A Learning-Based Approach for Train Timetable Rescheduling With Robustness Guarantee

Train timetable rescheduling (TTR) aims to promptly restore the original train operations in the event of unexpected disturbances. In recent years, learning-based methods, as a novel technology, have been widely studied, hoping to overcome the high computational cost of traditional methods. However, to our best knowledge, timetable robustness, as another critical factor in TTR, is often ignored in existing work. To fill this gap, this study proposes a learning-based method for rapidly restoring train delays with robustness guarantees. Compared with the existing work, our approach offers several benefits. First, we design a network architecture for learning-based methods when solving the robust TTR problem. The dual-headed network can deal with the reordering and re-timing operation simultaneously in the TTR problem. Second, we reformulate the robust TTR problem into a constrained Markov decision process (CMDP), and a Lagrange-based proximal policy optimization (LPPO) algorithm is employed to train the designed network. Through testing on real-scale problems, the learned model can not only provide the rescheduled timetable within seconds but also strike a fine balance between operation efficiency and the scheme’s robustness.