Evolution Strategies in Noisy Environments

Dirk V. Arnold

University of Dortmund, Computer Science XI

2001. 5. 26

Noise is a factor present in almost all real-world optimization problems. While it can potentially improve convergence reliability in multimodal settings by preventing convergence towards merely local optima, it is generally detrimental to the velocity with which an optimum is approached. Evolution strategies (ES) form a class of evolutionary optimization procedures that are believed to be able to cope quite well with noise. A number of interesting results have been obtained by analytically studying the performance of ES in very simple fitness environments. By restricting oneself to such simple environments, it becomes possible to single out the influence on the ability to deal with noise of different components of the strategies. In this talk, we review some of these recent results. In particular, we identify fitness overvaluation as a significant influence on ES performance, discuss the benefits of working with populations of individuals rather than with single candidate solutions, and study the effects of genetic repair in the presence of noise.

This page is maintained by Hee-Ju Chang (hjchang@scai.snu.ac.kr).
Last update: May 8, 2001