Heuristique pour l’apprentissage automatique décentralisé d’interactions dans des systèmes multi-agents réactifs. Heuristics for automatic decentralized interaction learning in reactive multi-agent systems

This article proposes an heuristics for the automatic computation of interactions in reactive multi-agent system. It describes first the Interac-DEC-POMDP formalism and then focuses on a algorithm based on reinforcement learning techniques and distribution of individual rewards among agents to build interaction and collective behaviour in a sub-class of DEC-POMDP. These techniques produce adaptative collective behaviour based on the use of direct