Coevolution of finite automata with errors

We use a genetic algorithm to simulate the evolution of error-prone finite automata in the repeated Prisoner’s Dilemma game. In particular, the automata are subjected to implementation and perception errors. The computational experiments examine whether and how the distribution of outcomes and genotypes of the coevolved automata change with different levels of errors. We find that the complexity of the automata is decreasing in the probability of errors. Furthermore, the prevailing structures tend to exhibit low reciprocal cooperation and low tolerance to defections as the probability of errors increases. In addition, by varying the error level, the study identifies a threshold. Below the threshold, the prevailing structures are closed-loop (history-dependent) and diverse, which impedes any inferential projections on the superiority of a particular automaton. However, at and above the threshold, the prevailing structures converge to the open-loop (history-independent) automaton Always-Defect (ALLD). Finally, we find that perception errors are more detrimental than implementation errors to the fitness of the automata. These resultsshow that the evolution of cooperative automata is considerably weaker than expected.