Selecting the optimal neural network architecture for predicting cotton yarn hairiness

Abstract

Models for predicting ring yarn hairiness are built using backpropagation neural network algorithm and linear regression analysis. An original approach dealing with overfitting and allowing the selection of the optimal neural network architecture is used. It is based on the estimation of the leverages, i.e. the influence of the training examples, on the parameters of the model and also on the calculation of the confidence intervals of the model predictions. The generalization error of the selected neural model is estimated. It reveals a very good performance in prediction, better than that of the linear model. The selected neural model is expected to be used as a ‘predictor model’ in a large scale industrial application.