Reinforcement Learning based Train Rescheduling on Event Graphs

This paper proposes a Reinforcement Learning (RL) formulation of the Train Timetable Rescheduling problem at a microscopic scale under unexpected events, particularly track blockages, with the goal of minimizing the overall delay of trains. We represent the operational timetable by means of Timed Event Graphs accounting for operational constraints and interdependencies among trains. We combine RL with Graph Neural Networks to make rescheduling decisions, such as reordering and re-routing of trains. The approach is evaluated on a medium-scale corridor of the German railway network, which includes multiple stations, tracks, and trains subject to various disruptions. The experiments validate the effectiveness of our RL formulation offering improved solution quality, practical applicability, and notable time efficiency compared to baselines, including First-In-First-Out and Branch-and-Cut.