Multi-agent reinforcement learning in stochastic single and multi-stage games

Abstract

In this paper we report on a solution method for one of the most challenging problems in Multi-agent Reinforcement Learning, i.e. coordination. In previous work we reported on a new coordinated exploration technique for individual reinforcement learners, called Exploring Selfish Reinforcement Rearning (ESRL). With this technique, agents may exclude one or more actions from their private action space, so as to coordinate their exploration in a shrinking joint action space. Recently we adapted our solution mechanism to work in tree structured common interest multi-stage games. This paper is a roundup on the results for stochastic single and multi-stage common interest games.