ADAPTIVE DISASTER RELIEF ROUTING USING DEEP REINFORCEMENT LEARNING UNDER DYNAMIC ROAD DISRUPTIONS

Natural disasters such as floods frequently disrupt transportation networks, significantly hindering emergency relief logistics and delaying the delivery of critical supplies. Traditional routing optimization methods, including shortest-path algorithms and metaheuristic approaches, typically assume static network conditions and struggle to adapt to rapidly evolving road disruptions. This paper proposes a Deep Reinforcement Learning (DRL) framework for adaptive disaster relief routing in dynamically changing environments. The problem is formulated as a Dynamic Vehicle Routing Problem (DVRP) modeled using a Markov Decision Process (MDP). A grid-based city simulation with nine demand nodes and stochastic road blockages is developed to emulate flood-induced transportation disruptions.Two DRL algorithms, Deep Q-Network (DQN) and Proximal Policy Optimization (PPO), are evaluated and compared against a Genetic Algorithm (GA) baseline. Experimental results demonstrate that the PPO-based routing policy achieves improved adaptability and reduced delivery delays under high disruption scenarios, outperforming traditional optimization methods. Furthermore, the DRL agents exhibit faster recovery and policy adjustment following sudden road blockages. The findings highlight the potential of reinforcement learning–based decision systems for real-time disaster logistics planning and adaptive emergency response.