Dynamic Graph Network with Spatial-Temporal Gated Attention: A Deep Reinforcement Learning Approach for Urban Logistics Optimization

With the development of smart transportation in cities, the dynamic vehicle routing problem has become increasingly important in urban logistics distribution, which requires optimizing vehicle routes in a dynamic environment where customer orders fluctuate in real-time and traffic conditions vary, aiming to achieve efficient resource utilization and cost minimization. Therefore, this paper focuses on solving the urban logistics distribution problem with dynamic traffic flow. The optimization objective is to minimize total travel time using a deep reinforcement learning method with dynamic graph network. In order to enhance the model's adaptability to dynamic real-time traffic conditions, gated temporal self-attention is introduced into the network. Meanwhile, distance spatial self-attention is incorporated into the decoder to improve the extraction of spatial structures among customers. The model is trained using the soft actor-critic algorithm and validated through experiments on real-world street traffic data. The results demonstrate that incorporating spatial-temporal gated self-attention in the model significantly improves the model's adaptability to dynamic traffic conditions while maintaining computational efficiency, leading to higher solution quality.