Novel approaches to confidence bound generation for neural network representations

With the continuing strategic interest in the application of neural networks for inferential estimation and control, it is essential that confidence bounds are placed around the resulting predictions. In this paper, two alternative approaches for the development of confidence bounds for neural network models are described. The first approach focuses upon the conjunction of two important aspects that determine model accuracy: the ability of the neural network to predict an output and the influence of the availability of the training data. In contrast, in the second approach, the confidence bounds result from the development of a robust neural network representations through the conjunction of multiple neural networks. Utilising the bootstrap technique, a number of training and test data sets are created from which neural network representations are developed and the models combined using principal component regression. Confidence bands for the model predictions are automatically produced as a result of the technique. The two approaches are compared by application to a batch polymerisation reactor.