New multi-objective algorithms for neural network training applied to genomic classification data

Marcelo Costa*, Thiago Rodrigues, Euler Horta, Antônio Braga, Carmen Pataro, René Natowicz, Roberto Incitti, Roman Rouzier, Arben Çela

*Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceedingChapterpeer-review

1 Scopus citations

Abstract

One of the most used Artificial Neural Networks models is the Multi-Layer Perceptron, which is capable to fit any function as long as they have enough number of neurons and network layers. The process of obtaining a properly trained Artificial Neural Network usually requires a great effort in determining the parameters that will make it to learn. Currently there are a variety of algorithms for Artificial Neural Networks's training working, simply, in order to minimize the sum of mean square error. However, even if the network reaches the global minimum error, it does not imply that the model response is optimal. Basically, a network with large number of weights but with small amplitudes behaves as an underfitted model that gradually overfits data during training. Solutions that have been overfitting are unnecessary complexity solutions.Moreover, solutions with low norm of the weights are those that present underfitting, with low complexity. The Multi-Objective Algorithmcontrols the weights amplitude by optimizing two objective functions: the error function and norm function. The high generalization capability of the Multi-Objective Algorithm and an automatic weight selection is aggregated by the LASSO approach, which generates networks with reduced number of weights when compared with Multi-Objective Algorithm solutions. Four data sets were chosen in order to compare and evaluate MOBJ, LASSO and Early-Stopping solutions. One generated from a function and tree available from a Machine Learning Repository. Additionally, the MOBJ and LASSO algorithms are applied to a microarray data set, which samples correspond to a genetic expression profile from DNA microarray technology of neoadjuvant chemotherapy (treatment given prior to surgery) for patients with breast cancer. Originally, the dataset is composed of 133 samples with 22283 attributes. By applying e probe section method described in the literature, 30 attributes were selected and used to train the Artificial Neural Networks. In average, the MOBJ and LASSO solutions were the same, the main difference is the simplified topology achieve by LASSO training method.

Original languageEnglish (US)
Title of host publicationFoundations of Computational, Intelligence Volume 1
Subtitle of host publicationLearning and Approximation
PublisherSpringer Verlag
Pages63-82
Number of pages20
ISBN (Print)9783642010811
DOIs
StatePublished - Jan 1 2009

Publication series

NameStudies in Computational Intelligence
Volume201
ISSN (Print)1860-949X

ASJC Scopus subject areas

  • Artificial Intelligence

Fingerprint

Dive into the research topics of 'New multi-objective algorithms for neural network training applied to genomic classification data'. Together they form a unique fingerprint.

Cite this