On the Effects of Outliers on Evolutionary Optimization

Most studies concerned with the effects of noise on evolutionary computation have assumed a Gaussian noise model. However, practical optimization strategies frequently face situations where the noise is not Gaussian, and sometimes it does not even have a finite variance. In particular, outliers may be present. In this paper, Cauchy distributed noise is used for modeling such situations. A performance law that describes how the progress of an evolution strategy using intermediate recombination scales in the presence of such noise is derived. Implications of that law are studied numerically, and comparisons with the case of Gaussian noise are drawn.