop perform the opposite action of what is desired. This produces large oscillations and bad tracking performance. The paper evaluates what combination of input and output data provides models with the best properties for the control loop. Plus, it proposes three extensions to the controller that makes it perform well, even when the model estimated would have degraded performance. Our proposed techniques are evaluated with an adaptive controller that provides latency targets for workloads on black-box computer services under a variety of conditions. The techniques are evaluated on two systems: a three-tier ecommerce site and a web server. Experimental results show that our best estimation approach improves the ability of the controller to meet the latency goals significantly. Previously oscillating workload latencies are with our techniques smooth around the latency targets. 1  </paper_abstract>
<query_num> 22501 </query_num>
<text>   atically maximizing the utility of data centers [22], and meeting performance goals in file systems [15, 16], 3-tier e-commerce sites [13, 16], disk arrays [3, 18, 23], databases [19] and web servers =-=[1, 20]-=-. For a solution to a specific management problem to be Magnus Karlsson and Michele Covell HP Labs Palo Alto, CA 94304, U.S.A. {magnus.karlsson,michele.covell}@hp.com 1 applicable to as many systems a   </text>
<query_num> 22502 </query_num>
<text>   automatically managing system with little or no human intervention. Examples of this includes, managing the energy consumption of servers [4, 14], automatically maximizing the utility of data centers =-=[22]-=-, and meeting performance goals in file systems [15, 16], 3-tier e-commerce sites [13, 16], disk arrays [3, 18, 23], databases [19] and web servers [1, 20]. For a solution to a specific management pro   </text>
<query_num> 22503 </query_num>
<text>   energy consumption of servers [4, 14], automatically maximizing the utility of data centers [22], and meeting performance goals in file systems [15, 16], 3-tier e-commerce sites [13, 16], disk arrays =-=[3, 18, 23]-=-, databases [19] and web servers [1, 20]. For a solution to a specific management problem to be Magnus Karlsson and Michele Covell HP Labs Palo Alto, CA 94304, U.S.A. {magnus.karlsson,michele.covell}@   </text>
<query_num> 22504 </query_num>
<text>   h managing computer systems have spurred a lot of interest in automatically managing system with little or no human intervention. Examples of this includes, managing the energy consumption of servers =-=[4, 14]-=-, automatically maximizing the utility of data centers [22], and meeting performance goals in file systems [15, 16], 3-tier e-commerce sites [13, 16], disk arrays [3, 18, 23], databases [19] and web s   </text>
<query_num> 22505 </query_num>
<text>   ite hosted on the 3-tier system is a version of the Java PetStore [11] that has been tuned in order to support a large number of concurrent users. The workload applied mimics real-world user behavior =-=[5]-=-, e.g., browsing, searching and purchasing behaviors including their respective time scales and probabilities. For the experiments here, we emulate 75 users partitioned into two classes. Each class is   </text>
<query_num> 22506 </query_num>
<text>   n intervention. Examples of this includes, managing the energy consumption of servers [4, 14], automatically maximizing the utility of data centers [22], and meeting performance goals in file systems =-=[15, 16]-=-, 3-tier e-commerce sites [13, 16], disk arrays [3, 18, 23], databases [19] and web servers [1, 20]. For a solution to a specific management problem to be Magnus Karlsson and Michele Covell HP Labs Pa . Classical non-adaptive control-theoretic methods are usually not adequate for two reasons. First, for many systems it is not even possible to use non-adaptive control as the system changes too much =-=[13, 15, 16]-=-. For example, the performance experience by a client of a three-tier e-commerce site depends on many things: what tier a request is served from, if it was served from the disk or the memory cache of  . 5 Related Work There are a few examples of self-tuning regulators used to control computer systems. Lu et al. [17] estimates a model between cache size and the hit ratio a workload receives. Triage =-=[15]-=- estimates a model between the total number of requests sent to the system and the latency. Karlsson et al. [16] estimates a model between shares and both latency and throughput. None of these approac   </text>
<query_num> 22507 </query_num>
<text>   ncludes, managing the energy consumption of servers [4, 14], automatically maximizing the utility of data centers [22], and meeting performance goals in file systems [15, 16], 3-tier e-commerce sites =-=[13, 16]-=-, disk arrays [3, 18, 23], databases [19] and web servers [1, 20]. For a solution to a specific management problem to be Magnus Karlsson and Michele Covell HP Labs Palo Alto, CA 94304, U.S.A. {magnus. . Classical non-adaptive control-theoretic methods are usually not adequate for two reasons. First, for many systems it is not even possible to use non-adaptive control as the system changes too much =-=[13, 15, 16]-=-. For example, the performance experience by a client of a three-tier e-commerce site depends on many things: what tier a request is served from, if it was served from the disk or the memory cache of   this paper. There are also a number of adaptive controllers that use ad-hoc estimation techniques that are not part of the estimation literature. For example, the control law and estimator of Yaksha =-=[13]-=- could probably benefit from our techniques. While RLS is the most widely used estimation technique for STR, there are others including extended leastsquares (ELS), total least-squares (TLS), the grad   </text>
<query_num> 22508 </query_num>
<text>   ue estimated by RLS. Allowing affine functions produces two new possible ways to form φ(k) referred to as RatioAffine and ThroughputAffine as shown in Table 2. A second trick used in image processing =-=[6]-=- is to add the sum of the throughputs to φ(k). When the throughput is changing, having this extra dimension gives a way for least squares to approximately scale the effects expected from T (k). It is  e employed have been successfully used in other fields. The trick of including scaling separately, as in the ThroughputSum method has been used for handling perspective projection in images and video =-=[6]-=-. To include a constant as input to least-squares has been used in many fields, for example in the off-line estimation of streaming media server performance [7], in order to get affine functions. The   </text>
</top>
<top>
<paper_num> 226 </paper_num>
<paper_title>   Three-Valued Abstraction for Continuous-Time Markov Chains.  </paper_title>
<paper_abstract>   Abstract. This paper proposes a novel abstraction technique for continuous-time Markov chains (CTMCs). Our technique fits within the realm of three-valued abstraction methods that have been used successfully for traditional model checking. The key idea is to apply abstraction on uniform CTMCs that are readily obtained from general CTMCs, and to abstract transition probabilities by intervals. It is shown that this provides a conservative abstraction for both true and false for a threevalued semantics of the branching-time logic CSL (Continuous Stochastic Logic). Experiments on an infinite-state CTMC indicate the feasibility of our abstraction technique. 1  </paper_abstract>
<query_num> 22601 </query_num>
<text>   . Prob l (Q) will be used as abbreviation for inf D∈Sched M Prob D (Q). We start with an algorithm for the approximation of probability bounds for timed reachability properties in uniform CTMDPs (see =-=[2]-=-). By Theorem 3, it suffices to consider extreme schedulers if one is interested in lower bounds. We interpret an ACTMC as a CTMDP, where each extreme distribution can be chosen by some action. From [ og(N) + N 3 ) + N) = O(N 5 ) for every iteration step. The following theorem, which states that the above algorithm yields an εaccurate approximation of reachability properties, follows directly from =-=[2]-=-. Theorem 4. For ACTMC M = (S,P l ,P u , Eunif, L), s ∈ S, B ⊆ S, t ∈ R&amp;gt;0 and error margin ε: Prob l (Reach≤t(s, B)) − ε ≤ q0(s) ≤ Prob l (Reach≤t(s, B)) Three-valued CSL-semantics. We define the sati   </text>
<query_num> 22602 </query_num>
<text>   a set of goal states (by avoiding bad states) within a maximal time span exceeds 7 8 . Existing abstraction techniques in this setting that have been applied in practice consider either bisimulation =-=[16]-=-, matrix bounding [6], simulation [24] or symmetry reduction [19]. (Due to the absence of nondeterminism, techniques such as partial-order reduction do not yield substantial reductions.) Despite the f   </text>
<query_num> 22603 </query_num>
<text>   ction Q : X × X → R≥0 let Q(Y, Y ′ ) = � y∈Y,y ′ ∈Y ′ Q(y, y′ ). The function Q(x, ·) is given by x ′ ↦→ Q(x, x ′ ) for all x ′ ∈ X. Furthermore a function f is called a distribution on X iff f : X → =-=[0, 1]-=- and f(X) := � x∈X f(x) = 1. Let AP be a fixed, finite set of atomic propositions and B2 = {⊥, ⊤} the two-valued truth domain. Definition 1 (DTMC). A DTMC is a tuple (S,P, L) with a finite non-empty s are “normalized” with respect to e. Thus, transitions occur with an average “pace” of e, uniform for all states. A CTMC is weak bisimilar to its uniformized CTMC [4]. Continuous Stochastic Logic. CSL =-=[1,3]-=- extends CTL by replacing existential and universal path quantification by a probability operator (as in PCTL) and by equipping the until-operator with a time bound (as in timed CTL): ϕ ::= true | a |  P ⊲⊳ p(Ψ)� = ⊤, iff Prob({σ ∈ Paths M s | �σ, Ψ� = ⊤}) ⊲⊳ p �σ, ϕ1U I ϕ2� = ⊤, iff ∃ t ∈ I : (�σ@t,ϕ2� = ⊤ ∧ ∀ t ′ ∈ [0, t) : �σ@t ′ , ϕ1� = ⊤) Table 1. Semantics of CSL where ⊲⊳ ∈ {&amp;lt;, ≤, ≥, &amp;gt;}, p ∈ =-=[0, 1]-=-, I = [0, t) | [0, t] | [0, ∞) for t ∈ R&amp;gt;0 and a ∈ AP. ϕ is a state-formula, whereas Ψ is a path-formula. State formulas are ranged over by ϕ, ψ, . . . and path formulas are ranged over by Ψ, Φ, . . .  8 10 v Fig.1. Abstraction for non-uniform CTMCs The functions P l (Ai, Aj, t) and P u (Ai, Aj, t) (considered as functions ranging over t) are in general not of the form p·(1 −e −E·t ) for fixed p ∈ =-=[0, 1]-=- and E &amp;gt; 0. Example 1. Consider the non-uniform CTMC M = ({s, u1, u2, u3, v},P, E, L) in Fig. 1 (left). We focus on the transition probabilities of the states u1, u2, u3 (indicated as labeled edges) a : Definition 4 (Abstract CTMC). An abstract CTMC (ACTMC for short) is a tuple M = (S,P l ,P u , Eunif, L) with a non-empty finite set of states S, transition probability functions P l ,P u : S × S ↦→ =-=[0, 1]-=- such that P l (s, S) ≤ 1 ≤ P u (s, S) componentwise for all s ∈ S. Eunif ∈ R&amp;gt;0 is the (global) exit rate for all states, and L : S × AP ↦→ B3 is a labeling function. 2 Recall that we consider the fra   </text>
<query_num> 22604 </query_num>
<text>   he approach is shown by considering abstractions of different granularity for an unbounded stochastic Petri net. Related work. Abstraction-refinement has been applied to reachability problems in MDPs =-=[10]-=-, partial-order reduction techniques using Peled’s ample-set method have been generalised to MDPs [13], abstract interpretation has been applied to MDPs [20], and various bisimulation equivalences and   </text>
<query_num> 22605 </query_num>
<text>   ical systems to systems biology. Model checking of CTMCs has been proved to extend and complement long-standing analysis techniques for Markov processes. Tools for stochastic Petri nets such as SMART =-=[8]-=- and GreatSPN [9], the PEPA Workbench [12] (a stochastic variant of the CWB), and Statemate [7] have adopted model checkers to analyse CTMCs, and temporal logics for Markov chains became prominent pro   </text>
<query_num> 22606 </query_num>
<text>   ime span exceeds 7 8 . Existing abstraction techniques in this setting that have been applied in practice consider either bisimulation [16], matrix bounding [6], simulation [24] or symmetry reduction =-=[19]-=-. (Due to the absence of nondeterminism, techniques such as partial-order reduction do not yield substantial reductions.) Despite the fact that fairly large reductions have recently been reported, mor  allow model aggregation prior to model checking, see e. g., [4, 23]. Recent techniques that have been proposed include abstraction of MDPs by two-player stochastic games [18], and symmetry reduction =-=[19]-=-. To our knowledge, threevalued abstraction of continuous-time stochastic models has not been considered.s2 Preliminaries Let X be a finite set. For Y, Y ′ ⊆ X and function Q : X × X → R≥0 let Q(Y, Y   </text>
<query_num> 22607 </query_num>
<text>   indefinite answer, the validity in the concrete model is unknown. The abstraction technique proposed here follows this three-valued approach. We consider abstractions for the branching-time logic CSL =-=[3]-=-, a real-time probabilistic variant of CTL. CSL is a powerful logic for expressing quantitative time-bounded constrained reachability properties such as the probability to reach a set of goal states ( are “normalized” with respect to e. Thus, transitions occur with an average “pace” of e, uniform for all states. A CTMC is weak bisimilar to its uniformized CTMC [4]. Continuous Stochastic Logic. CSL =-=[1,3]-=- extends CTL by replacing existential and universal path quantification by a probability operator (as in PCTL) and by equipping the until-operator with a time bound (as in timed CTL): ϕ ::= true | a |  s is omitted when s is clear from the context; the same applies to superscript M. Note that the probability measure of the set of infinite paths s0t0s1t1 . . . with � ∞ i=0 ti is converging, is zero =-=[3]-=-. The semantics of CSL is given in Table 1. ⊤ and ⊥ form a complete lattice such that ⊥ &amp;lt; ⊤ and meet ⊓ as well as complement · c are defined as usual. Measures of interest can now be expressed as CSL  elation allows us to reason about more concrete systems. For an ACTMC M, every scheduler D ∈ Sched M induces a probability space with a probability measure Prob D in the same manner as for CTMCs (see =-=[3]-=- for details). When interested in the infimum of probabilities of measurable sets with regard to all schedulers, it suffices to consider only extreme distributions. A scheduler which only chooses such   </text>
<query_num> 22608 </query_num>
<text>   n applied to reachability problems in MDPs [10], partial-order reduction techniques using Peled’s ample-set method have been generalised to MDPs [13], abstract interpretation has been applied to MDPs =-=[20]-=-, and various bisimulation equivalences and simulation preorders allow model aggregation prior to model checking, see e. g., [4, 23]. Recent techniques that have been proposed include abstraction of M   </text>
<query_num> 22609 </query_num>
<text>   nces and simulation preorders allow model aggregation prior to model checking, see e. g., [4, 23]. Recent techniques that have been proposed include abstraction of MDPs by two-player stochastic games =-=[18]-=-, and symmetry reduction [19]. To our knowledge, threevalued abstraction of continuous-time stochastic models has not been considered.s2 Preliminaries Let X be a finite set. For Y, Y ′ ⊆ X and functio   </text>
<query_num> 22610 </query_num>
<text>   pproximation. The system initially has no job to process, i.e. all three processors are available and the queue is empty. For m = 3, this is being formally described by the stochastic Petri net (SPN) =-=[5]-=- in Fig. 3(a). Numbers at edges denote that the corresponding transition consumes or produces the given number of tokens and can not be fired until there are enough token to consume. The semantics of   </text>
<query_num> 22611 </query_num>
<text>   ralised to MDPs [13], abstract interpretation has been applied to MDPs [20], and various bisimulation equivalences and simulation preorders allow model aggregation prior to model checking, see e. g., =-=[4, 23]-=-. Recent techniques that have been proposed include abstraction of MDPs by two-player stochastic games [18], and symmetry reduction [19]. To our knowledge, threevalued abstraction of continuous-time s   </text>
<query_num> 22612 </query_num>
<text>   ralised to MDPs [13], abstract interpretation has been applied to MDPs [20], and various bisimulation equivalences and simulation preorders allow model aggregation prior to model checking, see e. g., =-=[4, 23]-=-. Recent techniques that have been proposed include abstraction of MDPs by two-player stochastic games [18], and symmetry reduction [19]. To our knowledge, threevalued abstraction of continuous-time s . In unif (M) all rates of self-loops are “normalized” with respect to e. Thus, transitions occur with an average “pace” of e, uniform for all states. A CTMC is weak bisimilar to its uniformized CTMC =-=[4]-=-. Continuous Stochastic Logic. CSL [1,3] extends CTL by replacing existential and universal path quantification by a probability operator (as in PCTL) and by equipping the until-operator with a time b ff pl ≤ pu where pl, pu are the lower and upper bounds of time-abstract transition probabilities. Note that CTMC M and unif(M) are weak bisimilar, and as weak bisimulation preserves CSL equivalence 2 =-=[4]-=-, the shift to the uniformized CTMC is correct for CSL. Our abstract model now becomes: Definition 4 (Abstract CTMC). An abstract CTMC (ACTMC for short) is a tuple M = (S,P l ,P u , Eunif, L) with a n r(P l ,P u , S, s). To compare the behavior described by two ACTMCs, we introduce the notion of probabilistic simulation which is a variant of probabilistic simulation for CTMCs as it can be found in =-=[4]-=-. Definition 10 (Probabilistic simulation). Let M = (S,P l ,P u , Eunif, L) be an ACTMC. We call R ⊆ S × S a probabilistic simulation iff sRs ′ implies: 1. L(s ′ , a) �= ? ⇒ L(s ′ , a) = L(s, a) for a   </text>
<query_num> 22613 </query_num>
<text>   th m ′ additional processors then the requirement is not about m jobs anymore, but about m + m ′ . Note that CSL model-checking algorithms for quasi-birth-death processes have also been considered in =-=[21]-=-. Our abstraction technique, though, is not restricted to these (regular) infinite CTMCs. This paper presented a three-valued abstraction technique for CTMCs that is conservative for true and false re   </text>
<query_num> 22614 </query_num>
<text>   ystems biology. Model checking of CTMCs has been proved to extend and complement long-standing analysis techniques for Markov processes. Tools for stochastic Petri nets such as SMART [8] and GreatSPN =-=[9]-=-, the PEPA Workbench [12] (a stochastic variant of the CWB), and Statemate [7] have adopted model checkers to analyse CTMCs, and temporal logics for Markov chains became prominent property specificati   </text>
</top>
</query>
