Reinforcement learning with experience sharing promotes collective cooperation on higher-order networks

As a central form of prosocial behavior, collective cooperation is the cornerstone of social prosperity, and elucidating how it emerges among the self-interested population is of significant theoretical and practical importance. However, existing research of cooperative evolution on higher-order networks typically assumes that agents update decisions exclusively based on their own experiences, overlooking the pervasive diffusion of information within groups. In reality, decision-making is shaped not only by individual experience, but also by experiences acquired from neighboring agents. To this end, we develop a hypergraph-based reinforcement learning framework that integrates individual and neighborhood experiences to investigate how information sharing affects cooperation under higher-order interactions. Specifically, we utilize hypergraphs to characterize the interaction topology, within which agents update their strategies according to the Q-learning algorithm. Extending the standard learning paradigm, we introduce an experience sharing mechanism, where neighborhood information is aggregated to form a shared experience and subsequently fused with individual experience to guide decision-making. Through extensive simulations, we demonstrate that an intermediate level of experience sharing can significantly enhance the collective cooperation, particularly under severe social dilemmas, increasing the final cooperation level by up to approximately 0.43. In contrast, excessive experience sharing leads to behavioral homogenization among agents, thereby undermining the relative advantage of cooperation. Moreover, higher learning rates coupled with moderate discount factors are most conducive to cooperation. These findings shed new light on how experience sharing shapes cooperative dynamics in higher-order networks.