Emergence of cooperation in two-agent repeated games with reinforcement learning

Cooperation is the foundation of ecosystems and the human society, and the reinforcement learning provides crucial insight into the mechanism for its emergence. However, most previous work has mostly focused on the self-organization at the population level, the fundamental dynamics at the individual level remains unclear. Here, we investigate the evolution of cooperation in a two-agent system, where each agent pursues optimal policies according to the classical Q-learning algorithm in playing the strict prisoner’s dilemma. We reveal that a strong memory and long-sighted expectation yield the emergence of Coordinated Optimal Policies (COPs), where both agents act like “Win-Stay, Lose-Shift” (WSLS) to maintain a high level of cooperation. Otherwise, players become tolerant toward their co-player’s defection and the cooperation loses stability in the end where the policy “all Defection” (All-D) prevails. This suggests that tolerance could be a good precursor to a crisis in cooperation. Furthermore, our analysis shows that the Coordinated Optimal Modes (COMs) for different COPs gradually lose stability as memory weakens and expectation for the future decreases, where agents fail to predict co-player’s action in games and defection dominates. As a result, we give the constraint to expectations of future and memory strength for maintaining cooperation. In contrast to the previous work, the impact of exploration on cooperation is found not be consistent, but depends on composition of COMs. By clarifying these fundamental issues in this two-player system, we hope that our work could be helpful for understanding the emergence and stability of cooperation in more complex scenarios in reality.