An experimental study of different approaches to reinforcement learning in common interest stochastic games

Stochastic (a.k.a. Markov) Games pose many unsolved problems in Game Theory. One class of stochastic games that is better understood is that of Common Interest Stochastic Games (CISG). CISGs form an interesting class of multi-agent settings where the distributed nature of the systems, rather than adverserial behavior, is the main challenge to efficient learning. In this paper we examine three different approaches to RL in CISGs, embedded in the FriendQ, OAL, and Rmax algorithms. We show the performance of the above algorithms on some non-trivial games that illustrate the advantages and disadvantages of the different approaches.