Modeling Seismic Resilience and Hospital Evacuation: A Comparative Analysis of Multi-Agent Reinforcement Learning and Classical Evacuation Models

Hospitals in earthquake-prone regions must evacuate heterogeneous occupants rapidly while preserving operational continuity under disrupted conditions. However, many hospital-evacuation studies still rely on static routing assumptions or narrowly defined behavioral rules, which limits their value for building-level resilience planning. This paper develops a comparative hospital-campus evacuation framework that combines GIS-based geodesic routing, heterogeneous agent-based modeling, and reinforcement-learning-based decision policies. Puge County People’s Hospital in Sichuan, China, is used as the case study. Six algorithms are evaluated: three rule-based baselines—Shortest Path (SP), Random Walk (RW), and the Social Force Model (SFM) —together with a training-free density-aware heuristic, Density-Aware Gradient Routing (DAGR), and two reinforcement-learning approaches, Density-Aware Q-Learning (DAQL) and SARSA. Experiments cover three population scales (N∈50, 100, 200), normal daytime conditions, staffing-variation scenarios, and a blocked-exit disruption scenario, with 30 independent runs for each main condition. The results show that the rule-based and training-free methods remain the most reliable under full multi-agent evaluation: the SFM and RW achieve the highest completion ratios (approximately 100% and 93. 5%, respectively), while DAGR provides the strongest balance between completion and evacuation efficiency among the non-trained methods. In contrast, the trained RL agents perform substantially worse in direct multi-agent deployment with DAQL reaching approximately 37% completion and SARSA approximately 17%, highlighting a train–evaluation distribution shift associated with independent Q-learning. The ablation analysis further shows that collision avoidance is the most critical reward component, whereas density-avoidance shaping can unintentionally induce collective deadlock when all agents execute the learned policy simultaneously. Among the enhanced variants, DAQLRoleAware yields the best overall improvement, increasing the completion ratio to approximately 52% and reducing the 90th-percentile evacuation time to approximately 363 s. Overall, this paper clarifies both the promise and the present limitations of density-aware reinforcement learning for hospital evacuation while providing a more building-centred and reproducible basis for future coordination-aware evacuation design and emergency-planning research.