Self-Organized Subpopulation Based on Multiple Features in Genetic Programming on GPU

Abstract

Genetic Programming (GP) is an Evolutionary Computation (EC) algorithm. Controlling genetic diversity in GP is a fundamental requirement to obtain various types of local minima effectively; however, this control is difficult compared to other EC algorithms because of difficulties in measuring the similarity between solutions. In general, common subtrees and the edit distance between solutions is used to evaluate the similarity between solutions. However, there are no clear guidelines regarding the best features to evaluate it. We hypothesized that the combination of multiple features helps to express the specific genetic similarity of each solution. In this study, we propose a self-organized subpopulation model based on similarity in terms of multiple features. To reconstruct subpopulations, we introduce a novel weighted network based on each normalized feature and utilize network clustering techniques. Although we can regard similarity as a correlation network between solutions, the use of multiple features incurs high computational costs, however, calculating the similarity is very suitable for parallelization on GPUs. Therefore, in the proposed method, we use CUDA to reconstruct subpopulations. Using three benchmark problems widely adopted in studies in the literature, we demonstrate that performance improvement can be achieved by reconstructing subpopulations based on a correlation network of solutions, and that the proposed method significantly outperforms typical methods.