Train rescheduling method based on multi-agent reinforcement learning

Many multi-agent pathfinding algorithms have been raised to arrange trains' scheduling effectively and have reached excellent results. However, these algorithms usually focus on the fixed schedule and have a poor ability to deal with dynamic problems. This paper presents a train rescheduling method based on multi-agent reinforcement learning. A new observation is adopted for trains to better interact with the environment and other trains. The improved DQN network is implemented to train to obtain the best performance, such as avoiding conflicts, handling trains' breakdowns and generating new paths. According to simulation results, the model achieved an aggregate completion rate of over 70% of ten agents after training. Compared with the traditional multi-agent pathfinding algorithm CBS, this method was 20% higher in terms of completion rate when the malfunction rate was over 20%. Conclusively, the method has better handled unexpected situations and has excellent adaptability to problems such as sudden train breakdowns.