Apprentissage de la coordination multiagent. Une méthode basée sur le Q-learning par jeu adaptatif

Current algorithmes on multiagent learning are for almost limited since they cannot manage the multiplicity of Nash equilibria and thus converge to the Pareto-optimal. To alleviate this, we propose here a learning mechanism extending the Q-learning to non-cooperative stochastique games. This learning mechanism converges to Pareto-optimal equilibria in self-play. We present experimental results showing convergence of such learning mechanism. We then extend our approach to the case of non-stationarity of agents which is another important aspect of multiagent systems. Finally, we tackle the question of non stationarity in multiagent environments in its generality and we present in this context some research avenues which can lead to improve our preliminary results on adaptation.