Abstract—A common solution to improving the
generalization problem and increasing the efficiency of
different ANNs is to use ANN ensembles. These methods focus
on the possibility of generating different neural nets for a
dataset and combining the results for acquiring a more accurate
regression. In this paper, a new ensemble method called
machine learner fusion-regression (MLF-R) is proposed to
increase the accuracy of the results through focusing on difficult
samples. The architecture of MLF-R includes two different
parts: the first is a training phase from which final nets are
selected after a filtering process; the second part is a weighted
decision maker including a backpropagation structure which
fuses the different nets derived from the first step to predict the
outputs. The results demonstrate MLF-R is more efficient than
bagging, different boosting methods and the implementation of
single ANN methods with 18% to 51% higher accuracy.
Moreover, MLF-R offers more stable results compared to the
other methods which have been tested in this paper.
Index Terms—Adaptive threshold, ensemble, fusion, neural networks.
Ali Shamsoddini is with School of Surveying and Spatial Information Systems, University of New South Wales, Sydney, Australia (e-mail: a.shamsoddini@ student.unsw.edu.au).
John C. Trinder is Visiting Emeritus Professor at School of Surveying and Spatial Information Systems, University of New South Wales, Sydney, Australia (e-mail: email@example.com).
Cite:Ali Shamsoddini and John C. Trinder, "Neural Networks Fusion for Regression Problems," International Journal of Machine Learning and Computing vol.2, no. 4, pp. 511-516, 2012.