Machine Learning Models for Gain Regression Trained on 10 Million Yagi–Uda Antennas

Abstract

Statistical comparisons of two machine learning models, multilayer perceptrons and extreme gradient boosting, for predicting the forward gain of Yagi-Uda antennas from their geometry are presented. The primary concerns in the presented comparisons are prediction accuracy, training time, and model complexity. Two datasets, comprising 10 million antenna samples each and different ranges of antenna design parameters, are used in the experiments. An electromagnetic-based relationship between model prediction error and antenna gain error is established to facilitate setting training stopping criteria for specified gain prediction accuracy. The reported average training times and prediction accuracies are obtained from sets of independently trained machine learning models. Results are accompanied by 90 % confidence intervals estimated using Student’s $t$ -distribution. In general, extreme gradient boosting models require shorter training times, whereas models based on multilayer perceptrons achieve higher prediction accuracy. Furthermore, a classification-regression approach is proposed to exclude low-gain antenna samples before regression, thereby reducing training time while maintaining or improving prediction accuracy. The applicability of the proposed approach is demonstrated through antenna optimization and validated by full-wave electromagnetic simulations. The optimization time is reduced by two to three orders of magnitude compared with optimizations based on numerical electromagnetic solvers.