Small groups nurturing collective wisdom: The punishment-prediction reinforcement learning mechanism for multi-group cooperation

Understanding the evolution of collective cooperation is a core issue in multi-agent systems research, as modern interdependent systems require efficient mutual collaboration among agents. However, most previous works focused on single-group institutions. Even with some works on multi-group structures, they still remained limited within inter-group imitation learning, which requires the assumption of stable environments. Considering the dynamic characteristics of real-world environments, we turn to reinforcement learning frameworks to enable agents to dynamically adjust their strategies based on their own experience and long-term payoffs. In this paper, we proposes a novel reinforcement learning mechanism based on punishment-prediction (PunPreRL) to study cooperation behavior within the multi-group structure, enabling agents to dynamically adjust their strategies based on their own experiences and long-term payoffs. By integrating punishment and prediction, we construct a decision-optimization learning process under multi-group interactions. Furthermore, through pairwise approximation modeling of multi-group dynamics, we provide a theoretical foundation for understanding the cooperation evolution under the PunPreRL mechanism. Extensive simulations demonstrate that the PunPreRL mechanism significantly enhances group cooperation levels and stability across varying subgroup sizes, subgroup structures and initial cooperation conditions, thereby verifying the philosophy of “small groups nurturing collective wisdom”. Comparative experiments further confirm that the PunPreRL mechanism outperforms isolated prediction or punishment mechanisms, underscoring the critical role of double-mechanism synergy.