Evolutionary Algorithms in Noisy Environments: Theoretical Issues and Guidelines for Practice

This paper is devoted to the eeects of tness noise in EAs (evolutionary algorithms). After a short introduction to the history of this research eld, the performance of GAs (genetic algorithms) and ESs (evolution strategies) on the hyper-sphere test function is evaluated. It will be shown that the main eeects of noise { the decrease of convergence velocity and the residual location error R 1 { are observed in both GAs and ESs. Diierent methods for improving the performance are presented and hypotheses on their working mechanisms are discussed. The method of rescaled mutations is analyzed in depth for the (1;)-ES on the sphere model. It is shown that this method needs advanced self-adaptation techniques in order to take advantage of the theoretically predicted performance gain. The troubles with current self-adaptation techniques are discussed and directions for further research will be worked out.