Railway timetables have an important role in efficient and punctual railway operations. In particular, the robustness of the timetable has direct impact on the traffic's punctuality. To evaluate the robustness of a timetable, simulation is commonly used. A simulation study may indicate that some trains are too sensitive against minor delays, which may lead to that they fall out of their planned channel of operations (defined by their surrounding trains). We define this as robustness vulnerabilities of the timetable. In research, optimization-based methods are often used for improving timetable robustness. However, based on our previous experience, it seems that some robustness vulnerabilities may not be resolved by such methods. This work explores reinforcement learning (RL) as a method to resolve this. Which in the future can be used in combination with optimization-based methods or as a standalone method.

In this project, the objective is to find a new schedule for a single train in a given timetable without adjusting other trains. Our aim is that the optimized timetable should be sufficiently robust to allow the train to run within its planned channel of operations despite minor disturbances. To address this problem, we formulate a RL-based method for the robust railway timetabling problem and use the Actor-Critic method to train an agent. The timetables generated by the agent are compared with timetables generated using optimization-based methods from our previous work. The models are evaluated using microscopic RailSys simulation.

In our preliminary experiments we have seen that the RL method is able to generate feasible timetables with higher punctuality than the original timetable. The improvement comes with a cost in terms of longer travel times, which is in line with our previous research.

The method in this project is still in an early stage, and it is therefore not yet ready for deployment. However, the results are interesting, and we have identified several ways forward to further improve the method. By further improving the method, it seems possible that it can support planners in the future to construct more robust timetables.