Abstract
Train Timetable Rescheduling (TTR) is a crucial task in the daily operation of high-speed railways to maintain punctuality and efficiency in the presence of unexpected disturbances. However, it is challenging to promptly create a rescheduled timetable in real time. In this study, we propose a reinforcement-learning-based method for real-time rescheduling of high-speed trains. The key innovation of the proposed method is to learn a well-generalized dispatching policy from a large amount of samples, which can be applied to the TTR task directly. At first, the problem is transformed into a multi-stage decision process, and the decision agent is designed to predict dispatching rules. To enhance the training efficiency, we generate a small yet good-quality action set to reduce invalid explorations. Besides, we propose an action sampling strategy for action selection, which implements forward planning with consideration of evaluation uncertainty, thus improving search efficiency. Extensive experimental results demonstrate the effectiveness and competitiveness of the proposed method. It has been proven that the local policies trained by the proposed method can be applied to numerous problem instances directly, rendering it unnecessary to use human-designed rules.