Study and Comparison of Initial Populations on the Performance of Modified Differential Evolution Algorithm

Abstract

In nature, based on the survival of the fittest principle, the individual with better fitness has more chance to survive and produce a competitive offspring with inherited traits of the parent. Similarly, in evolutionary algorithms, individuals or solutions with better fitness can generate better offspring with higher survival rate. In evolutionary algorithms, the offspring are usually produced by variation operations (mutation and crossover). The generated offspring are compared with parents and better one either from the offspring or parent is selected in the next generation. The quality of initial population (group of individuals or solutions) can affect the performance of evolutionary algorithm. The present work aims to study the effect of initial population on the performance of modified Differential Evolution (DE) algorithm. Four types of initial population, generated by different techniques, are studied and compared for some standard benchmark functions. Four population-generation techniques, namely, random population, random population using the concept of design of experiments, population using symmetric Latin hypercube design and opposition-based learning population, are considered. The results are compared based on the performance measure such as success rate, feasibility rate, success performance, average number of evaluations to a solution, quality measure, enes and population measure.