Event-triggered Reinforcement Learning to Obtain Stable Train Operation Strategies

Energy-efficient train operation is a crucial issue in railway systems. Traditionally, dynamic programming has been used to optimize energy efficiency and solve operational control problems. However, dynamic programming has limitations due to the need to model the operational environment and scalability issues arising from the increasing number of variables. In contrast, the Deep Reinforcement Learning (DRL) approach is model-free, does not require a prior model of the environment, and can derive optimal operational control rules from arbitrary operating conditions by using function approximation to avoid scalability issues. Nevertheless, when DRL is applied to real problems like railway systems, challenges such as “sparse rewards” that can be set and “control instability” caused by sequential control based on the gradient method persist. This paper proposes an event-triggered reinforcement learning method to address these two problems and verify its effectiveness through computational experiments.