Evolution strategies applied to perturbed objective functions

We investigate the behavior of evolution strategies on noisy objective functions. We show for the simple sphere model that convergence velocity is not reduced as long as the noise level is small compared to the function value. If the noise level reaches a certain threshold, a size of the parent population greater than 1 improves the convergence precision significantly. Convergence reliability is tested for two nonconvex functions. Again the search process seems to be not influenced by low level noise. Interpreting the impact of noise purely as a modification of the selection process gives new insight into the role of selection in evolution strategies.<<ETX>>