Abstract

The intermediate (center of mass) recombination of object parameters is introduced in the evolution strategy with derandomized covariance matrix adaptation (CMA-ES). On various (unimodal) real space fitness functions convergence properties and robustness against distorted selection are tested for different parent numbers. Introducing (= I ; )-selection significantly improves robustness and has comparatively minor influence on the convergence speed of the CMA-ES. I. Introduction In evolutionary computation usually the aspect of global search is emphasized. In contrast, to approach a (global) optimum quickly, after its basin of attraction is found, convergence velocity of an evolutionary search algorithm is important. In our opinion the importance of this aspect is, at least in continuous parameter optimization problems, underestimated; in practical applications the optimum is not approached sufficiently--- 'premature convergence' can be observed. In evolution strategies (ESs) this pr...

Citations