A Reinforcement Learning-Based Iterative Method for Capacitated Hub Location Problems in UAV Networks

Though unmanned aerial vehicles (UAV) have recently attracted widespread attention and demonstrated significant potential in logistics, they are limited by their flight range and the need to recharge or replace batteries. One way to address this challenge is to design an efficient delivery network, deploy suitable charging stations or hub facilities. In this paper, a capacitated hub location problem is proposed for UAV delivery networks and formulated using the mixed integer linear programming with the objective of minimizing the total costs. To solve this problem, a novel reinforcement learning (RL) based iterative algorithm is developed, with the essence of boosting solutions iteratively via adaptive operator selection by the RL agent. Further, multiple graphs and node-specific feature representations are constructed to serve as inputs for specialized policy networks, which are architecturally based on Graph Neural Networks and Gated Recurrent Units (GRUs), facilitating the embedding of both demand and spatial patterns inherent in the solutions. Additionally, the actor-critic architecture Proximal Policy Optimization is employed as the training algorithm. Evaluation results across extensive simulation instances indicate the superior performance of the RL-based iterative algorithm in comparison with baseline methods and demonstrate its robust generalization capabilities across networks of various scales.