A diversified tour-driven deep reinforcement learning approach to routing for intelligent and connected vehicles

The routing problem for intelligent and connected vehicles has garnered significant attention because of its profound theoretical implications and wide-ranging practical applications. Despite advancements, existing learning-based methods often rely on training one single policy, which inadequately explores the solution space and leads to suboptimal performance. To address this limitation, we propose a diversified tour-driven deep reinforcement learning (DT-DRL) approach for solving vehicle routing problems (VRPs) across various scales. Our approach builds on the encoder‒decoder paradigm, with the encoder utilizing a multihead attention mechanism to derive informative node embeddings and a gate aggregation block to enhance state representation. During decoding, dynamic-aware context embedding is designed to capture real-time state transitions and graph variations, thereby offering comprehensive and timely information for decision-making. To promote solution diversity and expand the search space, multiple decoders with independent parameters are employed, coupled with a Kullback–Leibler divergence-based cross-entropy loss that regularizes the generation of diversified candidate tours. We validate the proposed DT-DRL through extensive experimentation on two representative routing problems for intelligent connected vehicles, namely, the traveling salesman problem (TSP) and the capacitated VRP (CVRP). The results demonstrate that DT-DRL consistently outperforms many heuristic and DRL-based methods, achieving up to a 7.54% improvement in the optimality gap, thereby establishing its effectiveness and robustness in tackling complex routing challenges for intelligent and connected vehicles.