Deep reinforcement learning based train door adaptive control in metro tunnel evacuation optimization

Abstract

Passenger evacuation in a metro tunnel is affected by distinct factors in the environment; thus, it is a huge challenge for both designers and managers throughout the metro system’s life cycle. A lateral evacuation platform is usually designed in a modern metro tunnel, and its evacuation efficiency is directly affected by train door opening strategy. By comparing the strategies of opening one, two, three, and four train doors, it is found that opening more doors improves evacuation efficiency. However, when two or more doors are accessed, this can lead to congestion on the evacuation platform near the train door. To reduce the congestion and improve evacuation efficiency, an adaptive train door control policy is proposed in this study, which features an adaptive dynamic programming method, namely, deep q network (DQN). The information of the environment and the state of the train doors are selected as the input and output of the control policy. By simulating scenarios under random, sequence, and DQN policies, the evacuation efficiencies are detailed. Results show that the proposed adaptive control method effectively improves evacuation efficiency. This method helps overcome the disadvantages associated with the fixed train door opening strategy and provides an optimal train door control policy during tunnel evacuation.