Emergence Of Coordination In Evolutionary Games

This paper presents a computational model of learning which is intended to capture some basic observations of recent studies of game experiments. Furthermore it should give a satisfactory explanation for the coordinating behavior in the context of evolutionary games.In a repeated simple coordination game boundedly rational heterogeneous agents face anonymous opponents. Every player chooses one out of his collection of behavioral rules. According to this rule he assesses probabilities to the opponent's strategies and takes a mixed strategy. If the realized payoff exceeds the player's aspiration level, the average payoff, the behavioral rule is reinforced, otherwise weakened in favor of his other rules. The agents update their collections of behavioral rules via some kind of genetic algorithm.The model combines elements from replicator dynamics and fictitious play; the weighting scheme for the agents' rules follows suggestions from studies of exponential fictitious play. The game environment is chosen following the experimental settings of Van Huyck, Battalio and Rankin (1997) who investigate the role of matching protocols and labels in a coordination game experiment. Against this experimental data our model is tested. Depending on the matching protocol and the label treatment distinct levels of coordination emerge, but even within the same settings quite different courses of the experimental sessions can be observed. Our model is able to capture the qualitative properties of the experimental behavior, especially it can explain the courses in terms of varying the size of agents' memory and the experimentation rate.