Benchmarking the Performance of a Machine Learning Classifier Enabled Multiobjective Genetic Algorithm on Six Standard Test Functions

Abstract

Previous work tested a multi-objective genetic algorithm that was integrated with a machine learning classifier to reduce the number of objective function calls. Four machine learning classifiers and a baseline “No Classifier” option were evaluated. Using a machine learning classifier to create a hybrid multiobjective genetic algorithm reduced objective function calls by 75–85% depending on the classifier used. This work expands the analysis of algorithm performance by considering six standard benchmark problems from the literature. The problems are designed to test the ability of the algorithm to identify the Pareto frontier and maintain population diversity. Results indicate a tradeoff between the objectives of Pareto frontier identification and solution diversity. The “No Classifier” baseline multiobjective genetic algorithm produces the frontier with the closest proximity to the true frontier while a classifier option provides the greatest diversity when the number of generations is fixed. However, there is a significant reduction in computational expense as the number of objective function calls required is significantly reduced, highlighting the advantage of this hybrid approach.