Learning correlated equilibria: An evolutionary approach

Correlated equilibrium (Aumann, 1974, 1987) is an important generalization of the Nash equilibrium concept for multiplayer non-cooperative games. In a correlated equilibrium, players rationally condition their strategies on realizations of a common external randomization device and, as a consequence, can achieve payoffs that Pareto dominate any of the game's Nash equilibria. In this paper we explore whether such correlated equilibria can be learned over time using an evolutionary learning model where agents do not start with any knowledge of the distribution of random draws made by the external randomization device. Furthermore, we validate our learning algorithm findings by comparing the end behavior of simulations of our algorithm with both the correlated equilibrium of the game and the behavior of human subjects that play that same game. Our results suggest that the evolutionary learning model is capable of learning the correlated equilibria of these games in a manner that approximates well the learning behavior of human subjects and that our findings are robust to changes in the specification and parameterization of the model.