Neural Model Identification Using Local Robustness Analysis

A local robustness approach for the selection of the architecture in multilayered feedforward artificial neural networks (MLFANN) is studied in terms of probability density function (PDF) in this work. The method is used in a non-linear autoregressive (NAR) model with innovative outliers. The procedure is proposed for the selection of the locally most robust (around a particular sample) MLFANN architecture candidate for exact learning of a finite set of the real sample. The proposed selection method is based on the output PDF of the MLFANN. As each MLFANN architecture leads to a specific output PDF when its input is a distribution with heavy tails, a distance between probability densities is used as a measure of local robustness. A Monte Carlo study is presented to illustrate the selection method.