Improving Combination Methods of Neural Classifiers Using NCL

In this paper the effect of diversity caused by Negative Correlation Learning (NCL) in the combination of neural classifier is investigated and an efficient way to improve combining performance is presented. Decision Templates and Averaging, as two non-trainable combining methods and Stacked Generalization as a trainable combiner are selected as base ensemble learner and NCL version of them are compared with them in our experiments. Utilizing NCL for diversifying the base classifiers leads to significantly better results in all employed combining methods. Experimental results on five datasets from UCI Machine Learning Repository indicate that by employing NCL, the performance of the ensemble structure can be more favorable compared to that of an ensemble use independent base classifiers.