A Reinforcement Learning-driven Iterated Greedy Algorithm for Traveling Salesman Problem

Abstract

This paper investigates a traveling salesman problem (TSP), which has important applications in real-world scenarios. A reinforcement learning-driven iterated greedy algorithm (RLIGA) is presented to address the TSP. A population initialization method based on the famous FRB2 heuristic is proposed to generate an initial population with high quality. To enhance the effectiveness of the RLIGA, the local search method and the destruction-construction mechanisms are designed for the city sequence. A generation method of sub-population based on current population sequence information is proposed to generate sub-population. An acceptance criterion is proposed to determine whether the offspring are adopted into the population. A re-destruction and re-construction method is proposed to avoid the proposed algorithm falling into local optimum. Lastly, the RLIGA is tested on the TSPLIB benchmark instances. The experimental results show that RLIGA is an effective algorithm to address the problem.