Coordination via genetic learning

This paper uses genetic algorithms to model the solution to a coordination problem through learning. The coordination problem arises out of the following two stage game. At core of the two stage game is a model of social interaction in which players use visible goods as signals about the identity of other players. If these signals are informative enough, players use them to condition their social interaction. Importantly, accurate signals are mutually beneficial. This game is then wrapped in another in which players choose their visible goods. There are many types of players and many visible goods that could be used to signal type. There are many equilibria of the two stage game, some of which allow individuals to perfectly signal their type in all social interactions, and others of which do not. The perfect signaling equilibria Pareto dominate the others, but since there are many of them, the players face a difficult coordination problem. We approach this coordination problem using genetic algorithms to simulate learning in this game. A player is a genetic code that evolves via selection.Questions of primary interest concern the set of parameter values such that the players manage to solve the coordination problem. The results of simulations indicate that the convergence of the genetic algorithm to a perfect signaling equilibrium depends on the values of the parameters that determine players' payoffs.We analyze two different scenarios. In one scenario, each player makes one (unconditional) decision to either use the visible goods displayed by other players as type signals, or to ignore the visible goods they display. In the other scenario, the decision to regard a player's visible good as a type signal is conditional on the good displayed. Our results indicate that it is easier for players to solve the coordination problem under the second scenario.