Comparing Reinforcement Learning and Evolutionary Based Adaptation in Population Games

In evolutionary game theory, the main interest is normally on the investigation of how the distribution of strategies changes along time and whether an stable strategy arises. In this paper we compare the dynamics of two games in which three populations of agents interact: a three-player version of matching pennies and a game with several Nash equilibria. We do this comparison by three methods: continuous replicator dynamics, an evolutionary approach, and reinforcement learning. We show how the convergence depends on the nature of the underlying method used, as well as on the pace of adjustments by the agents.