COMMENT ON A RECENT SENSITIVITY ANALYSIS OF RADIAL BASE FUNCTION AND MULTILAYER FEEDFORWARD NEURAL-NETWORK MODELS - RESPONSE

In our paper [1], the modeling capabilities of multi-layered feed-forw ard (MLF) and radial base function (RBF) networks were investigated on simulated data and well described experimental data from chemical ind ustry [4]. Since both networks are based on a different concept (that is, RBF in contrast to MLF shows more local modeling behaviour) both m odeling capability and robustness to input errors have been examined. The 'robustness' was expressed in terms of sensitivity of the network output units to random input perturbations by means of controlled pseu do-random noise. In this response paper, the comment of Faber et al., i.e., applying theoretical error propagation on artificial neural netw orks, and the consequences for the conclusions drawn in the original p aper [1], are addressed.