An efficient encoder-decoder network for the capacitated vehicle routing problem

The capacitated vehicle routing problem (CVRP) is of great importance to intelligent transportation systems. In recent, deep reinforcement learning (DRL) approaches have shown great potential in solving the CVRP efficiently. Specifically, encoder-decoder frameworks are trained via reinforcement learning with different schemes to construct solutions incrementally. As the total customer demands and remaining vehicle capacity are dynamic with time steps, it is still a challenge for this kind of methods to obtain optimal solutions. In this work, we develop an efficient encoder-decoder framework, termed the residual graph convolutional encoder and multiple attention-based decoders (RGCMA), which is trained by a reinforcement learning method with an elite baseline. The encoder produces powerful node representations while being dedicated to aggregating neighborhood features by a fitted dense residual edge and node features updating block. Compared to the popular single decoder strategy, our multiple decoders mechanism incrementally constructs a variety of solutions for any CVRP instance, which diversifies solution space, and eventually improves the solution quality. Extensive experiments demonstrate that RGCMA performs competitively with existing methods on variety CVRP datasets. RGCMA narrows the gap to LKH solver on four benchmark tasks, and its run time is less than 1 s. It also exhibits good generalization ability on large-scale tasks, well-known CVRPLIB dataset and real-world Jingdong Logistics distribution task.