Deep Reinforcement Learning for Adaptive Transportation Logistics Routing During Natural Disasters and Infrastructure Failures

Abstract

Natural disasters and cascading infrastructure failures disrupt transportation networks precisely when reliable logistics routing is most critical for delivering relief supplies, medical equipment, and evacuation support. Static shortest-path and classical vehicle-routing heuristics assume a largely intact network and adapt poorly to rapidly evolving road closures, congestion, and demand surges. This study presents an investigation of deep reinforcement learning (DRL) for adaptive, real-time logistics routing under disaster conditions, culminating in a proposed graph-attention multi-agent proximal policy optimization architecture (GAT-MAPPO). A synthetic disaster-response environment was constructed as a 60-node transportation network subject to stochastic edge disruption, congestion, and time-varying relief-demand surges representing earthquake, flood, hurricane, and combined multi-hazard scenarios. Six baseline routing strategies - static Dijkstra, A* heuristic, Clarke-Wright savings vehicle routing, deep Q-network (DQN), Double DQN, and asynchronous advantage actor-critic (A3C) - were benchmarked against proximal policy optimization (PPO), soft actor-critic (SAC), and the proposed GAT-MAPPO agent across 500 evaluation episodes per method. GAT-MAPPO achieved the highest demand fulfillment rate (93.7%) and lowest mean delivery delay (11.2 min), outperforming the strongest baseline (SAC, 85.1% fulfillment) and the static-routing baseline (58.2% fulfillment) by wide margins. The proposed agent retained a demand fulfillment rate above 80% at an edge-failure rate of 30%, generalized across all four disaster-scenario types with less than eight percentage points of performance variation, and exhibited graceful degradation under simulated sensor and communication noise. Multi-objective analysis mapped Pareto-optimal trade-offs between delivery time and fleet fuel consumption, and between responder risk exposure and delivery speed, while an ablation of the reward function confirmed that urgency weighting and failed-delivery penalties contributed most to overall performance. These results, illustrate a coherent workflow for developing, benchmarking, and stress-testing DRL-based disaster-logistics routing systems prior to real-world deployment.