Traffic Flow Optimization for Rail Transit Based on Multi-Agent Reinforcement Learning for Network-Based Train Dispatching

Abstract

Since the urban rail transit has developed rapidly, train scheduling and traffic flow optimization is a challenging task for improving the operational efficiency and service quality in developing networks and complicated networks. We can overcome the limitations of traditional centralized methods towards coping with massive-scale, high-density settings using multi-agent reinforcement learning (MARL) in network-based train dispatching, building a decentralized decision space where trains are treated as smart agents to optimize dispatching cooperatively through interaction with the environment. The paradigm maximizes traffic best, conserves computational complexity, and improves real-time disruption adaptability, achieving impact like reduced train delays and better energy efficiency. Experimental results indicate that the suggested MARL solution reduces average train delay (ATD) to 59% compared to FIFO during heavy-traffic scenarios, improves throughput (TP) to 148 trains/hour, lowers energy consumption (EC) by 11%, and lowers stability index (SI) by 43% without sacrificing near-real-time computation time (CT) to less than 1.5 seconds. The experimental results provide effective suggestions to intelligent rail transit dispatching and propose new directions for constructing intelligent transportation systems.