Social Learning in Recurring Games

In a recurring game, a stage game is played sequentially by different groups of players. Each group receives publicly available information about the play of earlier groups. Players begin with initial uncertainty about the distribution of types (representing the preferences and strategic behavior) of players in the population. Later groups of players are able to learn from the history of play of earlier groups. We first study the evolution of beliefs in this uncertain recurring setting and then study how the structure of uncertainty and information determine the eventual convergence of play. We show that beliefs converge over time and, moreover, that the limit beliefs are empirically correct: their forecast of future public information matches the true distribution of future public information. Next, we provide sufficient conditions to ensure that the play of any stage game is eventually close to that of a Bayesian equilibrium where players know the true type generating distribution. We go further to identify conditions under which play converges to the play of a trembling-hand perfect (Bayesian) equilibrium.