Investigation of the exponential population scheme for genetic algorithms
Journal
GECCO 2018 - Proceedings of the 2018 Genetic and Evolutionary Computation Conference
Pages
975-982
Date Issued
2018
Author(s)
Lin, Y.-J.
Abstract
Early development of GAs requires many parameters to be tuned. The tuning process increases the difficulty for inexperienced practitioners. Modern GAs have most of these parameters pre-determined, and therefore recent research concerning parameterless schemes has focused on population size. The techniques developed in this paper are mainly based on Harik and Lobo's work and the exponential population scheme (EPS), which double the population until the solution is satisfactory. In this paper, we modify EPS based on theoretical analyses. Specifically, we propose a new termination criterion and an optimized population multiplier. The experiment results show that our scheme reduces 33.4%, 19.1% and 29.6% number of function evaluations (NFE) on hBOA (the parameter-less hBOA), LT-GOMEA and DSMGA-II respectively when compared to Harik-Lobo scheme, and reduces 28.5%, 4.7% and 11.0% NFE on hBOA, LT-GOMEA and DSMGA-II respectively when compared to EPS. In addition, compared to EPS, our scheme empirically reduces the number of failures when using LT-GOMEA to solve the folded trap and MAX-SAT problems.
Type
conference paper