Controlled model assisted evolution strategy with adaptive preselection
Loading...
Date
2006-06
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
The utility of evolutionary algorithms for direct optimization of real processes or complex simulations is often limited by the large number of required fitness evaluations. Model assisted evolutionary algorithms economize on actual fitness evaluations by partially selecting individuals on the basis of a computationally less complex fitness model. We propose a novel model management scheme to regulate the number of preselected individuals to achieve optimal evolutionary progress with a minimal number of fitness evaluations. The number of preselected individuals is adapted to the model quality expressed by its ability to correctly predict the best individuals. The method achieves a substantial reduction of fitness evaluations on a set of benchmarks not only in comparison to a standard evolution strategy but also with respect to other model assisted optimization schemes.