Fast and Accurate Training of Multilayer Perceptrons Using anExtended

The training algorithm EKFNet uses an Extended Kalman Filter for supervised learning of feed-forward neural nets. The big diierence with respect to ordinary backpropagation methods is the calculation of a N N covariance matrix which considers the interdependence of the N weights that have to be optimised. Therefore computing time increases quadratically with the number of weights, thus restricting fast learning to problems with up to about 200 weights. In contrast to other optimization methods training is performed as learning-by-pattern to allow fast convergence in case of long sets of training data. EKFNet is demonstrated for the task of decoupling an industrial robot as well as for a problem of the PROBEN1 benchmark set. In most cases EKFNet proves to learn faster and more accurate than any other method tested.