Extrapolation-directed crossover for real-coded GA: overcoming deceptive phenomena by extrapolative search

Abstract

Proposes a new real-coded genetic algorithm (GA) using the combination of two crossovers: UNDX-m (unimodal normal distribution crossover - modified) and EDX (extrapolation-directed crossover). The search region of UNDX-m tends to be biased toward the inside of the area that the population of the GA covers. Because of this search bias, the GA using UNDX-m causes stagnation of its search if the cost surface has a certain kind of structure - viz. the so-called ridge structure or multiple-peak structure. In order to compensate for this fault of UNDX-m, we propose a new crossover - EDX - which has an extrapolative search area, and we show its effectiveness through numerical experiments.