Reinforcement learning for train timetable rescheduling under perturbation: A general value-based approach

This paper proposes a value-based deep reinforcement learning approach that is capable of handling train timetable rescheduling under both disturbed and disrupted situations. A railway environment is constructed to simulate the problem as a Markov decision process, where the optimization objective is integrated into the reward module and various constraints are incorporated into the conflict detection and avoidance module. To address the challenges of sparse rewards and large action space with limited legal actions, a value-based algorithm framework is proposed to efficiently select and effectively evaluate actions. Through the designed simulation and training procedures, the proposed approach is tested on several disturbance and disruption cases based on a real-world instance (i.e. a Chinese high-speed railway corridor). Experimental results show that the proposed method can obtain high-quality solutions within a reasonable computing time, and also outperforms handcrafted rules in terms of the optimality of solutions. Furthermore, the proposed method exhibits promising generalization capabilities in homogeneous perturbation scenarios (disturbance scenarios and disruption scenarios that share either the same affected location and start time or the same affected location and disrupted duration).