Evolutionary Optimization of the Weight Coefficients of a Multilayer Perceptron for Solving a Binary Classification Problem

Abstract

Currently, the main method used for training neural networks is the backpropagation method, which is based on selecting appropriate weight coefficients using gradient descent. However, this method has a major limitation: it may become trapped in local optima. The application of a genetic algorithm helps to overcome this problem, as it processes multiple alternative solutions in parallel and focuses on the most promising ones. This paper considers the application of a genetic algorithm to the problem of predicting customer purchase decisions. Improving the accuracy of such predictions can significantly increase the efficiency of pricing, assortment management, and customer needs analysis, as well as support the formation of decision-making strategies. The experimental results show that the use of genetic algorithms improves the quality of classification metrics. In the considered task, the genetic algorithm achieved an accuracy of 0.92, whereas the backpropagation method achieved an accuracy of 0.90