Building individual behaviors to solve collective problems is a major stake whose applications are found in several domains. DecPOMDP has been proposed as formalism for describing multi-agent problems. However, solving a Dec POMDP turned out to be a NEXP problem. In this study, we introduced the original concept of social action to get round the inherent complexity of DecPOMDP and we proposed three decentralized reinforcement learning algorithms which approximate the optimal policy in DecPOMDP. This article analyses the results obtained and argues that this new approach seems promising for automatic top-down collective behavior computation.