Structural design of multilayer neural networks on the basis of entropy approach

Abstract

The formalized algorithm of designing the structure of neural network models of complex objects and systems on the basis of theoretical-information approach to assessment of the complexity of neural network (NN) structure and training set used for adjustment (learning) of NN synaptic connection weights, is offered. The various variants of statement of NN models design (image recognition, prediction of temporal series, identification of dynamic objects) are considered. The example of construction of NN classificator of minimal complexity on the basis of multilayer perceptron with one and two hidden layers, illustrating the effectiveness of proposed algorithm of structural design, is presented.