Modern approaches to tuning neural network hyperparameters

Abstract

This study presents methods for using genetic algorithms in optimizing the hyper parameters of artificial neural networks. This study provides an analysis of the main types of GA, their mathematical models, and areas of practical application. At the same time, in the process of selecting hyper parameters, the effectiveness of Random Search, Grid Search, and genetic algorithms was compared in the MNIST classification task. The results of the experiment show that GAs provide an optimal balance in terms of calculation time and accuracy and show a higher accuracy compared to the random search method and a shorter calculation time compared to the grid search method. In conclusion, genetic algorithms are considered an effective approach for optimizing the hyper parameters of neural networks, particularly in cases where the parameter space is large and complex