Evolution towards efficient coordination in repeated games, preliminary version

We show that in long repeated games- or in infinitely repeated games with discount rate close to one- payoffs corresponding to evolutionary stable sets are asymptotically efficient, as intuition suggests. Actions played at the beginning of the game are used as messages that allow players to coordinate on Pareto optimal outcomes in the following stages. Strategies following some simple and intuitive "behavioral maxims" are shown to be able to drive out inefficient ones from a population. The result builds a bridge between the theory of repeated games and that of communication games that will be further investigated.