Joint Optimization of Passenger Flow Control and Train Skip-Stopping for Overcrowded Metro Lines: A Multi-Agent Reinforcement Learning Approach

Abstract During rush hours, the capacity of metro in megacities is insufficient to meet the travel demand, resulting in oversaturation and high risk on platform in stations, especially transfer stations. This paper addresses this problem through the joint optimization of some operational interventions, aiming to alleviate passenger overloads while maintaining travel efficiency. To make the model more realistic, the stochastic characteristics of passengers are considered, including the probability distribution of passenger arrival time, inbound and transfer walking times. To provide a high-quality solution for the complex constraint model, three cooperative agents—governing passenger inflow, transfer flows, and train skip-stopping mode—are architected within improved Double Deep Q learning Network (IDDQN) to form a multi-agent reinforcement learning solution. Empirical validation on Beijing Metro Line 13 and Changping Line demonstrates that the multi-agent framework proposed in this paper can eliminate 100% of passenger over-limit flow while reducing the average waiting time of passengers. It also has a significant improvement in reducing stochastic characteristic impact and accelerating convergence.