Performance of multi layer feedforward and radial base function neural networks in classification and modelling

Neural networks have been used in multiple applications, but as a kind of black box for dealing with problems where there is no a priori information about the data. This means that the model is constructed based solely upon information obtained from the data themselves. This seems to be a good property but makes it difficult to validate the models obtained. The classification properties of neural classifiers are usually described by the percentage of correctly classified objects in a test set. Since these straight methods are only based on discrimination, no information can be obtained in a statistical way. In this paper, on a simulated data set, two different types of neural networks, MLF (multi layer feedforward) and RBF (radial base function), are applied to solve a classification problem. The modelling ability, stability and reproducibility of this kind of networks are studied based on various different networks independently trained on the same data set with a predetermined value for the sensibility and specificity. Robustness to different kinds of error is also studied by means of Monte Carlo simulations adding noise at different levels and from different theoretical distributions. Further to this, an analysis based on principal components is carried out to study the apparently different networks obtained. The simulation studies reveal that both types of networks perform well enough to reproduce the input space. For RBF networks, due to the local approach, the study showed some properties related to sensibility and specificity which are relevant in practical problems.