Neural Network with a Self-Selection Ability of Input Variables for a Nonlinear System Identification

Abstract

The purpose of this study is to propose a neural network algorithm that has an ability of self-selection of useful input variables. For a nonlinear system whose structure is very complex and which contains very many input variables, we can not generally obtain a priori knowledge about useful input variables of a nonlinear system. The neural network algorithm that is proposed in this paper has an ability of self-selection of useful input variables. The neural network is constructed with some small sub-neural networks. If the sub-neural networks contain useless input variables, they are eliminated automatically from a neural network structure according to the prediction error criterion. So the neural network is constructed with only useful input variables and a good generation can be obtained. The neural network is applied to a nonlinear system identification problem and the results are compared with those which are obtained by using another neural network and GMDH (Group Method of Data Handling) algorithm. Finally, the neural network is applied to the short-term prediction problem of air pollution concentration and compared with the linear and nonlinear prediction models.