UDC 004.8: 004.94
METHOD OF PRELIMINARY SELECTION
OF MULTILAYER NEURAL NETWORK ARCHITECTURE
FOR POLYHARMONIC SIGNAL APPROXIMATION
V. V. Frolov, Dr. Sc. (Tech.), associate professor, professor, V. N. Karazin Kharkiv National University,Kharkov, Ukraine;
orcid.org/0000-0002-2770-3385, e-mail: This email address is being protected from spambots. You need JavaScript enabled to view it.
G. N. Zholtkevych, Dr. Sc. (Tech.), full professor, V. N. Karazin Kharkiv National University, Kharkov,Ukraine;
orcid.org/0000-0002-7515-2143, e-mail: This email address is being protected from spambots. You need JavaScript enabled to view it.
O. Yu. Prikhodko, Ph.D. (Tech.), associate professor, BSTU named after V. G. Shukhov, Belgorod, Russia;
orcid.org/0000-0002-6452-0465, e-mail: This email address is being protected from spambots. You need JavaScript enabled to view it.
Yu. V. Skuryatin, Ph.D. (Tech.), associate professor, BSTU named after V. G. Shukhov, Belgorod, Russia;
orcid.org/0000-0001-5555-8691, e-mail: This email address is being protected from spambots. You need JavaScript enabled to view it.
The proposed method for preliminary selection of the architecture for narrowing multi-layer arti-ficial
neural network of direct propagation without feedback with sigmoid activation functions is based on determining
the number of layers by the shape of a polyharmonic signal. The authors have pro-posed to characterize
signal shape by the number of inflection points. The article has experimentally proved that the number
of layers in multilayer network correlates with the number of inflection points according to the criterion of
minimizing absolute error when compared with a universal approxima-tor. The total number of neurons in a
multilayer network is determined from the condition that it can-not exceed the number of neurons in a universal
approximator. The essence of the method lies in the comparative analysis of absolute error for singlelayer
and multi-layer network. Resulting configura-tion of multilayer network can be used as an initial one
for further optimization of the structure.
Key words: artificial neural network, activation function, approximation, genetic algorithm, feedforward
neural network without feedbacks, discrete optimization, median, mean square, absolute error.