Optimizing Rail Traffic with Reinforcement Learning and Graph Attention Networks

Optimizing rail traffic and handling disturbances in the network are essential to minimizing delays and improving system throughput. The high computation burden of conventional solutions can be shifted to training time using machine learning techniques, particularly reinforcement learning (RL), to ensure real-time applicability. In this paper, we combine RL with Graph Attention Networks (GANs) to optimize traffic in multi-line railway networks. By leveraging graph-based representation and attention mechanisms, our model enhances feature extraction by assigning greater importance to traffic elements contributing more to delays. Our approach dynamically adapts both routes and train passing sequences in real-time, enabling more effective disturbance management and improved throughput.