Reinforcement Learning Approach to Velocity and Position Control of Metro Trains

Velocity and position control for metro trains is typically achieved by classical control methods (PID, etc). Challenges in this control problem include imprecise position sensing, time delay, and external disturbances due to weight changes, curves, and slopes of the rails. In order to achieve acceptable stop position of the trains at each station, the controller design often involves individual gain tuning for each sections in the route, which consumes much time and effort. As a means to reduce the effort, reinforcement learning approach is looked into for train control. Automatic Train Operation (ATO) simulator capable of realistic simulation of train dynamics along the Line 5 in Seoul Metro is used to investigate the feasibility of this approach. Results are discussed from the perspective of practicality.