Developing robust non-linear models through bootstrap aggregated neural networks

This paper presents a technique for building robust non-linear models by aggregating multiple neural networks. Data for building non-linear models are re-sampled using bootstrap techniques to form several different pairs of training and testing data sets. For each pair of training and testing data sets, a neural network model is developed. The developed neural network models are then combined together through principal component regression. Model generalisation capability can be significantly improved by using multiple neural networks. Confidence bounds for the neural network model predictions can also be obtained using bootstrap techniques. The technique has been successfully applied to several non-linear modelling problems including the building of software sensors for a batch polymerisation reactor. It is shown that models built from bootstrap aggregated neural networks are more accurate and robust than those built from single neural networks. (C) 1999 Elsevier Science B.V. All rights reserved.