The ILIAS seminars are held at Campus Kirchberg.
Speaker: Koen Hindriks (TU-Delft)
Title: (to be announced)
Abstract: (to be announced)
About the speakers:
Speaker: Paolo Turrini (Utrecht University)
Title: (to be announced)
Abstract: (to be announced)
About the speakers:
Speaker: Pietro Baroni and Massimiliano Giacomin (University of Brescia)
Title: (to be announced)
Abstract: (to be announced)
About the speakers:
Speaker: Nick Tinnemeier (Utrecht University)
Title: (to be announced)
Abstract: (to be announced)
Speaker: Eric Pacuit (University of Tilburg)
Title: (to be announced)
Abstract: (to be announced)
Speaker: Francesca Toni (Imperial College)
Title: Combining statistics and arguments to compute trust
Abstract: We propose a method for constructing Dempster-Shafer belief functions modeling the trust of a given agent (the evaluator) in another (the target) by combining statistical information concerning the past behaviour of the target and arguments concerning the target's expected behaviour. These arguments are built from current and past contracts between evaluator and target. We prove that our method extends a standard computational method for trust that relies upon statistical information only. We observe experimentally that the two methods have identical predictive performance when the evaluator is highly "cautious", but our method gives a significant increase when the evaluator is not or is only moderately "cautious". Finally, we observe experimentally that target agents are more motivated to honour contracts when evaluated using our model of trust than when trust is computed on a purely statistical basis.
Speaker: Andreas Herzig (IRIT)
Title: The dynamic logic of propositional control
Abstract: Along the lines of Coalition Logic of Propositional Control (CLPC), we propose a dynamic logic of propositional control DLPC. It extends classical propositional logic by a variety of modal operators of assignment, and modal operators of transfer of control over a propositional variable. We also present an epistemic extension. We establish the relationship of these logics with the existing CLPC and with its extension by control transfer DCLPC. We study their complexity and their proof theory.
Speaker: Jan Broersen (Utrecht University)
Title: A stit-logic analysis of intentional action
Abstract: We study intentional action in stit-logic. The formal logic study of intentional action appears to be new, since most logical studies of intention concern intention as a static mental state. We model an intended action as an action that possibly deviates from the actual action conducted by an agent. First, an actual action may deviate from an intended action because the agent is not able to carry out the intended action, and the actual action is used as a means to achieve the goal of the intended action. This explains differences such as the one between the actions `murder' and `manslaughter', and problems like the `dentist's' scenario of Cohen and Levesque. Second, an actual action may deviate from an intended action because the agent's environment behaves unexpectedly and the result of an action is not the one envisaged by the agent. So, the action is unexpectedly unsuccessful. We show how to deal with the distinction between successful and non-successful action by weakening the notion of `knowingly doing' to its `belief' equivalent. Finally, we briefly consider weakening the epistemic attitude towards action performance even further by discussing opportunities to model the notion of `attempt'.
Speaker: Hans van Ditmarsch (University of Sevilla)
Title: Security protocols using card deals
Abstract: In the mid-19th century, Thomas Kirkman and Jacob Steiner gave birth to what later became known as design theory, a subdiscipline of combinatorial mathematics. Tools and techniques from that area can be used to design unconditionally secure (communicative) protocols. The epistemic logician can also focus on the epistemic properties of such protocols, or reformulate their specifications (such as security requirements) in such terms. Individual, common and distributed knowledge all play a role. Following original contributions from the 1980s onward by Winkler, Fischer, Wright and others, we focus on the case of computationally unlimited agents using card deals. Given various agents or players, two among them (sender and receiver, say, but note that they both send and receive) may wish to communicate a secret. We present some results for the restricted case where secrets concern card ownership, and for the general case where any secret bit can be exchanged by means of such card deals.
Speaker: Emiliano Lorini (IRIT)
Title: Epistemic Games in Modal Logic: from individual rationality to social preferences
Abstract: I present in this work a sound and complete modal logic called EDLA (Epistemic Dynamic Logic of Agency) integrating the concepts of joint action, preference and knowledge and enabling to reason about epistemic games in strategic form. I provide complexity results for EDLA. In the second part of the paper, I study in EDLA the epistemic and rationality conditions of some classical solution concepts like Nash equilibrium and Iterated Deletion of Strictly Dominated Strategies (IDSDS). Moreover, I combine EDLA with Dynamic Epistemic Logic (DEL) in order to model epistemic game dynamics. In the conclusion, I show how EDLA can used to model social preferences and group/team reasoning.
About the speaker: Emiliano Lorini has received a PhD in Cognitive Sciences from the University of Siena (2007). In 2007-2009, he was a post-doctoral researcher at the Institut de Recherche en Informatique de Toulouse (IRIT), France. He is actually a CNRS researcher at the university of Toulouse and at IRIT. He is working on several topics in the area of logic, multi-agent systems and cognitive sciences: theory of intentional action, theory of emotions, logical foundations of game theory, deontic logic, theory of institutions and organizations, theory of collective action and collective intentionality.
Speaker: Tony Hunter (University College London )
Title: Computational Models of Argumentation for Richer Logics
Abstract: Abstract argumentation provides a natural starting point for modelling argumentation. In abstract argumentation, a graph is used to represent a constellation of arguments and the conflict between them. Each argument is represented by a node, and the attack of one argument on another is represented by a directed arc. Proposals by Phan Minh Dung and others then give principled ways for determining which arguments are acceptable based on the structure of the graph. However, abstract argumentation assumes the existence of such graphs and does not provide means for generating them. To address this, Henry Prakken and others, have shown how logic can be used to generate arguments and to identify conflict between arguments. However, most proposals for using logic, assume a simple defeasible logic. Yet, elsewhere in artificial intelligence, there is the need to use richer logics such as classical logics, description logics, modal logics, temporal logics, and probabilistic logics. In this talk, we consider a framework for harnessing richer logics in computational models of argument.
Speaker: Alexandru Baltag (Oxford University)
Title: Iterated Belief-Revision with Higher-Level Information: from doxastic cycles to learning the truth
Abstract:
I investigate and compare the long term-behavior of various belief-revision "methods" (or "belief-revision policies", or "belief upgrades", as they are sometimes called), under learning of higher-level doxastic information. By "learning" I mean iterated belief revision with new truthful information. By "higher-level doxastic information" I mean sentences that may refer to the agent's own beliefs, or to her belief-revision plans (her conditional beliefs), even if they also convey factual information. "The long-term behavior" refers to whether or not the agent's beliefs stabilize eventually, converging to a fixed point, or keep changing forever, in infinite cycles. If they do stabilize, we are interested to find natural conditions ensuring they converge to true beliefs (or even to true and complete beliefs).
The methods I consider are the following: conditioning; the "conservative upgrade" or "minimal revision" of Boutilier; a variant of minimal revision (re)formulated in terms of Spohn ordinals; lexicographic (or "radical") upgrade (initially proposed, and rejected, by Spohn, then generalized by Nayak, and justified by Glaister based on symmetry reasons, and by Baltag and Smets as a special case of the "Action-Priority Update Rule"); Spohn's ordinal analogue of Jeffrey conditioning; Darwiche and Pearl's natural modification of Spohn's method (called the "ratched method" by Kelly); Goldszmidt and Pearl's method etc.
My results are in general negative, with two exceptions: when starting with an initial finite Spohn model, the only methods that guarantee convergence of (simple) beliefs are conditioning and lexicographic upgrade (as well as its Nayak generalization). For these two types of upgrades, one can easily find natural conditions ensuring that a stream of upgrades converges to true and complete beliefs (the condition is to have a "maximally informative upgrade stream"). The positive result for lexicographic upgrades is non-trivial. I also give characterizations of the fixed points reached eventually (if reached), of the 2-cycles, 3-cycles etc.
In the end, I relate this discussion to Learning Theory, to the Surprise Examination Puzzle and (time-permitting) to the issue of belief merge via persuasive communication within a group, with applications to rationality announcements in games of perfect information.
This talk is based on joint work with Sonja Smets.
About the speaker:
EMPLOYMENT AND DEGREES:
1989: Masters in Mathematics, University of Bucharest.
1990-1992: Journalist.
1992-1993: Researcher at the Institute of Mathematics of the Romanian Academy.
1993-1998: Fulbright research fellowship at Indiana University.
1998: PhD in Mathematics, Indiana University, under the supervision of Jon Barwise and Larry Moss.
1998-2001: Post-doctoral Researcher, at CWI and ILLC (Amsterdam).
2001-present: University Lecturer, Oxford University Computing Laboratory.
WORK AND INTERESTS:
modal logic, dynamic logic, epistemic logic, temporal logic;
belief revision, models for multi-agent information flow;
quantum logic and quantum information flow;
coalgebras, non-well-founded sets, Universal Set Theory, models for self-reference, circularity and fixed-points;
rationality and action in Game Theory;
Epistemology, Philosophy of Information and Philosophy of Science.
Speaker: Juan Antonio Rodriguez (IIIA, Spain)
Title: On the emergence of social conventions in multi-agent complex networks
Abstract: Distributed mechanisms that regulate the behavior of autonomous agents in open multi-agent systems (MAS) are of high interest since we cannot employ centralized approaches relying on global knowledge. In actual-world societies, the balance between personal and social interests is self-regulated through social conventions that emerge in a decentralized manner. As such, a computational mechanism that allows to engineer the emergence of social conventions in MAS can become a highly promising tool to endow open MAS with self-regulating capabilities. To this end we propose a computational self-adapting mechanism that facilitates agents to distributively evolve their social behavior to reach the best social conventions. Our approach borrows from the social contagion phenomenon: social conventions are akin to infectious diseases that spread themselves through members of the society. We will experimentally show that our mechanism helps a MAS to regulate itself by searching and establishing (better) social conventions on a wide range of interaction topologies and dynamic environments. Moreover, we will also show that the mechanism can cope with large convention spaces. Finally, we will turn our attention to the robust emergence of social conventions. Current computational models for the emergence of conventions assume that there is no uncertainty regarding the information exchanged between agents. However, this assumption is too strong when unreliable information might be exchanged because agents may: (i) deliberately lie; (ii) make wrong assessments or misjudgements; or (iii) communicate through noisy channels. Hence, within these settings conventions may fail to emerge. We will show how to extend our mechanism for the emergence of conventions to help agents reach global consensus on conventions despite information uncertainty. The extension is based on allowing agents to self-protect against unreliable information.
Speaker: Guillaume Aucher (ILIAS)
Title: BMS revisited
Abstract: In knowledge representation, the insight of the BMS logical framework (proposed by Baltag, Moss and Solecki) is to representhow an event is perceived by several agents very similarly to the way one represents how a static situation is perceived by them: by means of a Kripke model. There are however some differences between the definitions of an epistemic model (representing the static situation) and an event model. In this talk I will restore the symmetry. Unlike any other logical framework, this one allows to express statements about ongoing events, which are very useful in knowledge representation. It also models the fact that our perception of events (and not only of the static situation) can also be updated due to other events. I axiomatize it and prove its decidability. Finally, I show that it embeds the BMS one (if one adds common belief operators).
About the speaker: Guillaume Aucher is one of the ICR team members.
Speaker: Angela Bovo (ONERA Toulouse)
Title: Validity of a speech reported by several consecutive sources
Abstract: We are trying to determine the degree of trust we can give to a speech that has been reported by several consecutive sources, in a multi-agent, non-cooperative context. We propose a modal logic modeling for reported speech. We then define its properties of validity and invalidity, and the criteria for validity. In addition, we will show how this work relates to the Agent Communication Languages (ACLs). This work has possible applications in intelligence.
About the speaker: Angela Bovo is a master student in artificial intelligence. She studied in the University of Toulouse 3 and in the engineering school Ecole Nationale Superieure de l'Aeronautique et de l'Espace. She is currently working in the ONERA (Toulouse) for her master internship. She is currently working on language modelling, epistemic and doxastic logic as well as non cooperative agent communication languages. Angela would like to work on cognitive sciences and possibly spiking neural networks in the future.
Speaker: Alexander Artikis (Demokritos, Athens)
Title: Dynamic Protocols for Open Agent Systems
Abstract:
Multi-agent systems where the members are developed by parties with
competing interests, and where there is no access to a member's internal
state, are often classified as 'open'. The specification of protocols
for open agent systems of this sort is largely seen as a design-time
activity. Moreover, there is no support for run-time specification
modification.
Due to environmental, social, or other conditions, however, it is often
required to revise the specification during the protocol execution.
To address this requirement, we present an infrastructure for 'dynamic'
protocol specifications, that is, specifications that may be modified at
run-time by agents. The infrastructure consists of well-defined
procedures for proposing a modification of the 'rules of the game' as
well as decision-making over and enactment of proposed modifications.
We evaluate proposals for rule modification by modelling dynamic
specifications as metric spaces. Furthermore, we constrain the enactment
of proposals that do not meet the evaluation criteria.
About the Speaker: Alexander Artikis is a research associate at the National Centre for Scientific Research "Demokritos", in Athens, Greece, and a visiting researcher at Imperial College London. He was awarded his PhD in Computing (Multi-Agent Systems) from Imperial College London in 2003. His research interests lie in the areas of temporal representation and reasoning, and distributed artificial intelligence. He has been publishing papers in related conferences and journals, such as the Artificial Intelligence Journal, and the ACM Transactions on Computational Logic.
Speaker: Piotr Kazmierczak, University of Warsaw
Title: On tableau systems for some modal logics
Abstract: This paper is an attempt to present and compare two different tableau systems for modal logic K. First we present general philosophical ideas behind theorem proving and modal logics, then we briefly describe tableau for PC, and further tableaux for modal logics. Tableau systems are indeterminate, but with some additional regulations they can be turned into decision procedures—we show how is it achieved. We also present a somewhat innovative way to construct a labelled tableau for K, and at the end we show implementations of theorem provers for PC and K written in SWI Prolog. What is yet to be presented in the paper (not yet written) is an efficienct comparison between two different tableaux for K. It will be a strictly practical comparison (like a benchmark), showing strenghts and weaknesses of these systems.
Speaker: Xavier Parent (ILIAS, ICR)
Title: Remedial interchange, contrary-to-duty obligation, and commutation
Abstract: The talk will focus on the relation between deontic logic and the study of conversational interactions. Special attention will be given to the notion of remedial interchange introduced by E. Goffman into the literature on conversational interactions. This notion appears to be close to the one of contrary-to-duty (reparational) obligation, which deontic logicians have been studying in its own right. I will investigate the question of whether some of the aspects of conversational interactions can fruitfully be described using formal tools originally developed in the study of iterated belief change. I will adapt the latter tools to deontic logic, and attempt an account of remedial interchange (and, more generally, contrary-to-duty reasoning) in terms of commutation. This account brings the dynamics of obligations to the fore.
About the speaker: Xavier Parent is a research associate in the Individual and Collective Reasoning (ICR) group at the University of Luxembourg. His current research interests are focused upon both fundamental and applied research related to deontic logic, with a special focus on dyadic deontic logic and the like. He obtained a Ph.D in philosophy with his thesis ``Non-monotonic Logics and Modes of Argumentation. The case of conditional obligation". Prior to joining the ICR group in Luxembourg, he was a research associate at the Computer Science Department of the King's College London, UK.
Speaker: Matteo Baldoni (University of Turino)
Title: Choice, Interoperability, and Conformance in Interaction Protocols and Service Choreographies (a joint work with C. Baroglio, A. K. Chopra, N. Desai, V. Patti, M. P. Singh)
Abstract: Many real-world applications of multiagent systems require independently designed (heterogeneous) and operated (autonomous) agents to interoperate. We consider agents who offer business services and collaborate in interesting business service engagements. We formalize notions of interoperability and conformance, which appropriately support agent heterogeneity and autonomy. With respect to autonomy, our approach considers the choices that each agent has, and how their choices are coordinated so that at any time one agent leads and its counterpart follows, but with initiative fluidly shifting among the participants. With respect to heterogeneity, we characterize the variations in the agents' designs, and show how an agent may conform to a specification or substitute for another agent. Our approach addresses a challenging problem with multiparty interactions that existing approaches cannot solve. Further, we introduce a set of edit operations by which to modify an agent design so as to ensure its conformance with others.
About the Speaker: Matteo Baldoni is an associate professor at the Department of Computer Science of the University of Torino since 2006 where he took his Ph.D. in Computer Science in May 1998. From July 1999 through September 2005 he has been researcher at the Department of Computer Science of the University of Torino. He has a background in computational logic, modal and non-monotonic extensions of logic programming, multimodal logics, reasoning by actions and change. His current research interests include issues in communication protocol design and implementation, conformance and interoperability for agents and web services, agent programming languages, web services and service oriented architecture, personalization by reasoning in the semantic web, e-learning.
Speaker: Hans Rott (University of Regensburg)
Title: The Ramsey Test for Conditionals and Iterated Theory Change
Abstract: According to the Ramsey Test, conditionals reflect changes of theories (or beliefs): A>B is accepted in T iff B is accepted in the minimal revision of T necessary to accommodate A. More than 20 years ago, the Ramsey test has come under heavy attack. A series of impossibility theorems ("triviality theorems") seemed to show that given standard models of theory change, the Ramsey test cannot serve as a viable analysis of conditionals. Other authors have come to its defence, arguing that it is rather the standard AGM-type model of theory change that is mistaken. In this talk I argue that an overly postulational approach to the semantics of (nested) conditionals should be avoided and that one should instead turn to an analysis in terms of constructive models of (iterated) theory change.
A crucial question is whether it is possible to use the Ramsey Test for the interpretation of conditionals and still respect the Preservation Condition according to which the original theory T should be fully retained after a revision by information that is consistent with T. Among the four most natural models for iterated belief change, I identify two solutions that indeed allow us to combine the Ramsey test with Preservation in languages containing only non-nested conditionals of the form A>B. These solutions, however, violate Preservation for nested conditionals of the form A>(B>C). I argue that by looking at the constructve models, we can understand why it has been wrong to expect that Preservation holds in languages containing nested conditionals.
Speaker: Dov Gabbay (King's College / University of Luxembourg)
Title: Matrix Abduction with Applications to Argumentation Theory and Paradoxes of Voting
Abstract We motivate and introduce a new method of abduction: Matrix Abduction. Given a matrix A with entries in {0, 1}, we allow for one or more blank squares in the matrix, say ai,j = ?. The method allows us to decide whether to declare ai,j = 0 or ai,j = 1 or ai,j = ? undecided. This algorithmic method is then applied to modelling several legal and practical reasoning situations including the Talmudic rule of Kal-Vachomer. We show that this new method can also be applied to the analysis of paradoxes in voting and judgement aggregation. In fact we have here a general method for executing non-deductive inferences. The method also offers a non deductive realisation of argumentation networks, where arguments are abduction matrices and the form of attack is matrix expansion.
About the Speaker: Dov Gabbay is one of the godfathers of the field of formal logic.
Speaker: Martin Caminada and Gabriella Pigozzi (ILIAS)
Title: On Judgement Aggregation in Abstract Argumentation
Abstract We present a paper that we have recently written for the special issue on computational social choice of the Journal of Autonomous Agents and Multi-Agent systems (JAAMAS). The topic is applying judgement aggregation to abstract argumentation
Speaker: Catherine Pelachaud (CNRS, TELECOM ParisTech)
Title: Virtual humanoid listener
Abstract Communicating with virtual humanoids imply they should have the capability to speak and to listen. In the past years we have concentrated our effort on creating a virtual humanoid able to communicate not only through speech but also through its nonverbal behaviours. It can display facial expression of emotion, it can gaze at a person, it can describe an object or show a point in space, etc. Lately we have started to work on a listener model. In a conversation not only the speaker but also the listener are active participants. Both interactants send verbal and nonverbal messages. Trough facial expression, head movement, or even gaze, the listener indicates what are its attitudes toward the speaker’s say, if it likes or dislikes, agrees or disagrees, understands or not, etc. In this talk we will present on-going work on the creation of a virtual humanoid able to communicate with human users. In particular we will describe our model of a virtual listener able to react to speaker’s speech and to indicate its mental state in real-time.
About the speaker Catherine Pelachaud is a Director of Research at CNRS in the laboratory LTCI, TELECOM ParisTech. In 1991 she obtained her PhD in Computer Graphics at the University of Pennsylvania. During 1993 and 1994 she was postdoc with a National Science Foundation grant, involved in the production of a system which automatically generates and animates conversations between multiple human-like agents with appropriate and synchronized speech, intonation, facial expressions, and hand gestures. Her current reserach topics are:
Speaker: Martin Rehak (Department of Cybernetics, Czech Technical University, Prague)
Title: CAMNEP: Collaborative Reduction of False Positives in Network Intrusion Detection
Abstract: Current network behavior analysis methods based on anomaly detection approaches suffer from comparatively higher error rate and low performance. We propose a framework system which addresses these issues by (i) using hardware-accelerated probes to collect unsampled NetFlow/IPFIX data from gigabit-speed network links and (ii) combining several anomaly detection algorithms by means of collective trust modeling, a multi-agent data fusion method. The data acquired on the network is preprocessed in the collector database and then passed to several anomaly detection methods to obtain several independent anomaly opinions for each flow. Each of these methods uses a distinct set of aggregate traffic features to determine the anomaly of each flow, which is determined by comparing the observed flows with a method-specific traffic prediction and/or a set of rules. The anomaly data is passed to several trust models to aggregate the current anomalies with past experience. Each of the trust models represents the flows in a high dimensional feature space, and progressively builds flow clusters which represent typical traffic behaviors. The clusters are labeled with the typical anomaly of the flow in the cluster, and this trustfulness (i.e. the anomaly aggregated over a long term) is used to evaluate the current flows, instead of the immediate anomaly determined for this flow. The trustfulness values provided by trust models based on different features are integrated into the final result and compared with a dynamically-defined threshold. Our experiments performed on real university and WAN networks suggest that the framework significantly improves the error rate while being computationally efficient, and is able to process the network speeds up to 1Gbps in online mode. Depending on the specific network, it can remove up to 95% of false positives when compared to individual anomaly detection methods, and up to 75% of false positives when compared to simple average of flow anomalies.
About the speaker:
Martin Rehak holds engineering degree from Ecole Centrale Paris. He is currently researcher at Agent Technology Group of the Gerstner laboratory and in the same time pursues his PhD studies at the Department of Cybernetics of the Czech Technical University. His current research interests are security, network intrusion detection, trust modeling and task allocation in adversarial environments.
Prior to his current position, Martin was member of the Mobile Communication Operations team of Schlumberger Smartcards (now Axalto), where he was working on definition, design and integration of novel location-based and other value added services for major European and African operators. Before obtaining his degree, he was a freelance contractor to diverse Czech and French companies, including General Electric Medical Systems. Martin is a 2005 McKinsey Scholar, and a member of AAMAS’06-08 program committee. In 2006, he was a visiting researcher with National Institute of Informatics, Tokyo.
Speaker: Davide Grossi (ILIAS)
Title: Correspondences in the theory of aggregation
Abstract In this talk I will investigate some interrelationships between the social-theoretic problems of preference and judgment aggregation from the perspective of formal logic. On the one hand, preference aggregation will be shown to be equivalent to the aggregation of some special types of judgments (logical formulae). On the other hand, the judgment aggregation problem will be proven equivalent to the aggregation problem of specific types of preferences. As a result, a detailed map of aggregation problems is obtained which highlights differences and similarities between the two research fields of preference and judgment aggregation, which allows the cross import of impossibility results between them, and which makes a systematization of the two fields possible.
About the speaker Davide Grossi has been a member of our ICR research group in 2008 and will soon start to work at the ILLC in Amsterdam.
Speaker: Juliana Bueno-Soler (UNICAMP, Brazil)
Title: Anodic and cathodic modalities: managing negatons in modal logic
Abstract: I will discuss the role of negaton in the domain of anodic logics (modal logics withouth negation) and cathodic logics (paraconsistent modal logics with some kind of negation). Motivations for studying such logics as connected to some deontic paradoxes will be given. Some difficulties about completeness for anodic logics will be emphasized, and completeness results, as well as an incompleteness result, will be sketched. This is a work in progress, and questions, problems and research lines wil be also assessed.
About the speaker: Juliana Bueno-Soler is a visiting PhD student from the Department of Philosophy and CLE at UNICAMP, Brazil.
Speaker: Jonathan Ben-Naim (IRIT)
Abstract: In many multi-agent systems, the user has to decide whether he (or she) sufficiently trusts a certain agent to achieve a certain goal. To help users to make such decisions, an increasing number of trust systems have been developed. By trust system, we mean a system that gathers information about an agent and evaluates its trustworthiness from this information. The aim of the present paper is to develop new trust systems that overcome limitations of existing ones. This is a challenging problem that raises questions such as: how trustworthiness may be represented, and from which information it may be estimated? We assume that a set of grades describing the past performances of the agent is given. With this common basis, two approaches are proposed. In the first one, the aim is to construct an interval that summarizes the grades. Such an interval gives a good account of the trustworthiness of the agent. We establish axioms that should be satisfied by summarizing methods, devise a particular method based on pulling, and check that it satisfies the axioms, which provides theoretical justifications for it. In the second approach, which is more briefly presented, a level of trust as the certainty that a future grade will be good, and a level of distrust as the fear that a future grade may be bad, are computed on the basis of the past grades. This approach is based on possibility theory and provides, thanks to the two levels, another view of trustworthiness, as well as summarizing intervals.
About the speaker: Jonathan Ben-Naim has received a PhD in Computer Science from the University of Aix-Marseille (2006). In 2006-2007, he was a post-doctoral researcher at the University of Luxembourg. Since October 2007, he is a CNRS researcher at the university of Toulouse. He is currently working on the development and the axiomatization of trust and reputation systems.
Speaker: Walter Carnelli (State University of Campinas, Brazil)
Title: Possible-Translations Semantics and Logics of Formal Inconsistency
Abstract: Besides its great importance for mathematical reasoning, standard logic, sometimes called "classical logic", cannot be seriously accepted as a basis for rationality. Indeed, rationally acceptable statements are not coincident with true statements, and contradictions in reasoning sometimes play a most important role. So, for instance, only contradictory allegations can make a judge decide whether there is any false statement around--there is no other reason why, in justice, people are interviewed in separate. Much to the contrary to what some philosophers maintain, in informal reasoning the use of contradictions is inherent and contradictions may contain precious information. It is a task of logic to offer a suitable formal model for the perfectly licit act of reasoning under contradictions, and paraconsistent logic accomplishes this. In particular, the wide family of "ogics of formal inconsistency" (LFI's) achieve this in a remarkably natural and elegant way. The possible-translations semantics (PTS's) were devised in 1990 in order to offer a palatable interpretation to some non-classical logics, and are very well adapted to LFI's. PTS's depart from the notion of translations as morphisms between logics (maps preserving their consequence relations). If translations are thought as different "world views", the concept of possible-translations semantics offer a way to interpret a given logic L as the combination of all possible "world views", technically viewed as an appropriate set of translations of the formulas of L into a class of "simpler" logics with known semantic characterization. In this way LFI's give a very natural account of understanding the phenomenon of a sentence and its negation being both true. However, PTS's can be given to non-paraconsistent logics as well, and also serve as a powerful tool to decompose logics in the spirit of the program of combining logics. Several examples will be discussed and some research directions will be suggested.
About the speaker: Walter Carnielli is professor of the Department of Philosophy and of the Centre for Logic, Epistemology and the History of Science at the University of Campinas, Brazil. He has a Master degree in Algebra (1978) and a Ph.D. in Logic (1982) from the University of Campinas, with post-doc positions at the University of California, Berkeley and in Germany (Münster and Bonn) as Alexander von Humboldt grantee. Carnielli's research is focused on foundations of non-classical logics and applications to common-sense reasoning and computation. He is presently visiting the Individual and Collective Reasoning group at CSC- University of Luxembourg on an FNR research grant.
Speaker: Mehdi Dastani (Utrecht Univerity)
Title: Normative Multi-Agent Programs and Their Logics
Abstract: Multi-agent systems are viewed as consisting of individual agents whose behaviors are regulated by an organization artefact. This paper presents a simplified version of a programming language that is designed to implement norm-based artefacts. Such artefacts are specified in terms of norms being enforced by monitoring, regimenting and sanctioning mechanisms. The syntax and operational semantics of the programming language are introduced and discussed. A logic is presented that can be used to specify and verify properties of programs developed in this language.
About the speaker: Mehdi Dastani is a lecturer in computer science at the Utrecht University. He has received his master degrees in computer science (1991) and philosophy (1992) from the University of Amsterdam and obtained his PhD degree (1998) at the institute for logic, language and computation, University of Amsterdam. He is currently working in the area of multi-agent systems, multi-agent programming, and agent logics for the last ten years and has published many papers on these subjects.
Speaker: Odinaldo Rodrigues (King's College)
Title: A multi-level framework for information change
Abstract: In belief revision, it is often the case that preference and consistency are intertwined concepts. The most general view is that of an input formula which is preferred to an old belief set when together they are inconsistent (in the classical logic sense). This is not always adequate for several reasons and many variations of belief revision techniques have been proposed to address specific issues, e.g., iteration of the revision process.
In this talk, we argue that a relaxation and separation of some the key concepts is issential for any practical application of belief revision. We then propose a multi-level model in which different aspects of the problem can be represented and dealt with separately according to the application area's particular needs. We illustrate the technique with an example used in the modelling of requirements specification in software engineering.About the speaker: Odinaldo Rodrigues has a BSc in Informatics from the University of Fortaleza, a MSc in Systems Engineering and Computation from the Federal University of Rio and a PhD in Computing from Imperial College London. He is currently a member of the Group of Logic, Language and Computation in King's College London. His research concentrates on the formalisation of common-sense reasoning, theory change and Artificial Intelligence techniques and their applications to computer science.
Speaker: Maria Biryukov (ILIAS, MINE)
Title: Language Classification of Personal Names with Application to DBLP
Abstract: We propose a new perspective for the data analysis in digital libraries, bibliographic and other databases containing personal names. Knowing language/cultural background of a person can be beneficial in many applications, however this information is often not present explicitly in the databases. We present in this talk a statistical tool for the automatic language detection of personal names. Our system does not require a dictionary of names for training and handles 14 different languages so far. General purpose corpora for all Western European, Chinese, Japanese and Turkish languages are used in order to build simple statistical models of the languages. The system is fine tuned to achieve precision and recall above 90% for many languages. On an example of a bibliographical database DBLP we show how our system can be used in tasks such as data cleaning and discovery of trends. In the second part of the talk we demonstrate how the co-author network, which is built from the bibliographic records, an be incorporated into the process of personal name language classification. Testing the model on the DBLP data set shows that the extension of the language classification process with the co-author network may help to refine the classification obtained from the author names considered independently. It may also lead to the discovery of dependencies between the elements of the co-author network, or participation of authors in scientific communities.
About the speaker: Maria has graduated in Master in Artificial Intelligence program from K.U. Leuven in September 2004. In 2005-2006 Maria was working in the University of Namur where she was involved in the project aiming at development and implementation of a domain independent tool for creation of question answering systems. In January 2007 Maria joined the University of Luxembourg as a Ph.D student and is working in the MINE group under the supervision of Prof. Christoph Schommer.
Speaker: Thierry Mamer
Title: Improving the use of ILP for Biological Grammar Acquisition
Abstract: This talk will present some of the recent work done in my PhD studies. In the first part of the talk, some basic principles of Inductive Logic Programming (ILP), Biological Grammar Learning and Minimum Description Length will be introduced. The second part will identify a shortcoming of a standard positive-only clause evaluation function within the context of learning biological grammars and explain how we propose to overcome it. We will describe L-modification, a modification to this evaluation function such that the lengths of individual examples in the training data are considered. We will show, among other things, that using L-modification to learn from a protein family called Neuropeptide Precursors results in induced grammars that have a better performance than that achieved when using the standard positive-only clause evaluation function.
About the speaker: Thierry is a research student at the Robert Gordon University in Aberdeen, Scotland. He has graduated with an honours degree in Computer Science (Artificial Intelligence) in 2005. He is currently conducting his last experiments and hopes to start writing up his PhD thesis soon. His current interests lie in the area of Machine Learning, Inductive Logic Programming, Bioinformatics, and loosely in Data Compression.
Speaker: Sascha Kaufmann (CSC)
Title: Making Web Portals Convivial
Abstract: In this talk, we want to motivate the idea of conviviality in web portals and argue that a convivial social being deeply depends on the implicit and explicit co-operation and co-laboration of natural users inside a community. Our belief is that an individual conviviality can benefit from the wisdom of the crowd, meaning that a continuously and dynamic understanding of the user's behaviour can heavily influences the inidividual well-being.
For this, we introduce the system CUBA, which stands for "Conviviality and User Behavior Analysis". The purpose of CUBA is to find novel ways to support users during their visits while discovering their interests. In this respect, CUBA comes up with certain recommendations and suggestions, which are partially based on a common behavior of participants in general. For example, concepts like time, space, and diverse user-based actions are taken into account. The talk closes with a report about the current state of the project.
About the speaker: Sascha Kaufmann received his diploma in Computer Science from the JW Goethe-University in Frankfurt/Main. He is now a doctoral student at the MINE Group, ILIAS Laboratory.
Speaker: Yining Wu (ILIAS)
Title: Towards an Argument Game for Stable Semantics
Abstract: We present a discussion game for argumentation under stable semantics. Our work is inspired by Vreeswijk and Prakken, who have defined a similar game for preferred semantics. We restate Vreeswijk and Prakken’s work using the approach of argument labellings and then show how it can be adjusted for stable semantics. The nature of the resulting argument game is somewhat unusual, since stable semantics does not satisfy the property of relevance.
The current presentation is a try-out for a talk at the CMNA workshop (Computational Models of Natural Argument) at the end of July.
About the speaker: Yining Wu has obtained a Master's degree of Computational Logic at the Technische Universität Dresden and is currently doing a PhD on Argumentation and Trust at the ILIAS group of the University of Luxembourg.
Speaker: Manfred Jaeger (Aalborg University)
Title: Probabilistic modeling and learning in relational domains
Abstract: In the last few years we have seen the emergence of the new field of "Statistical Relational Learning", also called "Probabilistic Logic Learning", or "Relational Data Mining". The plethora of labels results from the fact that this field is a confluence of several quite distinct strands of research: one line of research is the combination of probabilistic graphical models with more abstract, logic-based, knowledge representation languages. A second line of research contributing to the new field are probabilistic extensions of inductive logic programming techniques. Finally (and perhaps most significantly), probabilistic logic learning joins together these AI traditions with those areas in machine learning that are concerned with learning from structured data (graphs, sequences, relational databases, ...).
In this talk I will first introduce Probabilistic Logic Models as a formal semantic basis for a wide range of models and representation languages used in statistical relational learning, especially the language of Relational Bayesian Networks, which will also be described. Finally, I will present recent developments on parameter learning for Relational Bayesian Networks, and on relational feature discovery using Type Extension Trees.
About the speaker: Manfred Jaeger obtained a Diploma in Mathematics from Freiburg University in 1991, and a PhD in Computer Science from the University Saarbrücken in 1995. From 1995 to 2003 he was research associate at the Max-Planck Institute for Computer Science in Saarbrücken, and also spent time as postdoctoral researcher at Stanford University, Helsinki University, and Freiburg University. In 2003 he joined the Machine Intelligence group at Aalborg University as Associate Professor. His research centers around the integration of probabilistic and logical reasoning in Artificial Intelligence and Machine Learning.
Speaker: Jean-François Raskin (Université Libre de Bruxelles)
Title: An introduction to games played on graphs
Abstract: In this talk, I will review the fundamental notions underlying the notion of games played on a graph. I will consider both turn-based and concurrent games, different notions of strategies, winning conditions, and the foundations for algorithms to solve those games.
About the speaker: Jean-François Raskin was born in Belgium, in 1972. He received his master and Ph.D. degrees in computer science from the Université de Namur, Belgium, in june 1995 and april 1999, respectively. After long research stays at the University of California at Berkeley, Max-Planck Institute for Computer Science (Saarbrucken), and ENS Paris, he is now Associate Professor (chargé de cours) at the computer science department of the Université Libre de Bruxelles (ULB), Belgium. His research interests are the formal methods for the verification of concurrent, real-time, and hybrid systems. At the ULB, he teaches subjects in theoretical computer science : automatic proofs and proofs of programs, logic in computer science, computability and complexity, and analysis of hybrid systems. He is also the secretary of the Belgian F.N.R.S. contact group on Theoretical Computer Science.
Speaker: Henry Prakken (University of Utrecht / University of Groningen, the Netherlands)
Title: Integrating Different Modes of Reasoning
Abstract: Many reasoning problems involve subproblems of different kinds, which can be solved in different ways. In this talk an abstract formalism will be presented in which any set of specific problem solving methods satisfying some weak conditions can be combined. The formalism makes it possible to formally express dependencies between different subproblems, to define an overall solution on the basis of solutions to subproblems, and, most importantly, to manage alternative, incompatible solutions to (sub)problems. The formalism will be illustrated with a legal example that combines standard first-order logic, Bayesian probability theory and two versions of default logic.
About the speaker: Henry Prakken is a lecturer in the Intelligent Systems group of the computer science department at Utrecht University, and professor of Law and ICT at the law faculty of the University of Groningen. He has master degrees in law (1985) and philosophy (1988) from the University of Groningen. In 1993 he obtained his PhD degree at the Free University Amsterdam with a thesis titled Logical Tools for Modelling Legal Argument. Prakken's research concerns logical foundations of common-sense reasoning and argumentation, and the application of Artificial Intelligence and advanced IT to legal reasoning, dispute resolution, group decision making and negotiation.
Speaker: Martin Mozina (University of Ljubljana, Slovenia)
Title: Integration of Machine Learning and Argumentation
Abstract: Machine learning is concerned with development of algorithms for extracting implicit knowledge from data. As these algorithms need to select among several alternative explanations of data, the resulting models often seem incomprehensible to domain experts. On the other hand, since argumentation studies and formalises common-sense reasoning and can be therefore used to represent experts' knowledge in a natural way, we can expect that their integration would yield benefits in both areas. During my talk I will introduce basics of machine learning and point out several possible combinations of machine learning and argumentation. The focus will be on recently developed approach Argument Based Machine Learning (ABML), which uses arguments provided by experts to guide machine learning algorithms towards comprehensible explanations of data.
About the speaker: Martin Mozina has received a BSc degree in computer science at the Faculty of computer and information science in Ljubljana in year 2003, and is since a PhD student at the Artificial Intelligence Laboratory. In years 2004-2007 he worked as a reasearcher on european project ASPIC (Argumentation Service Platform with Integrated Components) and currently he works now on X-Media project (Large scale knowledge sharing and reuse across media). His research mainly includes machine learning, the use of argumentation in machine learning algorithms, visualization of learned models, and programs for automatic tutoring.
Speaker: Nikos Vlassis (Dept. of Production Engineering and Management Technical University of Crete, Greece)
Title: Distributed Decision Making for Large Teams of Agents
Abstract: We address the problem of optimal action selection and coordination of a large team of collaborative agents. We view the problem as a combinatorial optimization problem where we need to maximize a payoff function that is additively decomposed over a graph. We first review a dynamic programming approach that finds the optimal solution but has exponential worst-case time complexity. We then describe an approximate message-passing algorithm for action selection, similar to belief propagation for Bayesian networks, that exhibits a very good runtime/performance trade-off. We discuss extensions to distributed stochastic optimal control, and outline possible applications.
About the speaker: Nikos Vlassis received a diploma in Electrical and Computer Engineering in 1993, and a PhD in 1998, both from the National Technical University of Athens, Greece. In 1999 he was visiting researcher at the Electro-Technical Laboratory in Tsukuba, Japan. From 2000 until 2006 he held an Assistant Professor position at the University of Amsterdam, The Netherlands. Since 2007 he holds an Assistant Professor position at the Technical University of Crete, Greece. He has published one monograph on multiagent systems and several papers in the fields of robotics and machine learning. His current research interests include robotics, multiagent systems, machine learning, and stochastic optimal control.
Speaker: Andreas Herzig
Title: Grounding power on actions and mental attitudes
Abstract: The main objective of this work is to develop a logical framework called IAL (Intentional Agency Logic) in which we can reason about mental states of agents, action occurrences, agentive and group powers. IAL will be exploited for a formal analysis of different forms of power such as an agent's power of achieving a certain result, an agent's power to do a certain action and an agent's power over another agent.
Speaker: Régis Riveret
Title: Argumentation: Game Theory as Heuristics.
Abstract:
Over the years many dialogue games for argumentation have been proposed to shed
light on questions such as which conclusions are (defeasibly) justified, or how
procedures for debate and conflict resolution should be structured to arrive at
a fair and just outcome. An aspect of debates which has not yet received much
attention is the common sense observation that the outcome does not solely
depend on the premises of a case, but also on the strategies that parties in a
dispute actually adopt. The focus is on adjudication
debates in which two
participants aim to persuade each other to adopt a certain opinion and, a
neutral third party (for example, a judge) must decide at the end whether to
accept the statements that the opposing parties have made during the debate.
In this presentation, the problem studied is how to use game theory to
determine optimal strategies in dialogue games for argumentation. To do so, the
main notions of Prakken's abstract argument games are interpreted in game
theory, and optimal strategies are defined accordingly. Then, a specification
of the expected utility of a strategy is provided by combining the probability
of success of arguments with their associated costs and benefits. The
presentation will also discusses some related work and, suggest future
investigations.
Speaker: Carlos Ribeiro.
Title: Research on Multi-agent Systems and Complex Social Networks at the Technological Institute of Aeronautics.
Abstract:
In this talk I will first present some aspects of the Technological Institute of
Aeronautics (ITA) in the broad context of the Air Industry. ITA is the main
centre for research and education on Aerospace, Aeronautics and Air Force issues
in Brazil. Founded in 1950, it is one of the main responsibles for the important
role the brazilian aeronautics industry has today, with its main
company—Embraer—having produced more than 4,100 aircrafts which
operate in 69 countries.
The talk will then switch to the research carried out at the Artificial
Intelligence and Robotics Group at ITA, particularly in the topics on
Multi-Agent Systems and Complex Social Networks. On Multi-Agent Systems,
I will present our research on the design of overlay networks for incident
(failures and attacks) detection in a network model. The purpose of this
research is to devise mechanisms for topology control that considers na
autonomous negotiation process based on a consensus among nodes in a
hierarchical structure. Still in the context of Multi-Agent Systems, we present
a method for pairwise combination of partial robotic maps based on
paraconsistent logics which can be adapted for a consensus-based mobile robotic
network. Finally, we explore the issue of Social Network Analysis regarding
both its technological and theoretical aspects: a project on fraud and
irregularity detection for the Brazilian Inland Revenue and the Brazilian Air
Force) and the analysis of metrics for evaluating small-world properties in
large social networks.
Speaker: Mozhgan Tavakolifard
Title: A Distributed and Statistical based Trust and Reputation Management Framework
Abstract: The rapidly changing environments of the internet suffer from problems such a fragile trustworthiness of millions active entities on the internet, e.g. human and mobile agent. This problem is nontrivial as more and more commercial transactions get carried out over the internet. Therefore, devising an effective way for verifying the trustworthiness entity in such complex environments is essential. To this end, Ntropi framework was designed to facilitate the exchange of trust and reputation in information/business environments. The trust mechanism plays a key role in security of multi-agent systems. Also the trust establishment is nontrivial, since the traditional and social means of trust cannot be applied directly to virtual settings of these environments. Because, in many cases the involved parties have not interacted before. Therefore, the reputation techniques must be used to estimate the good qualities and behaviors in online markets and communities, and also sanction the poor qualities and bad behaviors of this. The Ntropi has classified the trust into direct (explicit) and recommended classes. The direct class is divided into five agent-based sections called "very trustworthy", "trustworthy", "moderate", "untrustworthy", and "very untrustworthy". But the recommended trustworthy is derived from word-of-mouth (e.g., opinions), which is called reputation, and can be translated to direct or regular trusts. This paper presents an automated and autonomous trust system using Bayesian inference and improved Dirichlet distribution. Also a re-trusting mechanism (forgiveness) is devised and incorporated into Ntropi framework. This simulation results indicate the effectiveness of our model and the percentage of improvements.
Speaker: Denis Zampunieris
Location: Room E012
Title: e-Learning and Proactive Computing
Abstract:
Learning Management Systems (LMS) or e-learning platforms are dedicated
software tools intended to offer a virtual educational and/or training
environment online. Although there are already a large number of
functions for a variety of different users acting specific roles in
these environments, current LMS are fundamentally limited tools. Indeed,
they are only reactive software developed like classical, user-action
oriented software. These tools wait for an instruction, most probably
given through a graphical user interface, and then react to the user
request.
Students using these online systems could imagine and hope for more help
and assistance tools, especially because those users are often
inexperienced ones and expect some guidelines (what to do and how to do
it) from the system instead of a static user interface. LMS should tend
to offer to these students some personal, immediate and appropriate
support based on an intelligent analysis of their (lack of) actions,
like teachers do in classrooms.
Moreover, certain users like e-tutors have to peruse lots of data in
order to try to efficiently manage other users' needs and expect some
clues (where to search and what to look for) from the system instead of
a passive database.
We introduce a new kind of Learning Management Systems: Proactive LMS,
designed to improve their users' online interactions by providing
programmable, automatic and continuous analyses of users (inter-)
actions augmented with appropriate actions initiated by the LMS itself.
Proactive systems [D. Tennenhouse, "Proactive Computing", Communications
of the ACM, 43 (5), 2000, pp. 43–50] adhere to two premises: working on
behalf of, or pro, the user, and acting on their own initiative, without
user's explicit command. Proactive behaviors are intended to cause
changes, rather than just to react to changes. This is a major change
from interactive computing, in which we lock a system into operating at
exactly the same frequency as humans do.
Our proactive LMS can, for example, automatically and continuously help
and take care of e-users with respect to previously defined procedures
rules, and even flag other users, like an e-tutor, if something "wrong"
is detected in their behaviors; it can also automatically verify that
awaited behaviors of e-users have been carried out, and it can react if
these actions did not happen.
Speaker: Eugen Staab
Location: Room A02
Title: Excusableness for Failing Agents or False Negatives in Trust Assessment
Abstract: To estimate how much an agent can be trusted, its trustworthiness needs to be assessed. Usually, poor performance of an agent leads to a decrease of trust in that agent. This is not always reasonable. If the environment interferes with the performance, the agent is possibly not to blame for the failure. We examine which failures can be called excusable and hence must not be seen as bad performances. Knowledge about these failures makes assessments of trustworthiness more accurate. In order to approach a formal definition of excusableness, we introduce a generic formalism for describing environments of Multi-Agent Systems. This formalism provides a basis for the definition of environmental interference. We identify the remaining criteria for excusableness and give a formal definition for it. Our analysis reveals that environmental interference and a strong commitment of the performing agent do not suffice to make a failure excusable.
Speaker: Paul Harrenstein
Location: Room A02
Title: Dominance in Social Choice and Coalitional Game Theory
Abstract: We consider dominance relations for social choice as based on the pairwise majority rule on the one hand and cooperative games with non-transferable utility (coalitional NTU games) on the other. As these dominance relations may fail to be transitive and even contain cycles, the notion of maximality becomes untenable as an analytical tool. Both in social choice theory and cooperative game theory, a number of concepts have been proposed to take over the role of maximality in the absence of transitivity. In 1953 McGarvey showed that any irreflexive and anti-symmetric relation can be obtained by the majority rule. In this paper, we address the analogous issue for finite NTU games. We find any irreflexive relation over a finite set can be obtained as the dominance relation of some ordinary, monotonic, and simple NTU coalitional game. We also show that any dominance relation can be induced by a non-cooperative game via beta-effectivity. Furthermore, we obtain a partial result for the case in which alpha-effectivity is used and consider the formal interrelationships between Smith sets, Schwartz sets, stable sets and the core in finite NTU games.
Speaker: Girma Berhe
Title: Autonomic management for distributed systems: Example of DIET
Abstract:
Today's increasingly complex architectures and distributed computing
infrastructures require highly skilled IT professionals to install,
configure, operate, tune and maintain them. This impedes effective and
efficient execution of business processes and the delivery of services
that relay on the IT infrastructure. In particular, IT management has
become too complex and costly. Moreover, the fact that management tasks
are performed by humans leads to many configuration errors and low
reactivity. Autonomic computing helps to address these problems by
providing self-managing capabilities to IT systems. In this presentation,
the foucs is on the use of JADE, a middleware for self-management of
distributed software environments.
The main principle is to wrap legacy systems(or software elements) in
components in order to provide a uniform management interface, thus
allowing the deployment of distributed applications and reconfigure them
autonomously as required. A particular use case and experiments in using
JADE will be presented.
Speaker: Sanjay Modgil
Title: Reasoning about Preferences in Argumentation Frameworks
Abstract:
A Dung argumentation framework consists of a set of arguments
related by a conflict based binary attack relation. A 'calculus of opposition'
is then applied to determine the justified or 'winning' arguments under
different extensional semantics. The framework abstracts from the underlying
logic in which the arguments and attack relation are defined. Dung's seminal
theory of argumentation has thus become established as a general framework
for non-monotonic reasoning, and, more generally, reasoning in the presence
of conflict.
To determine a unique set of justified arguments invariably requires
that a preference relation on arguments is available. However, preference
information is often itself defeasible, conflicting and so subject to
argumentation based reasoning. In this talk I will describe my work on
extending Dung's framework to include arguments that claim preferences
between other arguments. Specifically, the framework is extended to include
a second attack relation such that an argument that claims a preference
between two other arguments, attacks the binary attack between these two
conflicting arguments. I will describe how the the justified arguments of
an extended framework are evaluated under the full range of Dung's
extensional semantics, and discuss how the core results for Dung frameworks
scale up to extended frameworks. I will therefore propose that the extended
semantics provides a general a framework for non-monotonic reasoning
formalisms that accommodate defeasible reasoning about as well as with
preference information. To substantiate this proposal I will show how
existing reasoning formalisms can be formalised and extended in the
framework, and will illustrate with examples demonstrating agent
reasoning over beliefs goals and actions.
Speaker: Kittichai Lavangnanda
Location: Room C02
Title: Introduction to Evolutionary Computation and an application in knowledge discovery (a data mining program SARG)
Abstract:
Evolutionary Computation is one of the most popular techniques used in a sub-field of Artificial Intelligence (AI) known as Soft Computing or Computational Intelligence. Evolutionary Computation are used to offer solutions to problems where no known conventional methods or algorithms can guarantee optimal/best solution. Knowledge Discovery (or Data Mining) is a broad term which is used to describe techniques in discovering useful patterns hidden in a given dataset. One such techniques is to generate rules which discover these patterns by relating or associating one set of attributes to another hidden in the dataset. These set of rules are commonly known as Association Rules. The problem of generating efficient association rules can seen as search problem since many different sets of rules are possible from a given set of instances. As the application of evolutionary computation in searching is well studied, it is possible to utilize evolutionary computation in mining for efficient association rules.
In this presentation, Evolutionary Computation and its applications will be introduced in order to realize its potential. A program known as Self-adjusting Association Rules Generator (SARG) is described. SARG is a data mining program which can generate association rules for classification. It is an improvement of the data mining program called Genetic Programming for Inductive learning (GPIL). It has been tested on several benchmark data sets available in the public domain. Comparison between GPIL and SARG revealed that SARG achieved better performance and was able to classify these datasets with higher accuracy. The presentation will also discusses relevant aspects of SARG and suggest directions for future work.
Speaker: Maria Fasli
Location: Room B02
Title: Accounting for Social Order in Multi-agent Systems
Abstract: As agents are required to inhabit and operate in increasingly complex environments and interact with other agents as elements of larger systems, some means of regulating their behaviour to avoid disruption and to ensure smooth performance, fairness and stability is needed. To this end, theoretical research and in particular agent theories need to address issues such as stable group activity, regulation of behaviour and elements that contribute to social order in multi-agent systems. In this talk we will describe a formal framework that accounts for social order within a multi-agent system based on organizational and normative concepts. The fundamental building blocks of multi-agent systems are social agents whose structure can be formally characterized in terms of roles and relationships between them including authority relations. Social agents are bound together via commitments and different types of social agents may be bound by commitments in different ways. Agents are autonomous and may decide to deviate from prescribed behaviour and violate their commitments and obligations, but as a consequence they may have to face the sanctions as a result of the other agents exercising their rights.
Speaker: Richard Booth (Joint work with T. Meyer)
Location: Room B02
Title: On the dynamics of total preorders—Revising abstract interval orders
Abstract: Total preorders (tpos) are often used in belief revision to encode an agent's strategy for revising its belief set in response to new information. Thus the problem of tpo-revision is of critical importance to the problem of iterated belief revision. In a previouswork (with T. Meyer and K-S. Wong) we provided a useful framework for revising tpos which adds extra structure to guide the revision of the initial tpo. However, one shortcoming of this framework is that ith andles *single-step* tpo-revision only. In this talk I will briefly review this framework, before giving a way to extend it to consider *double-step* tpo-revision. This extension employs a new way to represent the structure required to revise a tpo, based on *abstract interval orders*.
Speaker: Mario Paolucci
Title: Repage—Reputation and Image for Partner Selection in Simple Markets
Abstract:
Reputation is a fundamental instrument of partner selection. Developed within the domain of electronic auctions, reputation technology isbeing been imported into other applications, from social networks to institutional evaluation. Its impact on trust enforcement isuncontroversial and its management is of primary concern for entrepreneurs and other economic operators.
In this talk, I will introduce Repage, a computational system that adopts a cognitive theory of reputation, based on theoretical research performed at LABSS in the course of several years. One of the main proposals is a fundamental difference between image (agents' believed evaluation of a target) and reputation (circulating evaluation, without reference to the evaluation source), which suggests a way out from the paradox of sociality, i.e. the trade-off between agents' autonomy and their need to adapt to social environment. On one hand, agents are autonomous if they select partners based on their social evaluations (images). On the other, they need to update evaluations by taking into account others'. Hence, social evaluations must circulate and be represented as "reported evaluations" (reputation), before and in order for agents to decide whether to accept them or not.
To represent this level of cognitive detail in artificial agents' design, there is a need for a specialised subsystem, which we have developed for the public domain.
After presenting the system, I will shortly report upon simulation-based studies on the role of reputation as a more tolerant form of social capital than familiarity networks. Whereas the latter exclude non-trustworthy partners, reputation is a more inclusive mechanism upon which larger and more dynamic networks are constructed.In the simulation, we model the spreading of information in a simple market with the presence of liars and the possibility of retaliation. While fear of retaliation inhibits the spreading of image, the detached character of reputation can be a cause of inaccuracy; The two forces could balance indifferent settings.
Final remarks and ideas for future works will conclude the paper, with special attention on the choice of developing the system algorithmically and not within a formal logic. I will also shorty present the current lines of development in the "eRep—Social Knowledge for e-Governance" project.
Speaker: Marcello D'Agostino (joint work with Corrado Sinigaglia)
Title: Forecasting Accuracy and Subjective Probability
Abstract: De Finetti's favourite justification of the probability laws, within his "subjectivist" account of probability theory, was in terms of "scoring rules". These are rules that are often employed in evaluating the accuracy of probabilistic forecasters and measuring their predictive success. De Finetti showed that if a specific scoring rule is adopted, the so-called Brier's rule, consisting in taking the mean squared error over a time series of predictions, then the score of a forecaster whose predictions are in accordance with the probabilistic laws dominates that of any forecaster whose predictions violate them. If this has to be read as a subjectivist justification of the probability rules, a natural question to ask is: why Brier's rule? Why couldn't we measure forecasting accuracy by means of any other reasonable rule, such as the one based on the mean absolute error? Several attempts have been made in the literature to characterize Brier's rule as the only proper scoring rule satisfying some general, more or less compelling, properties. In this talk we take a different approach. We embed the scoring problem into the general problem of measuring the distance between two times series of predictions and present a set of natural properties that a distance function between such time series should satisfy. We then show that these properties uniquely determine a metric which, coincides with Brier's scoring rule in the special case in which one of the two series is generated by an "infallible" forecaster. We then argue that, in this way, De Finetti's subjectivist approach to the justification of the probabilistic laws can be accomplished without appealing to the, somewhat misleading and inconclusive, arguments based on the elicitation problem and on the notion of "proper scoring rule" which have so far been the main, if not exclusive, concern of the literature.
Speaker: Wojtek Jamroga
Title: A Logic for Reasoning about Rational Agents (joint work with Nils Bulling)
Abstract: Game theory identifies a number of solution concepts (e.g., Nash equilibrium, undominated strategies, Pareto optimality) that can be used to define rationality of players. Then, the notion of rationality can be employed to discard "less sensible" behavior of players, and determine what should happen had the game been played by ideal agents. Game-theoretical solution concepts have been also characterized invarious modal logics, including dynamic logic, dynamic epistemic logic, and alternating-time temporal logic (ATL). On the other hand, modal logics of time, action and strategies are a natural choice when one wants to reason about behavior of arbitrary agents. In this work, we try to bridge these two concepts, and propose a logic for reasoning about the behavior of agents under rationality assumptions. In particular, we do not want to commit to any particular notion of rationality; instead, we prefer to allow for "plugging in" rationality assumptions in a flexible way.
Speaker: Stijn Vanderlooy
Title: The ROC Isometrics approach.
Abstract:
A wide variety of state-of-the-art machine-learning classifiers areavailable to be used in practice. Nevertheless, only few classifiers areemployed in application domains with high misclassification costs, e.g.,medical diagnosis and law enforcement.
In this talk I will outline the ROC isometrics approach. The approachprovides a framework to extend classifiers such that their performancecan be set by the user prior to classification. Hence, classifiersbecome reliable. I will show that the ROC isometrics approach ispractically useful and does not have strong assumptions. An extensiveempirical evaluation of the approach will verify the analysis.
From the analysis and experimental evaluation we may conclude that theROC isometrics approach results in classifiers that can safely beapplied in any domain.
Speaker: Jerome Lang.
Title: Sequential voting rules and multiple elections paradoxes (joint work with Lirong Xia and Mingsheng Ying).
Abstract: Multiple election paradoxes arise when voting separately on eachissue from a set of related issues results in an obviouslyundesirable outcome. Several authors have argued that a sufficient condition for avoiding multiple election paradoxes is the assumption that voters have separable preferences. We show that this extremely demanding restriction can be relaxed into the much more reasonable one: there exists a linear order x1 > ... > xp on the set of issues such that for each voter, every issue xi is preferentially independent of xi+1, ..., xp given x1, ..., xi-1. This leads us to define a family of sequential voting rules, defined as the sequential composition of local voting rules. These rules relate to the setting of conditional preference networks (CP-nets) recently developed in the Artificial Intelligence literature. We study in detail how these sequential rules inherit, or do not inherit, the properties of their local components. We focus on the case of multiple referenda, corresponding to multiple elections with binary issues.
Speakers: Foued Melakessou and Zdzislaw Suchanecki.
Title: On The Road Towards The Comprehension Of The Internet Traffic Behavior: Simulation And Analysis Of An End-To-End Connection With NS-2.
Abstract: This contribution is attempting to address the problem of understanding how different aspects of networks—physical topology and routes on that topology—along with packet arrival models lead to congestion and other observed network characteristics as its fractal and periodic behavior. We reveal topology and routing effects on data transmissions between two routers linked by a connection path in a well-defined network environment. Our statistical approach is based on log-normal distributions. We show that the fractal behavior of a network node traffic can be viewed as a consequence of the topology and routingeffects. This work is based on random simulations performed by the network simulator ns-2. We have shown that the longer the connection path, the more log-normal is the inter-arrival time distribution at the destination node. Moreover a spectrum analysis has been performed in order to emphasize the traffic periodic behavior rising from the most used transfer protocol TCP/Reno.
Speaker: Glenn Lawyer.
Title: Subtypes of controls and schizophrenia patients based on brain cortical thickness.
Abstract: Given the broad range of symptoms and known brain morphological disturbances in patients with schizophrenia, it is a priori possible that subtypes of the disease could be identified in localized cortical thickness measures. Recent advances in magnetic resonance imaging allow invivo measurement of cortical thickness at each vertex in a 1mm grid covering the entire cortex. Measurements were made on 106 patients and 96 controls. At each vertex, the number of clusters of individuals best explaining the data was determined. This was done for the patient group,the control group, and the two groups combined. It was found that the patients were primarily homogeneous with regard to cortical thickness. Exceptions were found in regions where the controls were also heterogeneous. This implies that these regions of heterogeneity are common to humanity. When the method was applied to the combined data, it largely agreed with findings of group difference previously established by significance testing.
Speaker: Marcin Seredynski.
Title: Evolution of a Strategy Driven Behavior in Ad Hoc Networks.
Abstract: We propose a strategy driven approach which aims at enforcing cooperation between network participants. Each node is using a strategy that defines conditions under which packets are being forwarded. Such strategy is based on the notion of trust and activity of the source node of the packet. These way network participants are enforced to forward packets and to reduce the amount of time of being in a sleep mode. To evaluate strategies we use a new game theory based model of an ad hoc network. This model also includes a simple reputation collection and trust evaluation mechanisms. A geneticalgorithm (GA) is applied to find good strategies.
Speaker: Bram Roth.
Title: How to preserve deterrence whilst abstaining from strategies?
Abstract: Deterrence is a relation that holds between strategies of players involvedin a game. Informally one strategy is deterrent vis-a-vis another if it canreasonably be played and there are reasonable alternatives to the deterredstrategy. More formally the reasonableness of strategies is captured by a notioncalled playability. Three systematic procedures are discussed for solving gamesof deterrence by abstaining from strategies, in order to find the playability ofeach strategy. The conditions are explored under which these procedures arevalid and their intuitive interpretations are provided. Abstaining from what iscalled a harmless strategy, for instance, is shown to be valid only if theharmless strategy does not happen to be the only safe strategy available to aplayer. An efficient algorithm for solving games is proposed on the basis of theiterative application of the procedures, and ways are suggested to explore themultiplicity of playability solutions to games of deterrence.