Semi-Hierarchical Deep Reinforcement Learning for Autonomous Railway Rescheduling
A new research paper published on arXiv introduces a semi-hierarchical deep reinforcement learning (RL) approach to address the Vehicle Rescheduling Problem (VRSP) in railway operations. Managing disruptions in increasingly dense rail networks is complex, often relying on human expertise due to the limitations of traditional Operational Research methods and existing RL models. This study proposes a novel formulation that separates dispatching from routing using dedicated action and observation spaces, allowing policies to specialize in distinct decision scopes. Evaluated on the Flatland-RL simulator with 7 to 80 trains across various difficulty levels, the method demonstrates significant improvements over heuristic baselines and monolithic RL approaches. Results indicate nearly double the number of trains reaching their destinations, with deadlock rates maintained below 5%. The system effectively handles heavy congestion by adaptively sequencing, delaying, or cancelling trains, offering a robust solution for real-time autonomous railway traffic management and improved resource utilization.
Wire timeline
Semi-Hierarchical Deep Reinforcement Learning for Autonomous Railway Rescheduling
A new research paper published on arXiv introduces a semi-hierarchical deep reinforcement learning (RL) approach to address the Vehicle Rescheduling Problem (VRSP) in railway operations. Managing disruptions in increasingly dense rail networks is complex, often relying on human expertise due to the limitations of traditional Operational Research methods and existing RL models. This study proposes a novel formulation that separates dispatching from routing using dedicated action and observation spaces, allowing policies to specialize in distinct decision scopes. Evaluated on the Flatland-RL simulator with 7 to 80 trains across various difficulty levels, the method demonstrates significant improvements over heuristic baselines and monolithic RL approaches. Results indicate nearly double the number of trains reaching their destinations, with deadlock rates maintained below 5%. The system effectively handles heavy congestion by adaptively sequencing, delaying, or cancelling trains, offering a robust solution for real-time autonomous railway traffic management and improved resource utilization.
cs.AI updates on arXiv.org