Orchestrating multiagent learning of penalty games

Abstract

In comparison to single agent learning, reinforcement learning in a multiagent scenario is more challenging, since there is an increase in the space of combination of actions that may have to be explored before agents learn an efficient policy. Among other approaches, there has been a proposition to address this problem by means of biasing the exploration. We follow this track using an organizational structure where low-level agents mainly use reinforcement learning, while also getting recommendations from agents possessing a broader view. These agents keep a base of cases in order to give such recommendations, orchestrating the process. We show that this approach is able to accelerate and improve learning in penalty games, a especial case of coordination games.