Combining Boosting with Negative Correlation Learning for Training Neural Network Ensembles

This paper presents a learning algorithm, boosting with negative correlation (NC) ensemble (BNCE), for training individual neural network (NN) in an ensemble. It combines boosting with NC learning. Unlike most previous studies on training ensembles, BNCE attempts to train individual NNs by using "resample" technique of boosting. Incremental training based on NC is used in BNCE to train individual NNs for different number of training epochs. The use of NC learning and different training sets for individual NNs generated by boosting reflects emphasis on diversity among individual NNs in an ensemble. BNCE has been tested extensively on several benchmark problems in machine learning and NNs. These are the breast cancer, the credit card, the glass, the letter recognition and the soybean problems. The experimental results show that BNCE can train NN ensembles that have good generalization ability in comparison with other ensemble algorithms.