Using statistical techniques to predict GA performance

Abstract

Tho design of models efficiently predicting the performance of a particular genetic algorithm on a given fitness landscape is a very important issue of practical interest. Virtual Genetic Algorithms (YGAs) constitute a statistical approach aimed at this objective. This work describes different improvements to the standard VGA model. These improvements are based on the use of a more representative dataset for the statistical analysis, the partitioning of this dataset into separate prediction models, and the utilization of a more sophisticated statistical model to grasp the distribution of fitnesses. The empirical evaluation of this enhanced model shows a more accurate fitness prediction. Furthermore, fast qualitative assessment of parameter changes is shown to be possible. © Springer-Verlag Berlin Heidelberg 2001.