Multi-Agent Reinforcement Learning Optimization for Virtual Coupled Train Set Problem in CBTC Systems

Modern urban rail transit systems face unprecedented challenges and opportunities due to accelerated urbanization and growing travel demand, which increase the need for operational efficiency and service quality. Traditional approaches based on Communication-Based Train Control (CBTC) systems are insufficient for fluctuating passenger flows. This study introduces a Virtual Coupled Train Set (VCTS) controlled by multi-agent reinforcement learning (MARL) to address these issues. VCTS enhances capacity and efficiency by dynamically adjusting train spacing and formations through train-to-train communication and onboard sensors. The MARL framework optimizes collaborative control in high-dimensional state spaces, complex action spaces, and dynamic environments. Simulation results demonstrate that the proposed MARL algorithm significantly improves train operation efficiency and safety compared to both traditional methods and previous reinforcement learning approaches. This research advances the application of the Internet of Things in rail transit, offering innovative solutions for safer, more efficient, and intelligent urban rail systems.