Online Multi-modal Evacuation during Passenger Flow Outburst in Urban Transit System: A Heterogeneous Multi-agent Reinforcement Learning Framework

With growing demand straining urban transit systems’ resilience in managing outburst passenger flows, existing approaches focused on offline and single-modal evacuations remain limited. This study proposes an online multi-modal evacuation framework that coordinates on-duty taxis, buses, and metros while minimizing impact on their regular services. We develop a data-driven agent-based environment to update multi-modal transit data and stranded passenger information in real time. Two coordination strategies are introduced: (1) an independent strategy using a distributed training and distributed execution algorithm, and (2) a collaborative strategy using a hybrid centralized training and distributed execution algorithm. To dynamically assess evacuation effectiveness, we design a resilience framework with three metrics: robustness, rapidity, and resourcefulness. These metrics are transformed into demand-responsive feedback at each time step, enabling agents to proactively generate resilient evacuation plans. In a real-world case study triggered by a railway disruption, our approach outperforms genetic algorithms and multiagent deep deterministic policy gradient algorithms in computation time and solution quality under offline conditions. Simulated new environments further validate its online applicability, demonstrating its potential for real-world deployment.