A CUDA-based Self-adaptive Subpopulation Model in Genetic Programming: cuSASGP

Created by W.Langdon from gp-bibliography.bib Revision:1.4192

  author =       "Keiko Ono and Yoshiko Hanada",
  title =        "A {CUDA}-based Self-adaptive Subpopulation Model in
                 Genetic Programming: {cuSASGP}",
  booktitle =    "Proceedings of 2015 IEEE Congress on Evolutionary
                 Computation (CEC 2015)",
  year =         "2015",
  editor =       "Yadahiko Murata",
  pages =        "1543--1550",
  address =      "Sendai, Japan",
  month =        "25-28 " # may,
  publisher =    "IEEE Press",
  keywords =     "genetic algorithms, genetic programming, GPU",
  DOI =          "doi:10.1109/CEC.2015.7257071",
  abstract =     "A parallel model encourages genetic diversity and
                 frequently shows a better search performance than do
                 single population models. In the parallel model,
                 individuals generally migrate to another subpopulation
                 based on their fitness values, where both the number of
                 individuals in each subpopulation and the topology are
                 fixed. To enhance the parallel model in the framework
                 of genetic programming (GP), it is important to
                 consider a balance between local and genetic search.
                 The incorporation of a local search method into the
                 parallel GP model is a promising approach to enhancing
                 it. In GP, individuals have various features because of
                 their structures, and therefore, it is difficult to
                 determine which feature is the most effective for local
                 search. Therefore, we propose a novel adaptive
                 subpopulation model based on various features of
                 individuals in each generation, in which subpopulations
                 are adaptively reconstructed based on a fitness value
                 and the distance between individuals. The proposed
                 method automatically generates a correlation network on
                 the basis of the difference between individuals in
                 terms of not only a fitness value but also node size
                 and generates subpopulations by network clustering. By
                 virtue of the reconstruction, individuals with similar
                 features can evolve in the same subpopulation to
                 enhance local search. Since, on the one hand, the
                 generation of a correlation network of individuals
                 requires considerable computational effort, and on the
                 other, calculating correlation among individuals is
                 very suitable for parallelization, we use CUDA to
                 construct a correlation network. Using three benchmark
                 problems widely adopted in studies in the literature,
                 we demonstrate that performance improvement can be
                 achieved through reconstructing subpopulations based on
                 a correlation network of individuals, and that the
                 proposed method significantly outperforms a typical
  notes =        "1005 hrs 15288 CEC2015",

Genetic Programming entries for Keiko Ono Yoshiko Hanada