Created by W.Langdon from gp-bibliography.bib Revision:1.7852
As science depends on accurate hypothesis testing to make progress, we perform a comparison and evaluation of statistical techniques used to evaluate evolutionary computation systems. Finding many of these wanting, with the exception of computational effort, we introduce two additional techniques, effective mean best fitness and the y-test. We also perform an extensive analysis of the computational effort, and identify some statistical cautions around the use of this key statistic. We provide an algorithm that carefully uses computational effort to determine the best values of population size and generation number for an EC treatment.
Finally, we identify several components that are of use with the scientist algorithm. We treat the use of multiobjective algorithms in GP, principal components analysis, and their combination. We demonstrate this by providing and testing an algorithm that makes evolved trees parsimonious. We introduce the notion of incremental evolution, and use it to make useful subroutines automatically from successful solutions to easy problems. We then use this to demonstrate scalable genetic programming on an integer sorting problem.",
NR23290
supervisor Franz Oppacher",
Genetic Programming entries for Steffen Christensen