Reinforcement and imitation learning for real-time train trajectory re-optimization

Rail transport is a key lever for reducing carbon emissions, offering a more sustainable alternative to individual cars and trucks for both passenger and freight transportation. However, the increased adoption of rail services is hindered by disruptions such as network failures and the resulting uncertainty in arrival times. While mathematical solvers are widely used in the railway industry for train scheduling and real-time re-optimization during disruptions, their performance deteriorates as the scale and complexity of the network grows. In this work, we propose a hybrid approach that combines reinforcement learning (RL) and imitation learning (IL) to accelerate train trajectory re-optimization. Our method leverages world models to create a compact and actionable representation of the railway environment, enabling efficient long-term planning. To facilitate coordination between multiple trains, we integrate a communication mechanism based on Transformers, allowing agents to exchange information and adapt to changing network conditions. Additionally, we use imitation learning from a mathematical solver as the expert,enabling the model to replicate the solver’s decision patterns while significantly speeding up the decision-making process. This hybrid approach aims to combine the reliability of mathematical optimization with the adaptability and speed of artificial intelligence, offering a scalable solution to improve railway efficiency and reduce the impact of disruptions.