Optimizing Humanitarian Logistics with Deep Reinforcement Learning and Digital Twins

This paper explores the optimization of humanitarian logistics, with a focus on the Multi-Commodity Network Flow (MCNF) problem in this setting. Our primary goal is to enhance the efficiency of aid distribution, minimizing long-term transportation costs while managing the complexities of demand and supply in disaster/emergency situations. To accomplish this, we suggest a deep reinforcement learning (DRL) method that incorporates graph neural networks to approximate the value function. Also, our approach involves the use of digital twins (DTs) for accurate modeling and simulation, reflecting the dynamic and stochastic nature of humanitarian logistics. Our computational experiments include a comparative analysis against traditional deterministic and heuristic methods. We examine the performance of our DRL approach across simulated MCNF problem instances. The results indicate DRL agent’s capability in optimizing logistics tasks, and the incorporation of DTs and DRL demonstrates effectiveness and adaptability in managing the humanitarian logistics.