Explore the scholarly record.

Abstract

Metro train timetable rescheduling needs to balance real-time responsiveness, solution quality, and generalization across network scales and operating conditions. Existing studies generally address disturbances and disruptions separately, although both commonly occur within the same operating period. We formulate the rescheduling problem as both a Decentralized Partially Observable Markov Decision Process (Dec-POMDP) and a mixed-integer linear programming (MILP) model. Based on these formulations, we propose When2Ask, a hierarchical decision framework that combines multi-agent reinforcement learning (MARL) for low-latency train-level control with MILP for joint system-level optimization. When2Ask activates the strategic layer when infrastructure restrictions, policy uncertainty, or passenger waiting indicate a need for coordinated decisions. The strategic layer integrates a large language model (LLM), a small language model (SLM), and the MILP solver. The LLM retrieves historical cases and specifies the focus region, planning horizon, and computational budget, while the SLM, trained through supervised fine-tuning (SFT), generates solver guidance. The MILP uses this guidance to determine feasible actions for selected trains, while the remaining trains continue under MARL control. Experiments on Shenzhen Metro Lines 20 and 1 show that the proposed method achieves the smallest reward gaps among all online methods (11.04% and 7.18%), with average decision times of 1.01s and 2.59s, corresponding to approximately 3.7-fold and 55.2-fold speedups over rolling-horizon MILP, respectively. The method also remains effective under cross-network transfer and unseen infrastructure-restriction and passenger-demand patterns without additional training. These results demonstrate that the proposed framework approaches the offline global optimization reference while preserving online responsiveness and generalization.