An Interactive Neighbor-Aware Double DQN Framework for Enhancing Cooperation in Multi-Agent Games

Cooperation is essential for enhancing the performance of complex systems. However, promoting sustained and efficient collaboration in networked evolutionary games remains challenging. Considering the importance of reinforcement learning (RL) and neighbor information in fostering cooperation, we propose a general interactive neighbor-aware double deep Q-network (double DQN) framework to explore its effects on cooperative rates in evolutionary games. This framework is suitable for static, symmetric two-player games, offering strong scalability and applicability to various game models. Using the classic game model—the prisoner's dilemma game on a grid network—as the experimental scenario, we conducted multiple comparative experiments to evaluate the effects of different reward information sources in double DQN, including (1) only using self-reward information without neighbor-reward information, (2) using both self and neighbor reward information, and (3) only using neighbor-reward information without self-reward information. The experimental results demonstrate that the interactive neighbor-aware double DQN framework significantly enhances cooperation in complex systems.