Title: An empirical evaluation of constructive neural network algorithms in classification tasks

Authors: Maria do Carmo Nicoletti, Joao R. Bertini Jr.

Addresses: Department of Computer Science – UFSCar – SP, Sao Carlos – 13560-905, SP, Brazil. ' Department of Computer Science – UFSCar – SP, Sao Carlos – 13560-905, SP, Brazil

Abstract: Unlike conventional Neural Network (NN) algorithms that require the definition of the NN architecture before learning starts, Constructive Neural Network (CoNN) algorithms enable the network architecture to be constructed along with the learning process. This paper presents and discusses the results of an empirical evaluation of seven two-class CoNN algorithms, namely Tower, Pyramid, Tiling, Upstart, Shift, Perceptron Cascade (PC) and Partial Target Inversion (PTI) in 12 knowledge domains. The way each particular algorithm approaches the growing of the network determines their differences. This paper also presents and analyses empirical results of five multiclass CoNN algorithms in five knowledge domains, namely MTower, MPyramid, MTiling, MUpstart and MPerceptron Cascade, which can be considered extensions of their two-class counterparts. Results obtained with the Pocket with the Ratchet Modification (PRM) algorithm, with its multiclass version, the PRMWTA algorithm and with the back propagation algorithm, are presented for comparison.

Keywords: ANNs; constructive neural networks; CoNN algorithms; network architectures; learning; classification; training.

DOI: 10.1504/IJICA.2007.013397

International Journal of Innovative Computing and Applications, 2007 Vol.1 No.1, pp.2 - 13

Published online: 25 Apr 2007 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article