Parameter Selection for Evolutionary Algorithm by Using Optimal Latin Hypercube Design

Abstract

In order to improve the accuracy of the Evolutionary Algorithm (EA), and present a guideline for choosing its appropriate parameters, we have investigated the error related to a variety of parameter setting. Four EA parameters are considered: population size, selection rate, recombination method and mutation rate. We analyzed the error of sixty experiment points sampled by using the Optimal Latin-Hypercube Design (OLHD) for fourteen test functions, and then ranked the error from low to high. We picked the parameter settings corresponding to the first 20 lowest error values, and made a box plot for them. After that we adopted the interquartile range (from lower quartile to the upper quartile) of the box plot as our confidence interval. According to the confidence interval, we obtained a recommended parameter range of three recombination methods which were considered. Among the three recombination method, the Intermediate Tendency (IT) recombination method had the lowest mean and standard deviation of errors. To validate the accuracy of the recommended parameter range, we tested the error of sixty experiment points using parameters from our recommended parameter range. Through a comparison of mean and standard deviation of errors, we found that the recommended parameter range offers better performance (with respect to error) than that of the initial parameter range.