Метод оптимізації парметрів штучних нейронних мереж

Abstract

The paper considers an evolutionary approach to training artificial neural networks based on a genetic algorithm for optimizing weight coefficients. The main stages of the algorithm implementation are analyzed, including population initialization, fitness evaluation, and the application of selection, recombination, and mutation operators. The advantages of the evolutionary method are substantiated, in particular the global search capability, independence from gradient computation, and effectiveness in optimizing complex nonlinear and non-differentiable objective functions. The expediency of using genetic algorithms to improve the efficiency of neural network training is demonstrated.