Title: A two-class constructive neural network algorithm for continuous domains: the OffTiling algorithm

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

Addresses: Institute of Mathematical Sciences and Computing, University of Sao Paulo, Av. Trabalhador Sao-Carlense 400, Sao Carlos, Brazil. ' Department of Computer Science, Federal University of Sao Carlos, Via Washington Luiz, km. 238, Sao Carlos, Brazil

Abstract: It is well-known that the architecture of a neural network (NN) plays a crucial role on its generalisation capabilities. While conventional NN training algorithms assume a pre-defined NN architecture before training starts, constructive neural network (CoNN) algorithms define the architecture of the NN along with its training. This paper describes the proposal of a new two-class CoNN algorithm named OffTiling, suitable for continuous valued domains and slightly inspired by two other CoNN algorithms found in the literature, the Offset, designed for learning in Boolean domains and the two-class CoNN Tiling algorithm. A comparative analysis of the OffTiling versus four other CoNN algorithms, namely the Pyramid, Tiling, PTI and the TilingLike in ten knowledge domains is presented, showing that the OffTiling can be a competitive algorithm as far as the performance of the network is concerned.

Keywords: constructive neural networks; data mining; two-class pattern classification; hybrid networks; Tiling; Offset.

DOI: 10.1504/IJKEDM.2010.034842

International Journal of Knowledge Engineering and Data Mining, 2010 Vol.1 No.2, pp.161 - 183

Available online: 24 Aug 2010 *

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