Complementary neural networks for regression problems
Kraipeerapun, P., Nakkrasae, S., Amornsamankul, S. and Fung, C.C. (2009) Complementary neural networks for regression problems. In: Eighth International Conference on Machine Learning and Cybernetics, 12-15 July 2009, Baoding, China.
|PDF - Published Version |
Download (379kB) | Preview
*Subscription may be required
In this paper, complementary neural networks (CMTNN) are used to solve the regression problem. CMTNN consist of a pair of opposite neural networks. The first neural network is trained to predict degree of truth values and the second neural network is trained to predict degree of falsity values. Both neural networks are complementary to each other since they deal with pairs of complementary output values. In order to predict the more accurate outputs, each pair of the truth and falsity values are aggregated based on two techniques which are equal weight combination and dynamic weight combination. The first technique is just a simple averaging whereas the second technique deals with errors occurred in the prediction. We experiment our approach to the classical benchmark problems including housing, concrete compressive strength, and computer hardware from the UCI machine learning repository. It is found that complementary neural networks improve the prediction performance as compared to the traditional single backpropagation neural network and support vector regression used to predict only truth values. Furthermore, the difference between the predicted truth value and the complement of the predicted falsity value can be used as an uncertainty indicator to support the confidence in the prediction of unknown input data.
|Publication Type:||Conference Paper|
|Murdoch Affiliation:||School of Information Technology|
|Copyright:||(c) 2009 IEEE|
|Notes:||Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be obtained from the IEEE.|
|Item Control Page|