Classification Methods I — Neural Networks

Abstract

As noted in Chap. 1, the reliability problem can be regarded under a classification viewpoint, because the limit-state function determines two well defined classes of samples, namely the safe and failure ones. Despite this clear association, the structural reliability community has paid little attention to solve this problem under such a paradigm. This is perhaps due to the fact that classical pattern recognition methods are dominated by the Bayesian approach to class discrimination, whose application in structural reliability is not adequate as shown later on. The development of computers, however, has fostered the development of new statistical methods that are characterized by an abandonment of the parametric approach that dominated Statistics in the most part of the twentieth century and the use of either non-parametric or adaptiveflexible models. Nowadays, there are many pattern recognition methods that could be investigated with regard to their applicability in structural reliability (See [69, 42] and [148] for a detailed description of these methods). In order to facilitate this task, it is valuable to group them in some classes, according to their treatment of the training samples.