Multi-objective optimisation for drone allocation and routing in humanitarian aid delivery: a reinforcement learning-based approach

Delivering humanitarian relief during natural disasters is a challenging and crucial task that calls for effective resource allocation to ensure timely assistance to the impacted areas. Recently, UAVs have emerged as a promising solution for rapid last-mile aid delivery in hard-to-reach places. This study presents a multi-objective mixed integer programming (MIP) framework to plan aid distribution by optimal routing and allocation of UAVs. Using a weighted sum approach, the proposed model balances multiple objectives, including aid delivery, priority aid delivery, locations served, and response times, providing decision-makers with practical and actionable insights. Two reinforcement learning-based solution approaches, Q-learning and Deep Q-learning (DQN), are used to solve the model. A case study based on the 2018 Kerala floods in India demonstrates the comparison of both learning-based solution approaches with the exact solution method. The experimental results highlight that the exact method achieves optimal solutions for small instances in less computational time. In contrast, the Q-learning approach outperforms the exact method and DQN approach in terms of computational efficiency for medium and large-scale scenarios. The DQN algorithm provides optimal solutions for most data instances comparable to the exact solution method. Furthermore, several managerial insights are derived by conducting sensitivity analysis.