Parental population sizing in evolutionary strategies

Abstract

One of the major differences of evolution strategies (ES) from other evolutionary computation (EC) is self-awareness of the distinction between parental and offspring populations. This feature is all the more prominent when it comes to the availability of two ES variants: COMMA and PLUS ES (i.e. (/spl mu/, /spl lambda/) and (/spl mu/+/spl lambda/) ES). Even though these two schemes capture, respectively, the antipodal emphases on exploration and exploitation, conventional practices of ES still suffer from the problem of premature convergence by the proposition of an adaptive parental population sizing approach, the authors wish to buttress the original ES scheme. This population sizing approach differs from previous ones in that only size of the parental population is adjusted. Simulation on several benchmark problems vindicates the virtue of this modification. Moreover, the modified algorithm also sheds light on convergence status of the population for a given optimization task, hence allowing further improvements on ES toward premature convergence checking and parallel search mechanisms.