Comparative study of model fitting by using neural network and regression

摘要

A comparative study of model fitting using neural networks and regression analysis based on the data set collected from an industrial production line is presented. This study used a five- layer back-propagation network implemented on an ANZA-plus neurocomputer by Hecht-Nielsen Corporation and a regression software package, the SAS, on a VAX/6350. The study has shown that there are some commonalities and differences in the two procedures of model fitting. Both commonalities and differences are discussed. The conclusion of our study is that a neural network is applicable to the model fitting problem. For the case where the function cannot be described parametrically, the neural network would provide a good means for performing a model fitting.