Applicability of genetic algorithms for determining the weighting coefficients of an artificial neural network with one hidden layer

In the training of an artificial neural network, one of the central problems is the initial initialization and adjustment of weighting coefficients associated with pseudo-random initialization of weighting coefficients. The article describes a basic genetic algorithm, as well as a method for determining weight coefficients using this algorithm. A combined method for determining weighting coefficients is also presented, which provides for initial initialization using a genetic algorithm at the first stage and the use of stochastic gradient descent at the second stage of training, the proposed methods are tested on a number of artificial neural networks of direct propagation for various tasks of binary classification of real and synthetic data, as well as for unambiguous multiclass classification of handwritten digits on images from the database MNIST data. Artificial neural networks are constructed on the basis of the Kolmogorov-Arnold theorem. This article presents a comparative analysis of two methods for determining weight coefficients – using a genetic algorithm and gradient descent. Based on the results of the comparative analysis, it is concluded that a genetic algorithm can be used to determine the weighting coefficients both as an algorithm for the initial initialization of an artificial neural network and as an algorithm for adjusting the weighting coefficients.