ILIKS SCIENTIFIC PROGRAMME 2008-2011
The ILIKS scientific programme for modelling interacting knowledge systems is
organized in four complementary topics.
The first topic is dedicated to the notion of cognitive agent, with its mental attitudes,
its reasoning to take decisions and plans, and its acting proper. The second topic is
dedicated to the fundamentals of multi-agent systems and to the social dimension of
interaction; it focuses on social relationships, on the notions of collective agents and
organizations, and on services. Interaction is modelled there in game-theoretical
terms. The third topic is dedicated to models of linguistic interaction, an essential
modality of interaction, focussing on the specificities of communicative actions and
on the structure of discourse and dialogue. The last and fourth topic is dedicated to the
lexical and knowledge resources to support interaction with artificial knowledge
The core of the research program, modelling interaction, is therefore dealt with in
topics 2 and 3. However, addressing it cannot be done without the fundamental
studies of topic 1, and developing concrete interacting systems requires the
infrastructure investigated in topic 4.
The tight interrelationship of these four topics is proven by the fact that most ILIKS
research groups are involved in several of them; there are even ILIKS members who
work in all the four.
- Cognitive agents: mental attitudes, decisions, actions and plans.
Modelling cognitive agents, artificial or not, is addressed by comparing and
establishing synergies between cognitive, social, economic, philosophical and logical
approaches. Work in this topic aims at understanding the nature of agents' mind,
especially mental attitudes, and the rules that control their coherence and dynamics.
Specifically, this addresses:
- ontology and general theories of mind, mental states and emotions. A current
challenge for developing theories of emotions and affective phenomena is to
develop formal (in particular, logical) frameworks which enable to specify and
characterize complex emotions like regret, relief, shame, guilt, reproach,
remorse, pride, envy, jealousy. These involve very sophisticated forms of
reasoning such as reasoning about beliefs, desires, goals and intentions of self
and others, selfattribution of responsibility, counterfactual reasoning,
reasoning about capabilities, and reasoning about norms and ideals.
- reasoning about belief and knowledge; the problems of belief change. We will investigate the computational issues of multiple forms of belief change in multiagent settings, based on our expertise on single-agent belief change (belief revision, update, reasoning about action) and on the models of multiagent belief change, especially dynamic logic. Other important issues to be addressed are: assessing the relative descriptive adequacy of competing Bayesian models of confirmation (or inductive strength); working out new and more satisfactory formal models of inductive reasoning; and studying conditional inferences under conditions of full uncertainty, in particular Jeffrey's rule as a descriptive model. We expect that this rule will prove to be descriptively inadequate, and plan to develop simple models to explain behavior.
- models of agents' intentions, goals and commitments. We shall address the following important questions: (i) on the basis of which beliefs do we arrive from activated goals to decisions and thus to formulate an 'intention', and how do 'intentions' guarantee some persistence and stability in goal-directed behavior (thanks to 'commitment' or other tricks)? (ii) how do we revise intentions due to new 'temptations' or after revising pertinent beliefs? (iii) what is the relationship between the "intention to do" (some action) and the "intention that" (some result holds in the world)? (iv) what is the difference between 'instrumental' goals and 'aims' or 'motives'? this is especially relevant when a goal is "motivating" and not just an expected positive result of the action; (v) how does the intention define and characterize the "action", even conveyed by just one and the same movement?
- reasoning about preferences, choices and decisions. The dynamics of preference (unlike the dynamics of beliefs) is an almost unexplored area and appears most promising. We will try to establish strong connections between preference change and belief change, which will help us rationalizing preference change and classifying the various forms of preference change processes. In particular, a current challenge is to provide a cognitive model capable of explaining well-known forms of time inconsistency in rational agents (time-dependent preference change, hyperbolic discounting, impulsivity, procrastination, etc.), and to analyze its implications for the theory of bounded rationality in human decision-making. We also aim here at studying the influence of incidental emotions and of various individual differences (e.g. the ability to solve simple arithmetic problems) on the perceived value of money (e.g. in a commercial offer).
As agency is based on the capacity of agents to perform actions, representing and reasoning about actions and plans is necessary as well. Work in this topic therefore also deals with:
- ontologies, theories and logics of intentional action, attempt, and agency. To obtain a complete account of action, a current challenge is to integrate three families of logics: those dealing with goals and intentions (BDI logics), those dealing with agency (esp. STIT logics) and those dealing with action effects (dynamic logics); steps in this direction have been taken, but no fully integrated formalism is yet available. A subsequent challenge is to combine such a reasoning framework with theories representing the eventual dimension (temporal and causal structure) of each action instance.
- the reasoning dimension of plans or sequences of future actions, in particular from the perspective of partial observability, with work on goal dependencies and goal conflicts and on planning in multi-agent environments. We will build new models, and then algorithms, for planning in partially observable, multi-agent domains, where agents may perform physical and epistemic actions as in single-agent planning, but also communication actions.
- Social relationships, game-theoretical interaction, collective agents and organizations, services .
Beyond the individual agent and its private mental states, we address the social dimension and the various notions of collective agent, here again combining several approaches in several disciplines. This provides the basis for a first family of models of interaction.
- Social reality is based on a number of social relationships creating links and dependencies between agents, such as trust, responsibility, obligation, delegation, reputation or cooperation, which are studied in cognitive sciences and philosophy and modelled within logical theories. An important aspect is the modelling of social norms and other conventions regulating relations between agents, for instance in terms of deontic logics.
- Security (of computer systems or networks, of transactions, of knowledge flow and processes within organisations, etc.) is an important application domain in which such social relationships are crucially involved.
- Interaction proper is then approached in game-theoretical terms. Relations between game-theory on the one hand and agency and action theories, studied in 1.e, on the other hand need to be addressed. Indeed, a challenge in the area of formal methods for multi-agent systems is to establish a clear connection between logics of agency and cooperation (ex. STIT, ATL, Coalition Logic) and game theory. In particular, are these logics sufficiently expressive to capture game-theoretic concepts? How can these logics be extended in order to meet this objective? Social reality notions seen just above are also involved in the theoretical and experimental study of social strategic games in cognitive psychology and economics. In this area, we aim at exploring, both from a theoretical and empirical point of view, mechanisms of trust and reciprocity in strategic interaction. We focus our attention to Prisoner's Dilemma-like interactions in which the implementation of self-interested actions requires some form of social approval. More precisely we want to study if the introduction of a peer-monitoring system is likely to mediate the choices of the agents, working as a social norm. In parallel, and still within this frame, we plan to investigate another psychological dimension of coordination within games. In particular we will investigate how the activation of mortality concerns affects both individual decision-making and strategic behavior in economic games. The objective is to assess the processes that link mortality to rationality, in terms of its impact on the goals and on the performances of the decision maker. A final objective is to study the interplay between emotions and analytical processes in the context of strategic choices. We expect to find that the more "primitive" reaction to an unfair offer is to accept.
- It is then necessary to study the specificities of collective agents ("we-attitudes", collective action and collective decision) and the institutional dimension of social reality, characterizing various kinds and structures of social groups, institutions, organizations and legal entities, e.g. in terms of roles and processes. We intend to analyze here to what extent it is possible to exploit a parallel with the design of artefacts and to conceive an organization as a multi-agent entity whose structure is determined by its plan to reach its goals. More formally, we intend to explore how the logics of ability, of (multi-agent) action and of delegation can be integrated to represent multi-agent plans at different levels of abstraction that correspond to the refinement steps of the design. Another challenge for the logical theories of normative systems and organizations is to explain how the dynamics of institutional facts (rules, norms, etc.) depend on how these facts are perceived, accepted, recognized by the members of the institution. There are few logical and formal works that have tried to provide an explanation of the relationships between the cognitive and social level (made of agents and groups of agents, and their individual and collective attitudes) and the institutional level (made of abstract concepts like rule, norm, etc.). Regarding group decision making, we will build axiomatic and computational models, thus contributing to the burgeoning field "computational social choice"; we will focus especially on the role that preference representation and logics play for group decision making, especially for fair division and voting on multi-issue domains.
- Finally the study and modelling of services and service composition, with applications spanning from web services, to processes internal to enterprises and administrations, and to e-government services for the citizen, requires the combination of these social concepts and structures with the notions of actions, processes and plans studied in topic 1. In fact, the current challenge is that of providing a solid foundation for a notion of service that could constitute the center of the future services science. The vision that should be built is an holistic one, in which services will be represented as complex systems of interrelated events involving entities of various nature (human and artificial agents, organizations, etc.) whose behavior and interaction is constrained by predefined commitment.
- Linguistic Interaction
This topic aims at the construction of formal models of linguistic interaction and communication and the development of implemented dialogue systems.
Formal models of linguistic interaction are being developed following two main families of approaches. The first are produced by the evolution of the formal semantic linguistic tradition to take into account discourse and dialogue. The second consist of specializations of multi-agent models developed above:
- Approaches focusing on the linguistic contents and structure of communication model the semantics and pragmatics of discourse and dialogue. They focus on their rhetorical structure (discourse relations, hierarchical structure, topic segmentation), the conventional rules governing turn-taking and possible speech acts, anaphora and grounding phenomena, and the interaction between attentional structure -including visual attention- and reference. Annotation in discourse and dialogue is an important step to both test theoretical hypotheses and to design discourse analysis systems following statistic approaches. Obtaining an as theoretically neutral as possible annotated corpus of discourse structure is a challenge we are currently addressing. This also implies to seriously address the coverage extension of discourse theories like Segmented Discourse Representation Theory (SDRT).
- The second family of approaches are mainly concerned with tracking the evolution of agents attitudes in dialogue, and analyze the interrelationships between speech acts and agents' intentions, beliefs, and emotions, considered singularly or collectively. This first implies the specialization and enrichment of general theories of action to characterize the basic actions involved in communicative interaction, i.e., speech acts and other epistemic actions. Then, dedicated logical frameworks are developed such as Public announcement logic, based on dynamic and BDI logics.
- A current challenge is to make these two lines of study meet, as they are concerned with complementary and inter-dependent facets of interaction. For instance, one of the objectives in the language domain is to look at making dialogue models less beholden to Gricean assumptions of cooperativity and to study strategic conversation in which dialogue agents' preferences (or utilities) may be at odds with each other. To this end, we need to integrate into a model of dialogue: a way of modeling the dynamic interaction between dialogue moves, commitments and agents' attitudes, preference elicitation from dialogue moves and the use of game theoretic techniques to compute best responses. We will address this first with the integration of Public Announcement Logics and SDRT. Another part of the investigations concerning strategic reasoning in dialogue will be preference representation and elicitation.
- The cognitive and multi-modal / multi-channel dimension of face-to-face communication and language is mainly addressed through the study of signed languages (SL), with a major focus on Italian (LIS) and French (LSF) SL, and of coverbal gestures in spoken communication by hearing people. The main challenge here is to define more appropriate models and tools for the representation of SL and coverbal gestures, a prerequisite for adequate analyses and descriptions of the multidimensional structure of face-to-face discourse in human language, and the extent to which it is or not influenced by the primary modality in which language is perceived and expressed (i.e. auditory-vocal in spoken languages vs. visual-gestural in SL). This will eventually bring us to develop interfaces in signed languages.
- Developing implemented dialogue systems is also part of this topic, in particular interactive question-answering systems in a multimodal dialogue framework. One challenge in this domain is to transform written procedures into oral, cooperative dialogue situations.
- Lexical and knowledge resources .
Communicative interaction is effective only when mutual understanding of the contents of the information exchanged is assured; in the case of communication between artificial agents this needs to be enforced. The
Semantic Web has imposed the use of computational lexicons and
ontologies (more than often limited to taxonomies), but system interaction --in other words system interoperability-- remains problematic.
- The last topic in the ILIKS research program thus first addresses the relationship between lexical semantics, cognition and formal ontology, for well-founded computational lexicons as well as for semantic negotiation and coordination.
- Techniques of ontology learning from texts and relation extraction are used to automatically build consensus ontologies.
- Formal ontology methods are then applied to "clean up" and improve the coherence of the resources thus obtained. A challenge which is still painfully open is to do the same in a thorough manner on standard lexical resources (WordNet and FrameNet), whose errors block the efficiency of many of their applications, including the tasks of ontology learning and relation extraction described in 4b themselves.
- Interoperability is also improved using ontology-based techniques of query reformulation, document annotation and categorization, and document collection browsing. These techniques will derive a benefit from the results established in 4.a : ontologies are all the more relevant for these purposes that their relationship with linguistic items is well grounded.
- Meaning negotiation and evolution issues are investigated to understand and represent the processes of negotiation, stabilization or modification of meaning.
- Finally, interoperability is not only impaired by differences in lexica or ontologies being used, but also by the difficulties of grounding, i.e., making sure the entities referred to are correctly identified. This issue is addressed through the development of a new kind of resource, an entity name system contributing to the
web of entities. Important research areas are: entity representation, entity matching, schema matching. Part of this work will be done in collaboration with the FP7 EU-funded OKKAM project .