coordination, minority game, market eciency, information, self-organization, reinforcement learning

C72, C91, C92

In minority games, players in a group must decide at each round which of two available options to choose, knowing that only subjects who picked the minority op- tion obtain a positive reward. Previous experiments on the minority and similar congestion games have shown that players interacting repeatedly are remarkably able to coordinate eciently, despite not conforming to Nash equilibrium behavior. We conduct an experiment on a minority-of-three game in which each player is a team composed by three subjects. Each team can freely discuss its strategies in the game and decisions must be made via a majority rule. Team discussions are recorded and their content analyzed to detect evidence of strategy co-evolution among teams playing together. Our main results of team discussion analysis show no evidence sup- porting the mixed strategy Nash equilibrium solution, and support a low-rationality, backward-looking approach to model behavior in the game, more consistent with reinforcement learning models than with belief-based models. Showing level-2 ratio- nality (i.e., reasoning about others' beliefs) is positively and signicantly correlated with higher than average earnings in the game, showing that a mildly sophisticated approach pays off. In addition, teams that are more successful tend to become more egocentric over time, paying more attention to their own past successes than to the behavior of other teams. Finally, we nd evidence of mutual adaptation over time, as teams that are more strategic (i.e., they pay more attention to other teams' moves) induce competing teams to be more egocentric instead. Our results contribute to the understanding of coordination dynamics resting on heterogeneity and co-evolution of decision rules rather than on conformity to equilibrium behavior. In addition, they provide support at the decision process level to the validity of modeling behavior using low-rationality reinforcement learning models.