Murdoch University Research Repository

Welcome to the Murdoch University Research Repository

The Murdoch University Research Repository is an open access digital collection of research
created by Murdoch University staff, researchers and postgraduate students.

Learn more

Porosity prediction using bagging of complementary neural networks

Kraipeerapun, P., Fung, C.C.ORCID: 0000-0001-5182-3558 and Nakkrasae, S. (2009) Porosity prediction using bagging of complementary neural networks. In: Sixth International Symposium on Neural Networks (ISNN 2009), 26-29 May 2009, Wuhan, China pp. 175-184.

Link to Published Version:
*Subscription may be required
Google Books Preview:
*Open access. Some pages may not be available


This paper presents a novel approach to the regression problem using bagging of complementary neural networks (CMTNN). A bagging technique is applied to an ensemble of pairs of feed-forward backpropagation neural networks created to predict degrees of truth and falsity values. In our approach, uncertainties in the prediction of the truth and falsity values are quantified based on the difference among all the predicted truth values and the difference among all the predicted falsity values in the ensemble, respectively. An aggregation technique based on uncertainty values is proposed. This study is realized to the problem of porosity prediction in well log data analysis. The results obtained from our approach are compared to results obtained from three existing bagging models. These three models are an ensemble of feed-forward backpropagation neural networks, an ensemble of general regression neural networks, and an ensemble of support vector machines. We found that our approach improves performance compared to those three existing models that apply a simple averaging technique based on only the truth porosity values in the ensemble.

Item Type: Conference Paper
Murdoch Affiliation(s): School of Information Technology
Publisher: Springer Berlin
Copyright: © Springer-Verlag Berlin Heidelberg 2008
Conference Website:
Item Control Page Item Control Page