Velocity Regulation for Automatic Train Operation via Meta-Reinforcement Learning

In this paper, we consider the velocity regulation for automatic train operation to maintain the velocity of the train at a desired value. Due to complicated railway conditions and uncertain dynamics of the system, the problem cannot be well solved by most of the model-based controllers. To this purpose, we formulate the velocity regulation problem as a sequence of stationary Markov decision processes (MDP) with unknown transition probabilities. Based on the meta-learning framework, we propose a model-free algorithm to learn an adaptive controller, which only requires a "small" amount of sampled data from the corresponding MDP. We illustrate with simulations that our model-free controller performs well and can well adapt to the dynamical environments.