Beyond cyclic dominance: Reinforcement learning promotes cooperation in the spatial rock–paper–scissors game

Cyclic dominance is widely recognized as a fundamental mechanism for sustaining cooperation in structured populations. However, existing studies primarily employ strategy updating mechanisms based on imitation, neglecting the role of individual learning from experience. In this study, we investigate the spatial rock–paper–scissors game with destructive agents on a two-dimensional lattice network, incorporating reinforcement learning as the strategy updating mechanism. Specifically, we examine two distinct learning paradigms: self-regarding Q-learning, where individuals update their strategies based solely on personal experience, and neighbor-regarding Q-learning, which integrates both individual and neighbor experiences. Monte Carlo simulation results demonstrate that when Q-learning governs strategy updates, cyclic dominance disappears, leading the system to converge toward a single dominant strategy. The specific dominant strategy depends on the adopted learning paradigm: when individuals rely exclusively on their own experiences, destructive agents ultimately dominate the system. In contrast, when individuals incorporate experiences from their neighbors, full cooperation emerges. These findings suggest that the significance of cyclic dominance in promoting cooperation may have been overestimated. Instead, cooperation can be maintained through experiential learning, as long as the learning perspective is not overly constrained.