Apprentissage par renforcement pour la conception de systèmes multi-agents réactifs

A new reinforcement learning (RL) methodology for the design of reactive multi-agent systems is presented. Although dealing with realistic situated agents with local perception does not belong to the framework where convergence of RL algorithm is guaranted, in our method each agent learns individually its local behavior. The progressive aspect of learning, which pits the agents against more and more complex sub-tasks, allows to go beyond the classical limitation of RL in this context. Our general framework is validated on a simulated environment where the agents have to coordinate themselves to reach a global goal. MOTS-CLÉS : processus décisionnels de Markov, agents réactifs, apprentissage par renforcement, apprentissage progressif, coordination, coopération