We investigate co-evolutionary GP that that co-evolvesfitness predictors in order to reduce the computationalcost of evolution and/or reduced the number ofevaluations required. Fitness predictors are lightobjects which, given an evolving individual,heuristically approximate its true fitness. Thepredictors are trained by their ability to correctlydifferentiate between good and bad solutions usingreduced computation. We apply coevolution of fitnesspredictors to symbolic regression and measure itsimpact. Our results show that a small computationalinvestment in co-evolving fitness predictors greatlyenhances both speed and convergence of individualsolutions while reducing the computational effortoverall. Finally we apply fitness prediction tointeractive evolution of pen stroke drawings. Theseresults show that fitness prediction is extremelyeffective at modelling user preference while minimisingthe sampling on the user to fewer than ten prompts.
CITATION STYLE
Schmidt, M. D., & Lipson, H. (2007). Coevolving Fitness Models for Accelerating Evolution and Reducing Evaluations. In Genetic Programming Theory and Practice IV (pp. 113–130). Springer US. https://doi.org/10.1007/978-0-387-49650-4_8
Mendeley helps you to discover research relevant for your work.