Intelligent Train Tracking Control Based on Monte Carlo Tree Method

Train tracking distance control is critical to ensure safety and enhance the transportation capacity of urban rail transit. The widely used control methods relying on the dynamics model of trains may not achieve ideal performance because of the inaccuracy of the dynamics model and delay of control commands execution, especially in the Urban Rail Transit scenarios. To overcome these limitations, this paper presents an intelligent tracking control method NN-MCTS combining a deep learning model with Monte Carlo Tree Search (MCTS) to cope with the tracking distance minimization. The deep learning model adopts a neural network (NN) to characterize the relationship between control command and actual train dynamic, while MCTS regards the control problem as a decision-making process leveraging the trajectory output of NN to obtain the reward of the corresponding control sequence. We extracted 20000 actual state trajectory segments of trains of Chengdu Metro Line 8 to train the deep-learning model, and additionally continuous trajectories as the known state of the train ahead to verify the effectiveness of the search algorithm. Experiments show the advantages of the proposed method in accuracy and stability compared with PID and real-time compared with MPC method.