Regularized Best Responses and Reinforcement Learning in Games

Abstract

We investigate a class of reinforcement learning dynamics in which each player plays a “regularized best response” to a score vector consisting of his actions’ cumulative payoffs. Regularized best responses are single-valued regularizations of ordinary best responses obtained by maximizing the difference between a player’s expected cumulative payoff and a (strongly) convex penalty term. In contrast to the class of smooth best response maps used in models of stochastic fictitious play, these penalty functions are not required to be infinitely steep at the boundary of the simplex; in fact, dropping this requirement gives rise to an important dichotomy between steep and nonsteep cases. In this general setting, our main results extend several properties of the replicator dynamics such as the elimination of dominated strategies, the asymptotic stability of strict Nash equilibria and the convergence of timeaveraged trajectories to interior Nash equilibria in zero-sum games.