Structural identification of the multi-layered neural networks by using GMDH-type neural network algorithm

In this study, structural identification of the multi-layered neural networks ( sigmoid function type neural networks) by using GMDH-type neural network algorithm is developed. The GMDH-type neural network algorithm can generate optimum multi-layered neural network architectures fitting the complexity of nonlinear system so as to minimize the error criterion defined as AIC (Akaike's Information Criterion). The GMDH-type neural networks are organized by using heuristic self-organization method proposed by A.G.Ivakhnenko. The GMDH-type neural networks have abilities of self-selecting the number of layers, the number of neurons in the hidden layers, the useful input variables and the optimum neuron architectures in the hidden layers. Therefore, it is very easy to find out the optimum neural network architectures fitting the complexity of nonlinear system by using the GMDH-type neural network algorithm.