Learning in a Black Box

We study behavior in repeated interactions when agents have no information about the structure of the underlying game and they cannot observe other agents' actions or payoffs. Theory shows that even when players have no such information, simple payoff-based learning rules eventually lead to equilibrium. Such rules have previously been documented for some forms of animal behavior, but to date they have not been tested empirically for humans. This paper analyzes human behavior in a laboratory setting and finds strong confirmation for key features of payoff-based learning that distinguish it from classical reinforcement models, and that are crucial for equilibrium convergence. By varying the amount of information we show that these features are also present even when subjects have full information about the game