An adaptive graph reinforcement learning method for scalable multi‐train cooperative control

Multi‐Train Optimal Control (MTOC) addresses the cooperative control problem of multi‐trains running on railway tracks through centralized or distributed controllers. However, two critical challenges emerge in solving MTOC problems: (1) the dynamic system dimensionality caused by time‐varying train numbers during station arrivals and departures and (2) the strong inter‐train command correlations in dense traffic scenarios. These complexities lead to computational challenges when scaling to extended railway networks with growing train populations, rendering conventional rule‐based methods ineffective. To address these challenges, we propose Graph Attention Soft Actor‐Critic (GASAC), a novel graph reinforcement learning algorithm integrating two core components: (1) A graph attention network (GAT) for efficient information aggregation from high‐dimensional train observations, and (2) A Soft Actor‐Critic (SAC) architecture serving as the centralized decision‐maker. The GAT module performs dimensionality reduction through feature attention mechanisms, effectively supporting the SAC module in deriving optimal control policies. Comparative evaluations against multi‐agent deep reinforcement learning baselines demonstrate that GASAC successfully synthesizes distributed train information to generate control commands, ensuring collision‐free and on‐time operations. Further sensitivity analysis shows the adaptability of the algorithm to different parameters.