Estimations of error bounds for RBF networks

The training and optimisation of neural networks to perform function approximation tasks is well documented in the literature [8, 3, 14]. The usefulness of neural networks will be enhanced if a further capacity is added to them: the ability to estimate the accuracy of the results which they generate. Not only will this provide users of neural networks with a confidence index, it will also enable the estimates from the neural networks to be included as part of an overall estimation scheme in which several estimates are combined in a Bayesian manner to guarantee the optimality (in terms of minimum variance) of the result. For example, it would enable the results from a neural network estimator to be included in a Kalman filter cycle with full mathematical rigour [14]. In this paper the suitability of a perturbation model to perform such a task will be examined.