Genetic algorithm for training dynamical object emulator based on RBF neural network

Article type: Research Article

Authors: Sergeev, S.A.^a | Mahotilo, K.V.^a | Voronovsky, G.K.^b | Petrashev, S.N.^b

Affiliations: [a] Kharkov State Polytechnic University, 21 Frunze St., Kharkov 310002, Ukraine E-mail: [email protected] | [b] Kharkov Combined Cycle Plant-5, Podvorky, Kharkov 312024, Ukraine E-mail: [email protected]

Abstract: The combination of Genetic Algorithms (GAs) and Artificial Neural Networks (ANNs) has already resulted in researchers advancing in quite a few real world applications but it is in control that this alliance yields such appreciable benefit. The paper reports a Radial Basis Function (RBF) network training technique which joins together global strategy of GAs and a local adjusting procedure typical for RBF networks. While activation function window centres and widths are processed via a “slow” numeric GA, output‐layer neurone synaptic weights are defined by a “fast” analytical method. The technique allows to minimize not only the network hidden-layer size but also the pattern set required for training the adequate dynamical object neuroemulator.

DOI: 10.3233/JAEM-1998-096

Journal: International Journal of Applied Electromagnetics and Mechanics, vol. 9, no. 1, pp. 65-74, 1998