Comparison of Machine Learning Classifiers for Reducing Fitness Evaluations of Structural Optimization

Abstract

Metaheuristic algorithms have been widely used to solve structural optimization problems. Despite their powerful search capabilities, these algorithms often require a large number of fitness evaluations. Constructing a machine learning classifier to identify which individuals should be evaluated using the original fitness evaluation is a great solution to reduce the computational cost. However, there is still a lack of a thorough comparison between machine learning classifiers when integrating into the optimization process. This paper aims to evaluate the efficiencies of different classifiers in eliminating unnecessary fitness evaluations. For this purpose, the weight optimization of a double-layer grid structure comprising 200 members is used as a numerical experiment. Six machine learning classifiers selected for assessment in this study include Artificial Neural Network, Support Vector Machine, k-Nearest Neighbor, Decision Tree, Random Forest, and Adaptive Boosting. The comparison is made in terms of the optimal weight of the structure, the rejection rate as well as the computing time. Overall, it is found that the AdaBoost classifier achieves the best performance.