Improving graph convolutional networks with reinforcement learning for pure electric logistics vehicle routing

Electric logistics vehicle route planning plays a crucial role in the operations of urban distribution service providers. By optimizing vehicle routes, not only can significant reductions in operating costs be achieved, but also enhanced customer satisfaction. However, the vehicle path problem is a classic NP-hard problem in the field of combinatorial optimization. Therefore, finding an efficient and accurate solution holds great practical significance. In response to the real-time operational requirements, this paper proposes a novel approach that improves the graph convolutional neural network model with reinforcement learning. This approach aims to learn effective strategies for generating vehicle route schedules. Through comprehensive experiments conducted on real geographic data, several key findings have been obtained: 1) Explicitly considering edge features within the model is essential; 2) Employing a joint learning strategy can expedite training convergence and improve solution quality, and 3) Comparative evaluations against existing algorithms in the literature demonstrate the superior performance of our proposed model.