Process modeling with neural networks using small experimental datasets

Abstract

This paper reports some work done to improve the modeling of complex processes when only small experimental data sets are available. Various solution strategies based on feed-forward and radial basis function (RBF) neural networks have been tested for three problems including two wood pulp applications. Experimental data sets obtained from D-optimal design and from a random selection throughout the experimental space were compared for their ability to lead to the proper model. In addition, the influence of activation functions, the number of levels in stacked neural networks and the composition of the training data sets have been studied. The study shows that designed training data sets are more desirable than random experimental sets, due to their higher orthogonality. The use of neural network is a powerful tool for modeling complex processes even when only a small set of data is available for training. However, special care must be exercised to insure that good predictive models are obtained.