Environment Representations of Railway Infrastructure for Reinforcement Learning-Based Traffic Control

The real-time railway rescheduling problem is a crucial challenge for human operators since many factors have to be considered during decision making, from the positions and velocities of the vehicles to the different regulations of the individual railway companies. Thanks to that, human operators cannot be expected to provide optimal decisions in a particular situation. Based on the recent successes of multi-agent deep reinforcement learning in challenging control problems, it seems like a suitable choice for such a domain. Consequently, this paper proposes a multi-agent deep reinforcement learning-based approach with different state representational choices to solve the real-time railway rescheduling problem. Furthermore, comparing different methods, the proposed learning-based approaches outperform their competitions, such as the Monte Carlo tree search algorithm, which is utilized as a model-based planner, and also other learning-based methods that utilize different abstractions. The results show that the proposed representation has more significant generalization potential and provides superior performance.