Dispersion-based population initialization

Abstract

Reliable execution and analysis of an evolutionary algorithm (EA) normally requires many runs to provide reasonable assurance that stochastic effects have been properly considered. One of the first stochastic influences on the behavior of an EA is the population initialization. This has been recognized as a potentially serious problem to the performance of EAs but little progress has been made in improving the situation. Using a better population initialization algorithm would not be expected to improve the many-run average performance of an EA, but instead, it would be expected to reduce the variance of the results, without loss of average performance. This would provide researchers the opportunity to reliably examine their experimental results while requiring fewer EA runs for an appropriate statistical sample. This paper uses recent advances in the measurement and control of a population's dispersion in a search space to present a novel algorithm for better population initialization. Experimental verification of the usefulness of the new technique is provided. © Springer-Verlag Berlin Heidelberg 2003.