A Deep Reinforcement Learning Based Real-Time Solution Policy for the Traveling Salesman Problem

The rapid development of logistics and navigation has led to increasing demand for solving route optimization problems in real-time. The traveling salesman problem (TSP) tends to require fast and reliable online solutions, which may not be met by traditional iterative optimization algorithms. In this work, a real-time solution policy is proposed for TSP. The idea is to build a mapping between city information and optimal solutions using deep neural networks. Therefore, when given a new set of city coordinates, the optimal route can be directly and quickly calculated without iteration. Considering the recent advancement in computer vision with deep convolutional neural networks (DCNNs), an image representation is proposed to convert TSP to a computer vision problem. A problem decomposition method is introduced to reduce the mapping complexity. Taking advantage of the powerful fitting capabilities of DCNN, a deep reinforcement learning method is designed without any labeling requirement. The proposed method is superior for real-time applications compared with other algorithms.