Evolving Cooperation in Complex Behavioral Interactions through Reputation

The iterated prisoner’s dilemma (IPD) has long been used to study the conditions that promote cooperative behaviors among selfish individuals. In particular, studies using the co-evolutionary learning framework have shown that cooperative behaviors can be learned through a process of adaptation of strategies based solely on direct interactions through repeated encounters in playing IPD. However, complex behavioral interactions (e.g., by humans) may involve more than just direct interactions between two individuals. In many real-world situations, it is impossible to interact with all other individuals. It is very important to study indirect interactions, e.g., in the form of estimating behaviors of future partners through their perceived reputation based on previous behaviors to other individuals. Here, we study the co-evolutionary learning of IPD with reputation where behavioral interactions for a pair of strategies depend not only on choices made in their previous moves, but also choices made to other strategies that are reflected by their reputation scores. We show that for more complex IPD interactions involving more choices where direct interaction alone is ineffective to promote cooperation, the addition of reputation helps to promote cooperation. We further show that different implementations of reputation estimation, which reflect the accuracy of reputation estimation from using memory of games from previous generation and more frequent updating of reputation scores, is for the evolution of cooperation. Finally, we also investigate the situation where strategies can misperceive their partner’s reputation and show that even in the circumstances that strategies misperceive reputation, evolution to cooperation is still possible. This implies that reputation helps to increase the robustness of co-evolutionary learning of cooperative strategies.