Committee c-mantec: A probabilistic constructive neural network

Abstract

C-Mantec is a recently introduced constructive algorithm that generates compact neural architectures with good generalization abilities. Nevertheless, it produces a discrete output value and this might be a drawback in certain situations. We propose in this work two approaches in order to obtain a continuous output network such as the output can be interpreted as the probability of a given pattern to belong to one of the output classes. The CC-Mantec approach utilizes a committee strategy and the results obtained both with the XOR Boolean function and with a set of benchmark functions shows the suitability of the approach, as an improvement over the standard C-Mantec algorithm is obtained in almost all cases. © 2013 Springer-Verlag Berlin Heidelberg.