Edge Encoded Attention Mechanism to Solve Capacitated Vehicle Routing Problem with Reinforcement Learning

The capacitated vehicle routing problem (CVRP), which is referred as NP-hard problem is a variant of Traveling Salesman Problem (TSP). CVRP constructs the route with the lowest cost without violating vehicle capacity constraints to meet demands of customer nodes. Following the advent of artificial intelligence and deep learning, the use of deep reinforcement learning (DRL) to solve CVRP is giving promising results. In this paper we proposed DRL model to solve CVRP. The transformer-based encoder of our proposed model fuses node and edge information to construct a rich graph embedding. The proposed architecture is trained using proximal policy optimization (PPO). Experiments using randomly generated test instances show that the proposed model gives rise to better results in comparison with the existing DRL methods. In addition, we also tested our model on locally generated real-world data to verify its performance. Accordingly, the results show that our model has a good generalization performance for both of random instance testing to real-world instance testing.