Lifting the performance of a heuristic for the time-dependent travelling salesman problem through machine learning

Abstract

In recent years, there have been several attempts to use machine learning techniques to improve the performance of exact and approximate optimization algorithms. Along this line of research, the present paper shows how supervised and unsupervised techniques can be used to improve the quality of the solutions generated by a heuristic for the Time-Dependent Travelling Salesman Problem with no increased computing time. This can be useful in a real-time setting where a speed update (or the arrival of a new customer request) may lead to the reoptimization of the planned route. The main contribution of this work is to show how to reuse the information gained in those settings in which instances with similar features have to be solved over and over again, as it is customary in distribution management. We use a method based on the nearest neighbor procedure (supervised learning) and the K-means algorithm with the Euclidean distance (unsupervised learning). In order to show the effectiveness of this approach, the computational experiments have been carried out for the dataset generated based on the real travel time functions of two European cities: Paris and London. The overall average improvement of our heuristic over the classical nearest neighbor procedure is about 5% for London, and about 4% for Paris.