Search

This work concerns cognitive agents, i.e. agents with mental states such as belief, goal, intention, internalized norms, etc. These mental states can be about facts of the world (including laws of action), about their own mental states, or about the mental states of other agents. Cognitive agents are able to reason about all their mental states. In what follows, we first give an overview of our contribution (Chapter 2), and then more details on the three most important parts of this contribution (Chapter 3). The first part concerns individual cognitive agents; the second part concerns social concepts (that is to say the groups of agents seen in a global way); and finally, the third part concerns the modeling of emotions. In the next chapter (Chapter 4) we list the publications that are related to each part of the previous chapter.

International audience; The 1990 papers of Cohen and Levesque (C&L) on rational interaction have been most influential. Their approach is based on a logical framework integrating the concepts of belief, action, time, and choice. On top of these they define notions of achievement goal, persistent goal, and intention. We here revisit their approach in a simplified, propositional logic, for which we give complete axiomatization. Within that logic we study the definition of achievement goals, refining C&L's analysis. Our analysis allows us to identify the conditions under which achievement goals persist. We then discuss the C&L definition of intention as well as a variant that has been proposed by Sadek and Bretier. We argue that both are too strong and propose a weakened version.

La formalisation de l'évolution des connaissances d'un agent est un problème crucial dès lors que cet agent peut percevoir des informations sur le monde dans lequel il évolue. Ces mécanismes sont d'autant plus complexe à décrire que le langage est riche. Cette contribution propose un cadre d'actes de langage simplifiés qui permet tout de même de traiter les actes nécessaires à un dialogue orienté-tâche. Pour cela, une caractérisation de la notion d'intention ainsi que de certaines lois coopératives est nécessaire, et permet d'inférer des requêtes et des questions fermées à partir d'assertions. Nous présentons les justificatifs philosophiques de ce mécanisme d'inférence, qui est formalisé dans une logique modale dynamique épistémique. Nous montrons enfin comment l'appliquer à une architecture BDI d'agent rationnel. Tout ce travail est dirigé dans le but d'obtenir une logique minimale et donc simple à mettre en oeuvre par des méthodes de preuve.

We propose a modal logic of belief and actions, where action might be nondeterministic, and there might be misperception. The agent must be able to revise his beliefs, because (contrarily to knowledge) observations might be inconsistent with his beliefs. We propose a new solution in terms of successor state axioms, which does not resort to orderings of plausibility. Our solution allows for regression in the case of deterministic actions.