Comparative assessment of Interval and Affine Arithmetic in neural network state prediction

Abstract

Two set theory methods, Interval and Affine Arithmetic, are used together with feedforward neural networks (FNN) in order to study their ability to perform state prediction in non-linear systems. Some fundamental theory showing the basic interval and affine arithmetic operations necessary to forward propagate through a FNN is presented and an application to a generic biotechnological process is performed confirming that due to the way the perturbations of the input data are considered, affine FNN perform better than interval ones. © Springer-Verlag Berlin Heidelberg 2005.