Neural Network with a Self-Selection Ability of Network Structure and Its Application to Nonlinear System Identification

Abstract

A neural network which has a self-selection ability of network structure is described. This neural network algorithm is a revised one of the neural network which can identify a nonlinear system whose structure is very large and complex. By using a criterion based on a prediction error, a neural network structure which is adapted for a complexity of a nonlinear system can be selected automatically. This neural network is applied to an identification problem of a rolling model and a physical and neural network combined model is identified. It is shown that the physical and neural network combined model gives better prediction results as compared with a conventional physical and statistical combined model.