Self-Adapting Payoff Matrices in Repeated Interactions

摘要

Traditional iterated prisoner's dilemma (IPD) assumed a fixed payoff matrix for all players, which may not be realistic because not all players are the same in the real-world. This paper introduces a novel co-evolutionary framework where each strategy has its own self-adaptive payoff matrix. This framework is generic to any simultaneous two-player repeated encounter game. Here, each strategy has a set of behavioral responses based on previous moves, and an adaptable payoff matrix based on reinforcement feedback from game interactions that is specified by update rules. We study how different update rules affect the adaptation of initially random payoff matrices, and how this adaptation in turn affects the learning of strategy behaviors.