Statistical, identification techniques for network topology selection

The problem of selecting the topology of a multilayer perceptron in order to obtain the best generalisation performance can be approached by means of statistical identification techniques. In the present paper three techniques for selecting the model order are described and their potential limitations when applied to the non linear feedforward neural models are evidenced On the basis of this study, an original implementation of two of these methods to multilayer perceptrons is proposed. A comparative analysis of the effectiveness of the resulting procedures is tested on both synthetic and real world data experiments.