Skip to Main Content
Automatic train regulation, which is a core function of the signaling system, concerns the headway/schedule adherence that dominates the transport capacity and punctuality of a metro line. The main difficulty in synthesizing a traffic regulator is that an accurate traffic model is inaccessible. This paper presents an adaptive optimal control (AOC) algorithm that can approximate the optimal traffic regulator by learning traffic data with artificial neural networks. The AOC algorithm is derived from the discrete minimum principle and organized in the critic-actor architecture of reinforcement learning to carry out sequential optimization forward in time. The critic network receives no signal from the traffic model so that the prediction of the future cost and the optimization of the traffic regulator are not biased by modeling errors. The efficacy of the AOC algorithm in the traffic regulation is verified in a simulated system using traffic data acquired from a real metro line.