Nonlinear Pattern Identification by Multi-layered GMDH-Type Neural Network Self-selecting Optimum Neural Network Architecture

Abstract

A revised Group Method of Data Handling (GMDH)-type neural network is applied to the nonlinear pattern identification. The GMDH-type neural network has both characteristics of the GMDH and the conventional multilayered neural network trained by the back propagation algorithm and can automatically organize the optimum neural network architecture using the heuristic self-organization method. In the GMDH-type neural network, many types of neurons described by such functions as the sigmoid function, the radial basis function, the high order polynomial and the linear function, can be used to organize neural network architecture and neuron characteristics, which fit the complexity of the nonlinear system, are automatically selected so as to minimize the error criterion defined as Akaike’s Information Criterion (AIC) or Prediction Sum of Squares (PSS). In this paper, the revised GMDH-type neural network is applied to the identification of the nonlinear pattern, showing that it is a useful method for this process.