Knowledge-guided hybrid deep reinforcement learning for the dynamic multi-depot electric vehicle routing problem

In this paper, we consider the dynamic multi-depot electric vehicle routing problem with time windows, proposing a novel hybrid framework, integrating knowledge-guided multi-agent deep reinforcement learning (MARL) and a variable neighborhood search (VNS) algorithm. The MARL component employs a double-deep Q-network for initial route generation, which is further refined by the VNS to enhance solution quality. Real-time decision-making and adaptive optimization enable the framework to respond effectively to dynamic changes in the environment, leading to improved efficiency, reduced costs, and enhanced overall performance. Extensive experiments on both synthetic and real-world benchmark datasets, demonstrate the framework’s superiority over state-of-the-art algorithms, showing significant improvements in total traveled distance, computation time, and scalability. The results indicate over 70% reduction in the average total traveled distance compared to state-of-the-art baselines on small-scale datasets. Importantly, the framework’s ability to handle large-scale problems effectively makes it a promising solution for real-world applications.