Graph-based Deep Reinforcement Learning Approach for UAV Multi-Warehouse Delivery Problem

With the rapid development of Unmanned Aerial Vehicle (UAV) technology, its application in logistics and distribution has become increasing widespread. Efficient task planning to minimize both time and cost has emerged as a critical challenge. However, traditional heuristic algorithms often suffer from limitations such as low search efficiency, inadequate solution accuracy, and a lack of flexibility in practical applications. To address these challenges, we propose a graph-based deep reinforcement learning (DRL) approach for solving the UAV multi-warehouse delivery problem. Firstly, the problem is formulated as a graph and subsequently transformed into a Markov decision process. Next, we design an encoder-decoder architecture, where GraphSAGE is employed as the encoder for node features embedding, and an attention mechanism is designed as the decoder to generate a feasible solution sequence step by step. Finally, we introduce a reinforcement learning approach enhanced by curriculum learning (CL) to update the model parameters, and the effectiveness and generalization of the proposed framework are verified through experiments on multi-warehouse delivery problem with varying scales.