A shepherd and a sheepdog to guide evolutionary computation?

Abstract

Memory is a key word in most evolutionary approaches. One trend in this field is illustrated by the PBIL method that stores statistical information on the values taken by genes of best individuals. The model we are presenting follows these lines, and stores information both from good individuals (attractor memory) and bad ones (repoussoir memory). We show the interest of this model on classical binary test instances. Next we propose to test variants of this strategy for solving a non binary opti- mization problem: the traveling salesperson problem (TSP). We discuss the difficulties that one must face to tackle a non binary representation, and present a solution adapted to the TSP. We put into evidence the lack of significant differences between results from these strategies, and argue about some characteristics of the TSP search space that could explain this behaviour.