RNN-Attention Hybrid Model to Solve Vehicle Routing Problem

Abstract

This paper presents a Recurrent Neural Network (RNN)-Attention hy-brid model for the Vehicle Routing Problem (VRP) through Reinforcement Learning (RL). The VRP, one of the well-known Combinatorial Optimization (CO) problems, minimizes the total distance traveled with some constraints. However, in online situations where Urban Air Mobility (UAM) is used, routes cannot be obtained quickly with conventional methods. The RL is a powerful heuristic solver that obtains suboptimal solutions even in complex situations. We propose an improved RL model combining Long-Short Term Memory (LSTM), a kind of RNN, and the attention model to solve the routing domain problems. The proposed framework provides trajectory information through the LSTM and non-visited nodes. We validated the proposed model by testing on Traveling Salesman Problem (TSP) and Capacitated Vehicle Routing Problem (CVRP) with different problem sizes and randomly generated instances. The proposed model performs better than the existing model across tested problem instances.