An evidence-theoretic neural network classifier
T. Denoeux · 2002
A new classifier based on the Dempster-Shafer theory of evidence is presented. The approach consists in considering the similarity to prototype vectors as evidence supporting certain hypotheses concerning the class membership of a pattern to be classified. The different items of evidence are represented by basic belief assignments over the set of classes and combined by Dempster's rule of combination. An implementation of this procedure in a neural network with specific architecture and learning procedure is presented. A comparison with LVQ and RBF neural network classifiers is performed.