A Deep Reinforcement Learning Approach for the Traffic Management of High-Speed Railways

In high-speed railway systems, unexpected disruptions may cause the delays of multiple trains and greatly affect the service quality to passengers. Our study proposes a deep reinforcement learning (DRL) approach for rescheduling the trains in case of disruptions. Specifically, in our DRL framework, the states are defined as positions of trains in the network, the actions are defined as the possible routes (e.g. going straight, using the side tracks, waiting for other passing trains, etc), and the reward functions are denoted by the delay time and possible conflicts according to the specific track structures. In addition, we develop a deep learning based value function approximation technique combined with a greedy algorithm, in order to further improve the training efficiency of the deep neural network. We use the Beijing-Zhangjiakou high-speed railway network as the simulation environment and conduct several sets of experiments. Our results demonstrate that the developed DRL can avoid possible conflicts and further reduce the train delay time compared with greedy algorithm.