Diploidy and Dominance in Artificial Genetic Search

Genet ic algor ithms (GAs) continue to receive increased attention as general-p ur pos e met hods for search and optimization . GAs are attrac t ive for search in complex spaces where the fun ction al relati onsh ips between paramet ers and ob jective function values are of unknown , arb itrary mat hemati cal charac te r. Desp ite t heir generally robust character, typical GAs are known to fare poorly on functions that var y with time, where the goal is to track nonstation ary opt ima. It has been theor ized that in natural gene t ics , diploidy and dominance increase the sur vivability of species in environments that vary wit h time. T his paper examines the effects of diploid representations and dominan ce op erators in genetic algorithms applied to nonstationary search problems. Ex pe rimental resu lt s indi ca te that t hese ad dit ions greatly increase t he efficacy of GAs in t ime-varying env ironments. T his increased performan ce is mad e possib le by abeyant recessive alleles. Analytical arguments show that these recessive alleles increase popul a tion diversity wit hout t he disruptive effects of high mu ta tion rates, t hus allowing t he GA to renew it s search process as the problem varies with time. An alysis also reveals that abeyant recessives are sens it ive to past env ironme nt al condit ions , and can therefore act as a form of distributed, probabilist ic memory of env ironmental cond itions t hat occur periodically. Final sections discuss exte nsions and implications of this work, including mult i-locus do minance und er genic cont rol, intrachromosom al dominance, and how diploidy may affect the inherently noisy GA search process, if t his noise is viewed as a nons tationary aspect of the objective fun ction . 'Electronic mail address: iDualix .ua , edu :robiDgalab2 . mh . ua. edu tE lectronic mail address: goldbergiDvmd . es o .ui ue .edu 252 Robert E. Smith and David E. Goldberg