GMDH-type neural network algorithm with a feedback loop for structural identification of RBF neural network

Abstract

In this paper, a Group Method of Data Handling (GMDH)-type neural network algorithm with a feedback loop for structural identification of Radial Basis Function (RBF) neural network is proposed. In case of the GMDH-type neural network, the network architecture is automatically organized by heuristic self-organization. Optimum architecture is evolved using one of the criterions, defined as Akaike's Information Criterion (AIC) or Prediction Sum of Squares (PSS), for minimizing the prediction error. In the conventional multilayered neural network, prediction error criteria defined as AIC and PSS cannot be used to determine the neural network architecture. In case of the GMDH-type neural network proposed in this paper, structural parameters such as the number of neurons, relevant input variables and the number of feedback loop calculations are automatically determined so as to minimize AIC or PSS. Furthermore, the GMDH-type neural network can identify RBF neural network accurately, since the complexity of the neural network is increased gradually by feedback loop calculations.