Multi-Agent Reinforcement Learning for railway rescheduling

Malfunctions, congestions, and accidents occur in every railway system from time to time, which influences the railway traffic on a given section of the system. The disturbance may cause inconvenience for several passengers and disruption in rail freight. Both the schedule and route of the affected trains must be modified to avoid further congestion and minimalize delays. The rigidity of the railway system (e.g., single tracks, vast distances without a service station, no viable alternative in case of malfunction) poses restrictions, unlike other transportation systems. Replanning schedules and train routes (called the railway rescheduling problem) is complex and demanding, even for human operators, as one must consider numerous factors. Thus, finding a satisfying solution poses a significant challenge. This paper presents a MARL-base (Multi-Agent Reinforcement Learning) solution that shows great potential for tackling this problem, even in the case of multiple connected stations.