Search Method of Number of Trees for Genetic Programming with Multiple Trees

Abstract

Genetic programming (GP), which is an evolutionary computational method, is known to be effective for agent problems because individuals are represented by a tree structure. As an extension method, GP with control nodes (GPCN) has been proposed. Because one individual has multiple tree structures, GPCN can efficiently evolve and obtain highly readable behavioral rules. However, the number of trees suitable for each problem has to be manually adjusted in advance and cannot be easily applied various problems. In the previous study, a method for automatically determining the number of trees have proposed. However, because the method of the previous study changes the fitness function and uses a special population, it cannot be combined with the extension methods to improve the evolution performance. In this study, a method for searching for the appropriate number of trees using three islands is proposed. The proposed method divides the population into three islands, but because the genetic operations and the fitness function of each island are not changed, it can be combined with the existing extension methods. In the experiments, they are compared these using two benchmark problems.