A deep reinforcement learning framework for coastal multi-hazard emergency response

Coastal multi-hazard events involving storm surge, extreme rainfall, damaging winds, and waves present decision-making challenges exceeding traditional emergency management capabilities. We present a simulation-based deep reinforcement learning framework using Proximal Policy Optimization to explore whether such approaches can serve as a viable foundation for multi-hazard emergency decision support. The framework integrates multi-hazard clustering, storm-conditioned policy heads, action masking, and curriculum learning within a simulated environment that enables controlled experimentation before real-world data integration. The best learned configuration protects 99.7% of the exposed population with no failed scenarios, outperforming all non-learning alternatives, including a hazard-aware rule-based plan (97.8%) and, by a smaller margin, random action selection under the same expert masks (99.6%). Ablation tests show that learning emergency responses jointly across storm types performs better than learning them separately using storm-specific policy heads, while expert action masks are essential for maintaining life-safety outcomes. An interactive demonstration interface enables inspection of learned behaviours across scenarios spanning minor storms to severe compound hazard emergencies. We discuss pathways toward real-world testing through efficient state representations for high-dimensional operational data, AI agent orchestration for explainable decision support, and stakeholder co-design.