Coordination of independent learners in cooperative Markov games

In the framework of fully cooperative multi-agent systems, independent agents learning by reinforcement must overcome several difficulties as the coordination or the impact of exploration. The study of these issues allows first to synthesize the characteristics of existing reinforcement learning decentralized methods for independent learners in cooperative Markov games. Then, given the difficulties encountered by these approaches, we focus on two main skills: optimistic agents, which manage the coordination in deterministic environments, and the detection of the stochasticity of a game. Indeed, the key difficulty in stochastic environment is to distinguish between various causes of noise. The SOoN algorithm is so introduced, standing for “Swing between Optimistic or Neutral”, in which independent learners can adapt automatically to the environment stochasticity. Empirical results on various cooperative Markov games notably show that SOoN overcomes the main factors of non-coordination and is robust face to the exploration of other agents.