THÈSE Pour l’obtention du grade de DOCTEUR D’UNIVERSITÉ Discipline: INFORMATIQUE

The goal of this thesis is to study the issue of rational BDI learning agents, situated in a multi-agent system. A rational agent can be defined as a cognitive entity endowed with intentional attitudes, e.g., beliefs, desires, and intentions (BDI). First, we study the concepts of agency and practical reasoning, allowing agents to induce from their intentional attitudes, a behavior identified as rational. These concepts are formalized using a multi-modal BDI logic, enabling us to specify and develop a BDI architecture implemented in Lisp. The architecture is fully operational, it allows the execution of multiple BDI agents in a multi-agent system and it improves the epistemic reasoning of each agent introducing first-order theorem proving competences. Then, we study the problem of learning in the agency context and we propose a hierarchy of learning situations based on the awareness of agents about the presence of other agents in the multi-agent system. We use the first-order induction of logical decision trees to enable our BDI agents to learn the “reasons” they have to form an intention. These reasons correspond in the architecture to the trigger conditions of the plans of the agent. This learning method is included in the architecture proposed, allowing BDI agents to learn from their own experience, as if they were isolated (first level of the hierarchy proposed), and communicating with other agents in the Multi-agent System, to profit in this way of the experience of others (second level in our hierarchy).