AML Talk

Coordination becomes a real challenge to reinforcement learning agents when more than one equilibrium strategy exists. In this paper, we propose a learning model (optimal fictitious learning) for solving this problem in cooperative multi-agent systems. Under the framework of identical interest stochastic games, OFL allows agents to assess the optimality of their joint actions according to the convergence rate of underneath learning algorithm. Optimal coordination can be achieved through fictitious play within the estimated optimal action set. A theoretical analysis is presented in the paper to study the optimal convergence of the approach. Examples are also given to empirically explore its convergence speed.