Parameter prediction based on features of evolved instances for ant colony optimization and the traveling salesperson problem

Abstract

Ant colony optimization performs verywell onmany hard optimization problems, even though no good worst case guarantee can be given. Understanding the reasons for the performance and the influence of its different parameter settings has become an interesting problem. In this paper, we build a parameter prediction model for the Traveling Salesperson problem based on features of evolved instances. The two considered parameters are the importance of the pheromone values and of the heuristic information. Based on the features of the evolved instances, we successfully predict the best parameter setting for a wide range of instances taken from TSPLIB.