A multi-agent reinforcement learning scheduling algorithm integrating state graph and task graph structural modeling for ride-sharing dispatching

Urban-scale ride-hailing dispatch faces critical challenges such as heterogeneous demand density, highly dynamic state transitions, and multi-agent coordination. Traditional rule-based or heuristic matching strategies struggle to maintain efficiency under large-scale spatiotemporal distributions. This paper proposes DualG-MARL, a graph-attentive multi-agent reinforcement learning framework that employs dual-path modeling of vehicle state graphs and task graphs. The framework extracts spatial structural features via multi-order diffusion kernels and introduces a feasibility mask combined with a Top-K filtering mechanism for cross-graph matching, thereby enhancing both decision-making efficiency and assignment quality. Empirical evaluations on real-world order datasets from Manhattan and Queens demonstrate that the proposed method outperforms the current state-of-the-art approach, CoopRide, by reducing the Average Waiting Time (AWT) by 0.27 and 0.35 minutes, increasing the Order Response Rate (ORR) by 2.3% and 2.7%, improving Vehicle Utilization Rate (VUR) by 3.5% and 3.9%, and lowering the Average Detour Ratio (ADR) by 0.05 and 0.06, respectively. These results establish new benchmarks in core dispatching metrics, and show that the proposed method maintains high responsiveness while effectively reducing matching redundancy and idle travel, offering a structure-aware paradigm for large-scale urban mobility systems.