Balanced Ensemble Learning with Adaptive Bounds

Different to other re-sampling ensemble learning, negative correlation learning trains all individual models in an ensemble simultaneously and cooperatively. In negative correlation learning, each individual could see all training data, and adapt its target function based on what the rest of individuals in the ensemble have learned. In this paper, two error bounds are introduced in negative correlation learning. One is the upper bound of error output (UBEO) which divides the training data into two groups. The other is the lower bound of error rate (LBER) which is set as a switch. Before the error rate of the learned ensemble is higher than LBER, all training data is learned by negative correlation learning. As soon as the learned ensemble has a lower error rate than LBER, negative correlation learning will be applied to one group only specified by UBEO in which data points are near to the current decision boundary. This paper will examine the differences among the individual models in the ensemble by negative correlation learning with two bounds to determine how LBER and UBEO should be adapted in negative correlation learning.