Lecture Notes in Artificial Intelligence Edited by R Goebel, J Siekmann, and W Wahlster Subseries of Lecture Notes in Computer Science 6040 Stasinos Konstantopoulos Stavros Perantonis Vangelis Karkaletsis Constantine D Spyropoulos George Vouros (Eds.) Artificial Intelligence: Theories, Models and Applications 6th Hellenic Conference on AI, SETN 2010 Athens, Greece, May 4-7, 2010 Proceedings 13 Series Editors Randy Goebel, University of Alberta, Edmonton, Canada Jörg Siekmann, University of Saarland, Saarbrücken, Germany Wolfgang Wahlster, DFKI and University of Saarland, Saarbrücken, Germany Volume Editors Stasinos Konstantopoulos Stavros Perantonis Vangelis Karkaletsis Constantine D Spyropoulos Institute of Informatics and Telecommunications NCSR Demokritos Ag Paraskevi 15310, Athens, Greece E-mail: {konstant, sper, vangelis, costass}@iit.demokritos.gr George Vouros Department of Information and Communication Systems Engineering University of the Aegean Karlovassi, Samos 83200, Greece E-mail: georgev@aegean.gr Library of Congress Control Number: 2010925798 CR Subject Classification (1998): I.2, H.3, H.4, F.1, H.5, H.2.8 LNCS Sublibrary: SL – Artificial Intelligence ISSN ISBN-10 ISBN-13 0302-9743 3-642-12841-6 Springer Berlin Heidelberg New York 978-3-642-12841-7 Springer Berlin Heidelberg New York This work is subject to copyright All rights are reserved, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, re-use of illustrations, recitation, broadcasting, reproduction on microfilms or in any other way, and storage in data banks Duplication of this publication or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965, in its current version, and permission for use must always be obtained from Springer Violations are liable to prosecution under the German Copyright Law springer.com © Springer-Verlag Berlin Heidelberg 2010 Printed in Germany Typesetting: Camera-ready by author, data conversion by Scientific Publishing Services, Chennai, India Printed on acid-free paper 06/3180 Preface Artificial intelligence (AI) is a dynamic field that is constantly expanding into new application areas, discovering new research challenges, and facilitating the development of innovative products Today’s AI tools might not pass the Turing test, but they are invaluable aids in organizing and sorting the ever-increasing volume, complexity, and heterogeneity of knowledge available to us in our rapidly changing technological, economic, cultural, and social environment This volume aims at bringing to the reader all the latest developments in this exciting and challenging field, and contains papers selected for presentation at the 6th Hellenic Conference on Artificial Intelligence (SETN 2010), the official meeting of the Hellenic Society for Artificial Intelligence (EETN) SETN 2010 was organized by the Hellenic Society of Artificial Intelligence and the Institute of Informatics and Telecommunications, NCSR ‘Demokritos’ and took place in Athens during May 4–7 Previous conferences were held at the University of Piraeus (1996), at the Aristotle University of Thessaloniki (2002), at the University of the Aegean (Samos, 2004, and Syros, 2008), and jointly at the Foundation for Research and Technology–Hellas (FORTH) and the University of Crete (2006) SETN conferences play an important role in disseminating innovative and high-quality scientific results by AI researchers, attracting not only EETN members but also scientists advancing and applying AI in many and diverse domains and from various Greek and international institutes However, the most important aspect of SETN conferences is that they provide the context in which AI researchers meet and discuss their work, as well as an excellect opportunity for students to attend high-quality tutorials and get closer to AI results SETN 2010 continued this tradition of excellence, attracting submissions not only from Greece but also numerous European countries, Asia, and the Americas, which underwent a thorough reviewing process on the basis of their relevance to AI, originality, significance, technical soundness, and presentation The selection process was hard, with only 28 papers out of the 83 submitted being accepted as full papers and an additional 22 submissions accepted as short papers This proceedings volume also includes the abstracts of the invited talks presented at SETN 2010 by four internationally distinguished keynote speakers: Panos Constantopoulos, Michail Lagoudakis, Nikolaos Mavridis, and Demetri Terzopoulos As yet another indication of the growing international influence and importance of the conference, the EVENTS international workshop on event recognition and tracking chose to be co-located with SETN 2010 And, finally, SETN 2010 hosted the first ever RoboCup event organized in Greece, with the participation of two teams from abroad and one from Greece The Area Chairs and members of the SETN 2010 Programme Committee and the additional reviewers did an enormous amount of work and deserve the VI Preface special gratitude of all participants Our sincere thanks go to our sponsors for their generous financial support and to the Steering Committee for its assistance and support The conference operations were supported in an excellent way by the ConfMaster conference management system; many thanks to Thomas Preuss for his prompt responding with all questions and requests Special thanks go to to Konstantinos Stamatakis for the design of the conference poster and the design and maintenance of the conference website We also wish to thank the Organizing Committee and Be to Be Travel, the conference travel and organization agent, for implementing the conference schedule in a timely and flawless manner Last but not least, we also thank Alfred Hofmann, Anna Kramer, Leonie Kunz, and the Springer team for their continuous help and support March 2010 Stasinos Konstantopoulos Stavros Perantonis Vangelis Karkaletsis Constantine D Spyropoulos George Vouros Organization SETN 2010 was organized by the Institute of Informatics and Telecommunications, NCSR ‘Demokritos’, and EETN, the Hellenic Association of Artificial Intelligence Conference Chairs Constantine D Spyropoulos Vangelis Karkaletsis George Vouros NCSR ‘Demokritos’, Greece NCSR ‘Demokritos’, Greece University of the Aegean, Greece Steering Committee Grigoris Antoniou John Darzentas Nikos Fakotakis Themistoklis Panayiotopoulos Ioannis Vlahavas FORTH and University of Crete University of the Aegean University of Patras University of Piraeus Aristotle University Organizing Committee Alexandros Artikis Vassilis Gatos Pythagoras Karampiperis Anastasios Kesidis Anastasia Krithara Georgios Petasis Sergios Petridis Ioannis Pratikakis Konstantinos Stamatakis Dimitrios Vogiatzis Programme Committee Chairs Stasinos Konstantopoulos Stavros Perantonis NCSR ‘Demokritos’ NCSR ‘Demokritos’ Programme Committee Area Chairs Ion Androutsopoulos Nick Bassiliades Athens University of Economics and Business Aristotle University of Thessaloniki VIII Organization Ioannis Hatzilygeroudis Ilias Maglogiannis Georgios Paliouras Ioannis Refanidis Efstathios Stamatatos Kostas Stergiou Panos Trahanias University of Patras University of Central Greece NCSR ‘Demokritos’ University of Macedonia University of the Aegean University of the Aegean FORTH and University of Crete Programme Committee Members Dimitris Apostolou Argyris Arnellos Alexander Artikis Grigorios Beligiannis Basilis Boutsinas Theodore Dalamagas Yannis Dimopoulos Christos Douligeris George Dounias Eleni Galiotou Todor Ganchev Vassilis Gatos Efstratios Georgopoulos Manolis Gergatsoulis Nikos Hatziargyriou Katerina Kabassi Dimitris Kalles Kostas Karatzas Dimitrios Karras Petros Kefalas Stefanos Kollias Yiannis Kompatsaris Dimitris Kosmopoulos Constantine Kotropoulos Manolis Koubarakis Konstantinos Koutroumbas Michail Lagoudakis Aristidis Likas George Magoulas Filia Makedon Manolis Maragoudakis Vassilis Moustakis Christos Papatheodorou Pavlos Peppas Sergios Petridis University of Piraeus University of the Aegean NCSR ‘Demokritos’ University of Ioannina University of Patras IMIS Institute/‘Athena’ Research Center University of Cyprus University of Piraeus University of the Aegean TEI Athens University of Patras NCSR ‘Demokritos’ TEI Kalamata Ionian University National Technical University of Athens TEI Ionian Hellenic Open University Aristotle University of Thessaloniki TEI Chalkis City Liberal Studies National Technical University of Athens CERTH NCSR ‘Demokritos’ Aristotle University of Thessaloniki National and Kapodistrian University of Athens National Observatory of Athens Technical University of Crete University of Ioannina Birkbeck College, University of London (UK) University of Texas at Arlington (USA) University of the Aegean Technical University of Crete Ionian University University of Patras NCSR ‘Demokritos’ Organization Stelios Piperidis Vassilis Plagianakos Dimitris Plexousakis George Potamias Ioannis Pratikakis Jim Prentzas Ilias Sakellariou Kyriakos Sgarbas John Soldatos Panagiotis Stamatopoulos Giorgos Stoilos Ioannis Tsamardinos George Tsichrintzis Nikos Vasilas Michalis Vazirgia Maria Virvou Spyros Vosinakis Dimitris Vrakas ILSP-Athena RC University of Central Greece FORTH and University of Crete FORTH NCSR ‘Demokritos’ Democritus University of Thrace University of Macedonia University of Patras AIT National and Kapodistrian University of Athens Oxford University (UK) University of Crete and FORTH University of Piraeus TEI Athens Athens University of Economics and Business University of Piraeus University of the Aegean Aristotle University of Thessaloniki Additional Reviewers Charalampos Doukas Anastasios Doulamis Giorgos Flouris Theodoros Giannakopoulos Katia Kermanidis Otilia Kocsis Eleytherios Koumakis Anastasia Krithara Pavlos Moraitis Nikolaos Pothitos Spyros Raptis Vassiliki Rentoumi Evangelos Sakkopoulos Themos Stafylakis Sophia Stamou Andreas Symeonidis Vassilios Vassiliadis Dimitrios Vogiatzis IX University of the Aegean Technical University of Crete FORTH NCSR ‘Demokritos’ Ionian University University of Patras Technical University of Crete NCSR ‘Demokritos’ Paris Descartes University (France) National and Kapodistrian University of Athens ILSP-Athena RC NCSR ‘Demokritos’ University of Patras ILSP-Athena RC University of Patras Aristotle University of Thessaloniki University of the Aegean NCSR ‘Demokritos’ Table of Contents Invited Talks Digital Curation and Digital Cultural Memory Panos Constantopoulos RoboCup: A Challenge Problem for Artificial Intelligence Michail G Lagoudakis Robots, Natural Language, Social Networks, and Art Nikolaos Mavridis Artificial Life Simulation of Humans and Lower Animals: From Biomechanics to Intelligence Demetri Terzopoulos Full Papers Prediction of Aircraft Aluminum Alloys Tensile Mechanical Properties Degradation Using Support Vector Machines Nikolaos Ampazis and Nikolaos D Alexopoulos Mutual Information Measures for Subclass Error-Correcting Output Codes Classification Nikolaos Arvanitopoulos, Dimitrios Bouzas, and Anastasios Tefas 19 Conflict Directed Variable Selection Strategies for Constraint Satisfaction Problems Thanasis Balafoutis and Kostas Stergiou 29 A Feasibility Study on Low Level Techniques for Improving Parsing Accuracy for Spanish Using Maltparser Miguel Ballesteros, Jes´ us Herrera, Virginia Francisco, and Pablo Gerv´ as A Hybrid Ant Colony Optimization Algorithm for Solving the Ring Arc-Loading Problem Anabela Moreira Bernardino, Eug´enia Moreira Bernardino, Juan Manuel S´ anchez-P´erez, Juan Antonio G´ omez-Pulido, and Miguel Angel Vega-Rodr´ıguez Trends and Issues in Description Logics Frameworks for Image Interpretation Stamatia Dasiopoulou and Ioannis Kompatsiaris 39 49 61 410 N Pothitos and P Stamatopoulos Conclusions and Further Work In this work, it has been shown that we can achieve a much better lower memory bound for the representation of a domain, than the actual memory consumption of Constraint Programming systems An improved way of storing a domain, through new data structures and algorithms was proposed This methodology naturally applies to various problems with wide domains, e.g Bioinformatics problems that come along with large genome databases In future, hybrid data structures can contribute towards the same direction For example, variable size bit vectors could be integrated into binary tree nodes Everything should be designed to be as much generic as possible, in order to exploit at any case the plethora of known algorithms for generic CSPs Acknowledgements This work is funded by the Special Account Research Grants of the National and Kapodistrian University of Athens, in the context of the project ‘C++ Libraries for Constraint Programming’ (project no 70/4/4639) We would also like to thank Stavros Anagnostopoulos, a Bioinformatics expert, for his valuable help in our understanding of various biological problems and data References Codognet, P., Diaz, D.: Compiling constraints in clp(FD) The Journal of Logic Programming 27(3), 185–226 (1996) ECLi PSe constraint programming system (2008), http://eclipse-clp.org Gent, I., Jefferson, C., Miguel, I., Nightingale, P.: Data structures for generalised arc consistency for extensional constraints In: AAAI 2007: 22nd National Conference on Artificial Intelligence, pp 191–197 AAAI Press, Menlo Park (2007) ILOG S.A.: ILOG Solver 4.4: User’s Manual (1999) Pothitos, N.: Naxos Solver (2009), http://www.di.uoa.gr/~ pothitos/naxos Sabin, D., Freuder, E.C.: Contradicting conventional wisdom in constraint satisfaction In: Borning, A (ed.) PPCP 1994 LNCS, vol 874, pp 125–129 Springer, Heidelberg (1994) Schulte, C., Carlsson, M.: Finite domain constraint programming systems In: Handbook of Constraint Programming, pp 495–526 Elsevier Science, Amsterdam (2006) Th´ebault, P.: MilPat’s user manual (2006), http://carlit.toulouse.inra.fr/MilPat Th´ebault, P., de Givry, S., Schiex, T., Gaspin, C.: Searching RNA motifs and their intermolecular contacts with constraint networks Bioinformatics 22(17), 2074– 2080 (2006) 10 Watson, J., Baker, T., Bell, S., Gann, A., Levine, M., Losick, R.: Molecular Biology of the Gene, ch 6, 5th edn Pearson/Benjamin Cummings (2004) 11 Zytnicki, M., Gaspin, C., Schiex, T.: A new local consistency for weighted CSP dedicated to long domains In: SAC 2006: Proceedings of the 2006 ACM symposium on Applied computing, pp 394–398 ACM, New York (2006) A Collaborative System for Sentiment Analysis Vassiliki Rentoumi1,2 , Stefanos Petrakis3, Vangelis Karkaletsis1, Manfred Klenner3 , and George A Vouros2 Inst of Informatics and Telecommunications, NCSR “Demokritos”, Greece University of the Aegean, Artificial Intelligence Laboratory, Samos, Greece Institute of Computational Linguistics, University of Zurich, Switzerland vrentoumi@iit.demokritos.gr, petrakis@cl.uzh.ch, vangelis@iit.demokritos.gr, klenner@cl.uzh.ch, georgev@aegean.gr Abstract In the past we have witnessed our machine learning method for sentiment analysis coping well with figurative language, but determining with uncertainty the polarity of mildly figurative cases We have shown that for these uncertain cases, a rule-based system should be consulted We evaluate this collaborative approach on the ”Rotten Tomatoes” movie reviews dataset and compare it with other state-of-the-art methods, providing further evidence in favor of this approach Introduction In the past we have shown that figurative language conveys sentiment that can be efficiently detected by FigML[2], a machine learning (ML) approach trained on corpora manually annotated with strong figurative expressions1 FigML was able to detect the polarity of sentences bearing highly figurative expressions, where disambiguation is considered mandatory, such as: (a)“credibility sinks into a mire of sentiments” On the other hand, there exist cases for which FigML provided a classification decision based on a narrow margin between negative and positive polarity orientation, often resulting in erroneous polarity evaluation It was observed that such cases bear mild figurativeness, which according to [4] are synchronically as literal as their primary sense, as a result of standardized usage, like: (b) “this 10th film in the series looks and feels tired” Here, fatigue as a property of inanimate or abstract objects, although highly figurative, presents an obvious negative connotation, due to standardized usage of this particular sense, therefore sentiment disambiguation is not necessary Such regular cases could be more efficiently treated by a rule-based system such as PolArt[1] In fact, in this paper we extend the work presented in [8] where we have indeed shown that cases of mild figurative language are better treated by PolArt, while cases of strong figurative language are better handled by FigML In [8], a novel collaborative system for sentiment analysis was proposed and managed Subsets from the AffectiveText corpus (SemEval’07) and the MovieReviews sentence polarity dataset v1.0, annotated with metaphors and expanded senses: http://www.iit.demokritos.gr/~ vrentoumi/corpus.zip S Konstantopoulos et al (Eds.): SETN 2010, LNAI 6040, pp 411–416, 2010 c Springer-Verlag Berlin Heidelberg 2010 412 V Rentoumi et al to outperform its two subcomponents, FigML and PolArt, tested on the AffectiveText corpus Here, we try to verify the validity of this approach on a larger corpus and of a differenet domain and style In addition and most importantly, another dimension of complementarity between a machine learning method and a rule-based one is explored: the rule-based approach handles the literal cases and the - already introduced - collaborative method treats the cases of figurative language Results show that integrating a machine learning approach with a finer-grained linguistically-based one leads to a superior, best-of-breed system Methodology Description The proposed collaborative method involves four consecutive steps: (a)Word sense disambiguation(WSD): We chose an algorithm which takes as input a sentence and a relatedness measure[6] The algorithm supports several WordNet based similarity measures among which Gloss Vector (GV)[6] performs best for non-literal verbs and nouns [5] Integrating GV in the WSD step is detailed in [2] (b)Sense level polarity assignment(SLPA): We adopted a machine learning approach which exploits graphs based on character n-grams[7] We compute models of positive and negative polarity from examples of positive and negative words and definitions provided by a enriched version of the Subjectivity Lexicon2,3 The polarity class of each test sense, is determined by computing its similarity with the models as detailed in [2] (c)HMMs training: HMMs serve two purposes Computing the threshold which divides the sentences in marginal/non-marginal and judging the polarity(positive/ negative) of non-marginal sentences We train one HMM model for each polarity class The format of the training instances is detailed in [2] For computing the threshold, the training data are also used as a testing set Each test instance is tested against both models and the output is a pair of log probabilities of a test instance to belong to either the negative or the positive class For each polarity class we compute the absolute difference of the log probabilities We then sort these differences in ascending order and calculate the first Quartile (Q1) which separates the lower 25% of the sample population from the rest of the data We set this to be the threshold and we apply it to the test instances Marginal cases are the ones for which the absolute difference of log probability is below that threshold In our experiments we use a 10-fold cross validation approach to evaluate our results (d) Sentence-level polarity detection: The polarity of each sentence is determined by HMMs [2] for non-marginal cases and by PolArt[1] for marginal http://www.cs.pitt.edu/mpqa/ For each positive or negative word entry contained in the Subjectivity Lexicon, we extracted the corresponding set of senses from WordNet, represented by their synsets and gloss examples; in this way we tried to reach a greater degree of consistency between the test and the training set A Collaborative System for Sentiment Analysis 413 ones PolArt employs compositional rules and obtains word-level polarities from a polarity lexicon, as described in detail in [1] The Collaborative system’s total performance is then given by adding up the performances of FigML and PolArt Experimental Setup 3.1 Resources We ran our experiments on the MovieReviews corpus4 This corpus was split into different subsets according to our experimental setup in two different ways: – Expanded Senses/Metaphors/Whole: The corpus was enhriched with manually-added annotations for metaphors and expanded senses inside sentences We produced an expanded senses dataset and a metaphorical expressions one Furthermore, we treated the entire corpus as a third dataset, ignoring the aforementioned annotations The produced datasets are: • Expanded senses: 867 sentences, 450 negative and 417 positive ones • Metaphors: 996 sentences, 505 negative and 491 positive ones • Whole: 10649 sentences, 5326 negative and 5323 positive ones – Literal/Non-literal: We group all figurative sentences (metaphors/expanded senses) as the non-literal set The rest of the sentences we call the literal set • Non-literal: 1862 sentences5 , 954 negative and 908 positive ones • Literal: 8787 sentences, 4372 negative and 4415 positive ones We run numerous variations of PolArt, modifying each time the polarity lexicon it consults: – SL+: This is the subjectivity lexicon6 with manually added valence operators – Merged: The FigML system produces automatically sense-level polarity lexica (AutSPs), one for each dataset or subset For the non-literal, metaphors and expanded senses, these lexica target non-literal expressions, metaphors and expanded senses accordingly For the entire MovieReviews dataset (Whole), all word senses are targeted Various Merged lexica are produced by combining and merging the SL+ lexicon with each of the AutSPs We used the sentence polarity dataset v1.0 from http://www.cs.cornell.edu/People/pabo/movie-review-data/ One sentence belonged to both the metaphors and expanded senses subsets, and was included only once here http://www.cs.pitt.edu/mpqa/ 414 3.2 V Rentoumi et al Collaborative Method Tested on MovieReviews Dataset We tested our Collaborative method originally presented and evaluated in [8], with the extended MovieReviews corpus, in order to test its validity Table presents scores for each polarity class, for both variants of our method, the CollaborativeSL+ (using the SL lexicon) and CollaborativeMerged (using the Merged Lexica), across all three datasets For the majority of cases, CollaborativeSL+ has better performance than CollaborativeMerged Comparing the performance of CollaborativeSL+ for the MovieReviews with that of CollaborativeSL+ for the AffectiveText corpus [8], for the Whole corpus (f-measure: neg: 0.62, pos: 0.59), we noticed that the performance remains approximately the same This is evidence that the method is consistent across different datasets Table MovieReviews: Performance scores for full system runs recall Whole precision f-measure recall Met precision f-measure recall Exp precision f-measure 3.3 CollaborativeSL+ neg pos 0.682 0.537 0.596 0.628 0.636 0.579 0.724 0.735 0.737 0.722 0.731 0.728 0.640 0.623 0.647 0.616 0.643 0.619 CollaborativeMerged neg pos 0.656 0.536 0.586 0.609 0.619 0.570 0.697 0.704 0.708 0.693 0.702 0.699 0.642 0.623 0.648 0.617 0.645 0.620 The Collaborative Approach Treats Non-literal Cases as a Whole: Complementarity on the Literal/Non-literal Axis We have so far shown that our Collaborative method is performing quite well on the expanded senses and metaphors datasets Although we consider them as distinct language phenomena, they both belong to the sphere of figurative connotation To support this we tested our claim collectively, across non-literal expressions in general, by merging these two datasets into one labelled nonliterals As a baseline system for assessing the performance of the collaborative method we use a clean version of PolArt (i.e without added valence shifters) In Table 2, we compare BaselinePolart with CollaborativeSL+ (using the SL lexicon) and CollaborativeMerged (using the Merged Lexica), tested upon the non-literals dataset We observe that our proposed method outperforms the baseline and proves quite capable of treating non-literal cases collectively By assembling the non-literals into one dataset and treating it with our collaborative method we set aside its complementary dataset of literals Since our method is more inclined to treat figurative language, we not expect that it should treat literal cases optimally, or at least as efficiently as a system that is more inclined to treat literal language Therefore, assigning the literals to PolArt and the nonliterals to Collaborative, would provide a more sane system architecture and result in better performance for the entire MovieReviews dataset In Table we present the performance of both variants of the new system architecture (PolartwithCollaborativeSL+, PolartwithCollaborativeMerged) In A Collaborative System for Sentiment Analysis 415 Table MovieReviews: Performance scores for the non-literals subset CollaborativeSL+ neg pos recall 0.710 0.646 Nonliterals precision 0.678 0.680 f-measure 0.694 0.662 CollaborativeMerged neg pos 0.681 0.644 0.668 0.658 0.674 0.651 BaselinePolart neg pos 0.614 0.667 0.659 0.622 0.636 0.644 Table MovieReviews: Performance scores for full system runs recall Literals/nonliterals precision f-measure Whole recall precision f-measure PolartwithCollaborativeSL+ neg pos 0.608 0.659 0.641 0.627 0.624 0.642 CollaborativeSL+ neg pos 0.682 0.537 0.596 0.628 0.636 0.579 PolartwithCollaborativeMerged neg pos 0.603 0.659 0.638 0.624 0.620 0.641 CollaborativeMerged neg pos 0.656 0.536 0.586 0.609 0.619 0.570 both versions pure PolArt treats literal cases, while CollaborativeSL+ and CollaborativeMerged treat non literals cases This new architecture is compared to the one concerning the treatment of the whole corpus (Whole) by both variants of the proposed method (CollaborativeSL+, CollaborativeMerged) It is observed that the performance of this modified system is better for the majority of cases This fact leads us to the conclusion that a system which treats sentiments in a more language-sensitive way, can exhibit improved performance We further compared our system with a state-of-the-art system by Andreevskaia and Bergler[3], tested on the MovieReviews corpus Their system employs a Naive Bayes Classifier for polarity classification of sentences, trained with unigrams, bigrams or trigrams derived from the same corpus This state-of-the-art system’s accuracy was reported to be 0.774, 0.739 and 0.654 for unigrams, bigrams and trigrams Our two alternative system architectures, CollaborativeSL+ and PolartwithCollaborativeSL+, scored 0.609 and 0.633 The performances of both our alternatives are clearly lower than the state-ofthe-art system’s when the latter is trained with unigrams or bigrams, but they get closer when it is trained with trigrams The main point is that the CollaborativeSL+ method performs quite well even for the case of a corpus containing mainly literal language We expect CollaborativeSL+ to perform optimally when applied on a corpus consisting mainly of non-literal language It is also worth noting that since PolArt deals with the majority of cases it is bound to heavily affect the overall system performance Additionally PolArt’s dependency on its underlying resources and especially the prior polarity lexicon is also a crucial performance factor Thus, the observed moderate performance of the system can be attributed to the moderate PolArt’s performance, probably due to the incompatibility of the Subjectivity Lexicon with the idiosyncratic/colloquial language of the Movie Reviews corpus 416 V Rentoumi et al All in all, the overall performance is still quite satisfactory Consequently, if we provide PolArt with a more appropriate lexicon, we expect a further boost Conclusions and Future Work In this paper we further extend and examine the idea of a sentiment analysis method which exploits complementarily two language specific subsystems, a rule-based (PolArt) for the mild figurative, and a machine learning system (FigML) for the strong figurative language phenomena[8] By further examining the validity of such an approach in a larger (and of different domain) corpus (Movie Reviews corpus), in which strong figurative language co-exists with mild figurative language, we observed that this Collaborative method is consistent We also explored another dimension of complementarity concerning literal/ non-literal cases of language, where PolArt is treating the literal cases and the Collaborative method the non-literal cases We get empirical support from the performance obtained that utilizing the special virtues of the participating subsystems can be a corner-stone in the design and performance of the resulting system We will test the collaborative method on a more extensive corpus bearing figurative language We intend to dynamically produce sense-level polarity lexica exploiting additional machine learning approaches (e.g SVMs) References Klenner, M., Petrakis, S., Fahrni, A.: Robust compositional polarity classification In: Recent Advances in Natural Language Processing (RANLP), Borovets, Bulgaria (2009) Rentoumi, V., Giannakopoulos, G., Karkaletsis, V., Vouros, G.: Sentiment analysis of figurative language using a word sense disambiguation approach In: Recent Advances in Natural Language Processing (RANLP), Borovets, Bulgaria (2009) Andreevskaia, A., Bergler, S.: When specialists and generalists work together: overcoming domain dependence in sentiment tagging In: Proceedings of ACL 2008: HLT, pp 290–298 (2008) Cruse, D.A.: Meaning in language Oxford University Press, Oxford (2000) Rentoumi, V., Karkaletsis, V., Vouros, G., Mozer, A.: Sentiment Analysis Exploring Metaphorical and Idiomatic Senses: A Word Sense Disambiguation Approach In: International Workshop on Computational Aspects of Affectual and Emotional Interaction, CAFFEi 2008 (2008) Pedersen, T., Banerjee, S., Patwardhan, S.: Maximizing Semantic Relatedness to Perform Word Sense Disambiguation Supercomputing Institute Research Report UMSI, vol 25 (2005) Giannakopoulos, G., Karkaletsis, V., Vouros, G., Stamatopoulos, P.: Summarization system evaluation revisited: N-gram graphs ACM Transactions on Speech and Language Processing (TSLP) (2008) Rentoumi, V., Petrakis, S., Klenner, M., Vouros, G., Karkaletsis, V.: A Hybrid System for Sentiment Analysis To appear in LREC 2010 (2010) Minimax Search and Reinforcement Learning for Adversarial Tetris Maria Rovatsou and Michail G Lagoudakis Intelligent Systems Laboratory Department of Electronic and Computer Engineering Technical University of Crete Chania 73100, Crete, Greece mariarovatsou@gmail.com, lagoudakis@intelligence.tuc.gr Abstract Game playing has always been considered an intellectual activity requiring a good level of intelligence This paper focuses on Adversarial Tetris, a variation of the well-known Tetris game, introduced at the 3rd International Reinforcement Learning Competition in 2009 In Adversarial Tetris the mission of the player to complete as many lines as possible is actively hindered by an unknown adversary who selects the falling tetraminoes in ways that make the game harder for the player In addition, there are boards of different sizes and learning ability is tested over a variety of boards and adversaries This paper describes the design and implementation of an agent capable of learning to improve his strategy against any adversary and any board size The agent employs MiniMax search enhanced with Alpha-Beta pruning for looking ahead within the game tree and a variation of the Least-Squares Temporal Difference Learning (LSTD) algorithm for learning an appropriate state evaluation function over a small set of features The learned strategies exhibit good performance over a wide range of boards and adversaries Introduction Skillful game playing has always been considered a token of intelligence, consequently Artificial Intelligence and Machine Learning exploit games in order to exhibit intelligent performance A game that has become a benchmark, exactly because it involves a great deal of complexity along with very simple playing rules, is the game of Tetris It consists of a grid board in which four-block tiles, chosen randomly, fall from the top and the goal of the player is to place them so that they form complete lines, which are eliminated from the board, lowering all blocks above The game is over when a tile reaches the top of the board The fact that the rules are simple should not give the impression that the task is simple There are about 40 possible actions available to the player for placing a tile and about 1064 possible states that these actions could lead to These magnitudes are hard to deal with for any kind of player (human or computer) Adversarial Tetris is a variation of Tetris that introduces adversity in the game, making it even more demanding and intriguing; an unknown adversary tries to S Konstantopoulos et al (Eds.): SETN 2010, LNAI 6040, pp 417–422, 2010 c Springer-Verlag Berlin Heidelberg 2010 418 M Rovatsou and M.G Lagoudakis hinder the goals of the player by actively choosing pieces that augment the difficulty of line completion and by even “leaving out” a tile from the entire game, if that suits his adversarial goals This paper presents our approach to designing a learning player for Adversarial Tetris Our player employs MiniMax search to produce a strategy that accounts for any adversary and reinforcement learning to learn an appropriate state evaluation function Our agent exhibits improving performance over an increasing number of learning games Tetris and Adversarial Tetris Tetris is a video game created in 1984 by Alexey Pajitnov, a Russian computer engineer The game is played on a 10 × 20 board using seven kinds of simple tiles, called tetraminoes All tetraminoes are composed of four colored blocks (minoes) forming a total of seven different shapes The rules of the game are very simple The tiles are falling down one-by-one from the top of the board and the user rotates and moves them until they rest on top of existing tiles in the board The goal is to place the tiles so that lines are completed without gaps; completed lines are eliminated, lowering all the remaining blocks above The game ends when a resting tile reaches the top of the board Tetris is a very demanding and intriguing game It has been proved [1] that finding a strategy that maximizes the number of completed rows, or maximizes the number of the lines eliminated simultaneously, or minimizes the board height, or maximizes the number of tetraminoes placed in the board before the game ends is an N P-hard problem; even approximating an optimal strategy is N P-hard This inherent difficulty is one of the reasons this game is widely used as a benchmark domain Tetris is naturally formulated as a Markovian Decision Process (MDP) [2] The state consists of the current board and the current falling tile and the actions are the approximately 40 placement actions for the falling tile The transition model is fairly simple; there are seven equiprobable possible next states, since the next board is uniquely determined and the next falling piece is chosen uniformly The reward function gives positive numerical values for completed lines and the goal is to find a policy that maximizes the long-term cumulative reward The recent Reinforcement Learning (RL) Competition [3] introduced a variation of Tetris, called Adversarial Tetris, whereby the falling tile generator is replaced by an active opponent The tiles are now chosen purposefully to hinder the goals of the player (completion or lines) The main difference in the MDP model of Adversarial Tetris is the fact that the distribution of falling tiles is non-stationary and the dimension of the board varies in height and width Furthermore, the state is produced like the frames of the video game, as it includes the current position and rotation of the falling tile in addition to the configuration of the board and the player can move/rotate the falling tile at each frame The RL Competition offers a generalized MDP model for Adversarial Tetris which is fully specified by four parameters (the height and width of the board and the adversity and type of the opponent) For the needs of the competition 20 instances of this model were specified with widths ranging from to 11, heights ranging from 16 to 25, and different types of opponents and opponent’s adversity Minimax Search and Reinforcement Learning for Adversarial Tetris 419 Designing a Learning Player for Adversarial Tetris Player Actions In Adversarial Tetris the tile is falling one step downwards every time the agent chooses one of the low-level actions: move the tile left or right, rotate it clockwise or counterclockwise, drop it, and nothing Clearly, there exist various alternative sequences of these actions to achieve the same placement of the tile; this freedom yields repeated board configurations that lead to an unnecessary growth of the game tree Also, playing at the level of the lowlevel actions ruins the idea of a two-player alternating game, as the opponent’s turn appears only once after several turns of the player Lastly, the branching factor of would lead to an intractable game tree, even before the falling tile reaches a resting position in the board These observations led us to consider an abstraction of the player’s moves, namely high-level actions that bring the tile from the top of the board directly to its resting position using a minimal sequence of low-level actions planned using a simple look-ahead search The game tree now contains alternating plies of the player’s and the opponent’s moves, as a true twoplayer alternating game; all unnecessary intermediate nodes of player’s low-level actions are eliminated The actual number of high-level actions available in each state depends on the width of the board and the number of distinct rotations of the tile itself, but they will be at most × wb, where wb is the width of the board (wb columns and rotations) Similarly, the opponent chooses not only the next falling tile, but also its initial rotation, which means that he has as many as × = 28 actions However, not all these actions are needed to represent the opponent’s moves, since in the majority of cases the player can use low-level actions to rotate the tile at will Thus, the initial rotation can be neglected to reduce the branching factor at opponent nodes from 28 to just In summary, there are about 4wb choices for the player and choices for the opponent Game Tree The MiniMax objective criterion is commonly used in two-player zero-sum games, where any gain on one side (Max) is equal to the loss on the other side (Min) The Max player is trying to select its best action over all possible Min choices in the next and future turns In Adversarial Tetris, our player is taken as Max, since he is trying to increase his score, whereas the adversarial opponent is taken as Min, since he is trying to decrease our player’s score We adopted this criterion because it is independent of the opponent (it produces the same strategy irrespectively of the competence of the opponent) and protects against tricky opponents who may initially bluff Its drawback is that it does not take risks and therefore it cannot exploit weak opponents The implication is that our agent should be able to play Tetris well against any friendly, adversarial, or no-care opponent The MiniMax game tree represents all possible paths of action sequences of the two players playing in alternating turns Our player forms a new game tree from the current state, whenever it is his turn to play, to derive his best action choice Clearly, our player cannot generate the entire tree, therefore expansion continues up to a cut-off depth The utility of the nodes at the cut-off depth is estimated by an evaluation function described below MiniMax is aided by Alpha-Beta Pruning, which prunes away nodes and subtrees not contributing to the root value and to the final decision 420 M Rovatsou and M.G Lagoudakis Evaluation Function The evaluation of a game state s whether in favor or against our agent is done by an evaluation function V (s), which also implicitly determines the agent’s policy Given the huge state space of the game, such an evaluation function cannot be computed or stored explicitly, so it must be approximated We are using a linear approximation architecture formed by a vector of k features φ(s) and a vector of k weights w The approximate value is k computed as the weighted sum of the features, V (s) = i=1 φi (s)wi = φ(s) w We have issued two possible sets of features which will eventually lead to two different agents The first set includes features for characterizing the board: a constant term, the maximum height, the mean height, the sum of absolute column differences in height, the total number of empty cells below placed tiles (holes), and the total number of empty cells above placed tiles up to the maximum height (gaps) The second set uses a separate block of these features for each one of the tiles of Tetris, giving a total of 42 features This is proposed because with the first set the agent can learn which boards and actions are good for him, but cannot associate them to the falling tiles that these actions manipulate The same action on different tiles, even if the board is unchanged, may have a totally different effect; ignoring the type of tile leads to less effective behavior This second set of features alleviates this problem by simply weighing the base features differently for different falling tiles Note that only one block of size is active in any state, the one corresponding to the current falling tile Learning In order to learn a good set of weights for our evaluation function we applied a variation of the Least-Squares Temporal Difference Learning (LSTD) algorithm [4] The need for modifying the original LSTD algorithm stems from the fact that the underlying agent policy is determined through the values given to states by our evaluation function, which are propagated to the root; if these values change, so does the policy, therefore it is important to discard old data and use only the recent ones for learning To this end, we used the technique of exponential windowing, whereby the weights are updated in regular intervals called epochs; each epoch may last for several decision steps During an epoch the underlying value function and policy remain unchanged for collecting correct evaluation data and only at the completion of the epoch are the weights updated In the next epoch, data from the previous epoch are discounted by a parameter μ Therefore, past data are not completely eliminated, but are weighted less and less as they become older and older Their influence depends on the value of μ which ranges between (no influence) to (full influence) A value of leads to singularity problems due to the shortage of samples within a single epoch, however a value around 0.95 offers a good balance between recent and old data with exponentially decayed weights A full description of the modified algorithm is given in Algorithm (t indicates the epoch number) In order to accommodate a wider range of objectives we used a rewarding scheme that encourages line completion (positive reward), but discourages loss of a game (negative reward) We balanced these two objectives by giving a reward of +1 for each completed line and a penalty of −10 for each game lost We set the discount factor to (γ = 1) since rewards/penalties not loose value as time advances Minimax Search and Reinforcement Learning for Adversarial Tetris 421 Algorithm LSTD with Exponential Windowing (wt , At , bt ) = LSTD-EW(k, φ, γ, t, Dt , wt−1 , At−1 , bt−1 , μ) if t == then At ← 0; bt ← else At ← μAt−1 ; bt ← μbt−1 end if for all samples (s, r, s ) ∈ Dt At ← At + φ(s) φ(s) − γφ(s ) ; bt ← bt + φ(s)r end for −1 wt ← (At ) bt return wt , At , bt Related Work There is a lot of work on Tetris in recent years Tsitsiklis and Van Roy applied approximate value iteration, whereas Bertsekas and Ioffe tried policy iteration, and Kakade used the natural policy gradient method Later, Lagoudakis et al applied a least-squares approach to learning an approximate value function, while Ramon and Driessens modeled Tetris as a relational reinforcement learning problem and applied a regression technique using Gaussian processes to predict the value function Also, de Farias and Van Roy used the technique of randomized constraint sampling in order to approximate the optimal cost function Finally, Szita and Lă orincz applied the noisy cross-entropy method In the 2008 RL Competition, the approach of Thiery [5] based on λ-Policy Iteration outperformed all previous work at the time There is only unpublished work on Adversarial Tetris from the 2009 RL Competition, where only two teams participated The winning team from Rutgers University applied look-ahead tree search and the opponent in each MDP was modeled as a fixed probability distribution over falling tiles, which was learned using the cross entropy method Results and Conclusion Our learning experiments are conducted over a period of 400 epochs of 8,000 game steps each, giving a total of 3,200,000 samples The weights are updated at the end of each learning epoch Learning is conducted only on MDP #1 (out of the 20 MDPs of the RL Competition) which has board dimensions that are closer to the board dimensions of the original Tetris Learning takes place only at the root of the tree in each move, as learning at the internal nodes leads to a great degree of repetition biasing the learned evaluation function Agent (6 features) learns by backing up values from depth (or any other odd depth) This set of features ignores the choice of Min and thus it would be meaningless to expand the tree one more level deeper at Min nodes, which are found at odd depths The second agent (42 features) learns by backing up values from depth (or any other even depth) This set of basis functions takes the action choice of the Min explicitly into account and thus it makes sense to cut-off the search at Max nodes, which are found at even depths The same cut-offs apply to testing M Rovatsou and M.G Lagoudakis 600 L Change in Weights 400 30 Average Lines per Game 422 Steps per Game 500 300 200 100 400 300 200 100 0 100 200 300 0 400 100 350 300 Steps per Game 200 Epoch 10 100 300 400 250 200 150 50 200 300 400 Epoch 12 100 100 15 0 400 Average Lines per Game L Change in Weights 300 20 Epoch Epoch 0 200 25 100 200 Epoch 300 400 10 0 100 200 300 400 Epoch Fig Learning curve, steps and lines per update for Agents (top) and (bottom) Learning results are shown in Figure Agent clearly improves with more training epochs Surprisingly, Agent hits a steady low level, despite an initial improvement phase In any case, the performance of the learned strategies is way below expectations compared to the current state-of-the-art A deeper look into the problem indicated that the opponent in Adversarial Tetris is not very aggressive after all and the MiniMax criterion is way too conservative, as it assumes an optimal opponent In fact, it turns out that an optimal opponent could actually make the game extremely hard for the player; this is reflected in the game tree and therefore our player’s choices are rather mild in an attempt to avoid states where the opponent could give him a hard time Agent avoids this pitfall because it goes only to depth 1, where he cannot “see” the opponent, unlike Agent Nevertheless, the learned strategies are able to generalize consistently to the other MDPs (recall that training takes place only on MDP #1) For each learned strategy, we played 500 games on each MDP to obtain statistics Agent achieves 574 steps and 44 lines per game on average over all MDPs (366 steps and 16 lines on MDP #1), whereas Agent achieves 222 steps and 11 lines (197 steps and lines on MDP #1) Note that our approach is off-line; training takes place without an actual opponent It remains to be seen how it will perform in an on-line setting facing the exploration/exploitation dilemma References Breukelaar, R., Demaine, E.D., Hohenberger, S., Hoogeboom, H.J., Kosters, W.A., Liben-Nowell, D.: Tetris is hard, even to approximate International Journal of Computational Geometry and Applications 14(1-2), 41–68 (2004) Tsitsiklis, J.N., Roy, B.V.: Feature-based methods for large scale dynamic programming Machine Learning, 59–94 (1994) Reinforcement Learning Competition (2009), http://2009.rl-competition.org Bradtke, S.J., Barto, A.G.: Linear least-squares algorithms for temporal difference learning Machine Learning, 22–33 (1996) Thi´ery, C: Contrˆ ole optimal stochastique et le jeu de Tetris Master’s thesis, Universit´e Henri Poincar´e – Nancy I, France (2007) A Multi-agent Simulation Framework for Emergency Evacuations Incorporating Personality and Emotions Alexia Zoumpoulaki1, Nikos Avradinis2, and Spyros Vosinakis1 Department of Product and Systems Design Engineering, University of the Aegean, Hermoupolis, Syros, Greece {azoumpoulaki,spyrosv}@aegean.gr Department of Informatics, University of Piraeus, Greece avrad@unipi.gr Abstract Software simulations of building evacuation during emergency can provide rich qualitative and quantitative results for safety analysis However, the majority of them not take into account current surveys on human behaviors under stressful situations that explain the important role of personality and emotions in crowd behaviors during evacuations In this paper we propose a framework for designing evacuation simulations that is based on a multi-agent BDI architecture enhanced with the OCEAN model of personality and the OCC model of emotions Keywords: Multi-agent Systems, Affective Computing, Simulation Systems Introduction Evacuation simulation systems [1] have been accepted as very important tools for safety science, since they help examine how people gather, flow and disperse in areas They are commonly used for estimating factors like evacuation times, possible areas of congestion and distribution amongst exits under various evacuation scenarios Numerous models for crowd motion and emergency evacuation simulations have been proposed, such as fluid or particle analogies, mathematical equations estimated from real data, cellular automata, and multi-agent autonomous systems Most recent systems adopt the multi-agent approach, where each individual agent is enriched with various characteristics and their motion is the result of rules or decision making strategies [2, 3, 4, 5] Modern surveys indicate that there is number of factors [8, 9] influencing human behavior and social interactions during evacuations These factors include personality traits, individual knowledge and experience and situation-related conditions like building characteristics or crowd density, among others Contrary to what is believed, people don’t immediately rush towards the exits but take some time before they start evacuating, performing several tasks (i.e gather information, collect items) and look at the behaviors of others in order to decide whether to start moving or not Also route and exit choices depend on familiarity with the building Preexisting relationships among the individuals also play a crucial role upon behavior as members of the same S Konstantopoulos et al (Eds.): SETN 2010, LNAI 6040, pp 423–428, 2010 © Springer-Verlag Berlin Heidelberg 2010 424 A Zoumpoulaki, N Avradinis, and S Vosinakis group like friends and members of a family will try to stay together, move with similar speeds, help each other and aim to exit together Additionally, emergency evacuations involve complex social interactions, where new groups form and grow dynamically as the egress progress New social relations arise as people exchange information, try to decide between alternatives and select a course of actions Some members act as leaders, committed to help others, by shouting instructions or leading towards the exits while others follow [10] Although individuals involved in evacuations continue to be social actors, and this is why under non-immediate danger, people try to find friends, help others evacuate or even collect belongings, stressful situations can result to behaviors like panic [11] During an emergency, the nature of the information obtained, time pressure, the assessment of danger, the emotional reaction and observed actions of others are elements that might result to catastrophic events, such as stampedes The authors claim that above factors and their resulting actions should be modeled, for realistic behaviors to emerge during an evacuation simulation The proposed approach takes in consideration recent research not only in evacuation simulation models but also in multi agent system development [7], cognitive science, group dynamics and surveys of real situations [8] In our approach, decision making is based on emotional appraisal of the environment, combined with personality traits in order to select the most suited behavior according to the agents’ psychological state We introduce an EP – BDI (Emotion Personality Beliefs Desires Intentions) architecture that incorporates computational models of personality (OCEAN) and emotion (OCC) The emotion module participates in the appraisal of information obtained, decision making and action execution The personality module influences emotional reactions, indicates tendencies to behaviors and help address issues of diversity Additionally we use a more meaningful mechanism for social organization, where groups form dynamically and roles emerge due to knowledge, personality and emotions We claim that these additions may provide the necessary mechanisms for simulating realistic human like behavior under evacuation Although the need for such an approach is widely accepted, to our knowledge no other evacuation simulation framework has been designed incorporating fully integrated computational models of emotion and personality The Proposed Framework The proposed agent architecture (Fig.1) is based on the classic BDI (Beliefs-DesiresIntentions) architecture enriched with the incorporation of Personality and Emotions The agent’s operation cycle starts with the Perception phase, where the agent acquires information on the current world state through its sensory subsystem Depending on the agent’s emotional state at the time, its perception may be affected and some information may possibly be missed The newly acquired information is used to update the agent’s Beliefs Based upon its new beliefs, the agent performs an appraisal process, using its personality and its knowledge about the environment in order to update its emotional state The agent’s Decision making process follows, where current conditions, personality and agent’s ... Sublibrary: SL – Artificial Intelligence ISSN ISBN-10 ISBN-13 030 2-9 743 3-6 4 2-1 284 1-6 Springer Berlin Heidelberg New York 97 8-3 -6 4 2-1 284 1-7 Springer Berlin Heidelberg New York This work is subject... advancement of the state-of-the-art in Artificial Intelligence and Robotics S Konstantopoulos et al (Eds.): SETN 2010, LNAI 6040, p 3, 2010 c Springer-Verlag Berlin Heidelberg 2010 Robots, Natural... innovative and high-quality scientific results by AI researchers, attracting not only EETN members but also scientists advancing and applying AI in many and diverse domains and from various Greek and international