Метод еволюції архітектури штучних нейронних мереж

Abstract

This paper considers a method for evolving the architecture of artificial neural networks based on the NEAT (NeuroEvolution of Augmenting Topologies) algorithm. The approach involves simultaneous optimization of connection weights and network topology through evolutionary operators of selection, recombination, and mutation. It is shown that the use of neuroevolution enables automated synthesis of adaptive artificial neural networks structures without manual design. The advantages of the method are identified, including preservation of population diversity and gradual complexification of the architecture, as well as inherent limitations related to computational cost and convergence speed.