Planification multi-agents dans un cadre markovien : les jeux stochastiques à somme générale

Planning agent’s actions in a dynamic and uncertain environment has been exten-sively studied. The framework of Markov decision process provides tools to model andsolve such problems. The field of game theory has allowed the study of strategic inter-actions between multiple agents for a given game. The framework of stochastic games isconsidered as a generalization of the fields of Markov decision process and game theory.It allows to model systems with multiple agents and multiple states.However, planning in a multi-agent system is considered difficult : agent’s decisionsdepend not only on its actions but also on actions of the other agents. The work presentedin this thesis focuses on decision making in distributed multi-agent systems. Existingworks in this field allow the theoretical resolution of stochastic games but place severerestrictions and ignore some crucial problems of the model. We propose a decentralizedplanning algorithm for the model of stochastic games. Our proposal is based on the Value-Iteration algorithm and on the concept of Nash equilibrium. To improve the resolutionprocess and to deal with large problems, we sought to ease decision making and limit theset of joint actions at each stage. The proposed algorithm was validated on a coordinationproblem including several agents and various experiments were conducted to assess thequality of the resulting solution.