Transition learning for creating diverse neural networks

Abstract

Ensemble approaches have been widely applied to many real world problems as they have been growing into more complex. It is essential to create a set of different subsystems which subdivide the task. Negative correlation learning (NCL) and balanced ensemble learning (BEL) have been proposed to train a number of neural networks simultaneously and cooperatively in an ensemble. It has been found that the individual neural networks created by NCL are less different than those by BEL although NCL often displayed better performance than BEL on noisy data sets. This paper examines two types of transition learning based on NCL and BEL to observe how diversity among the individual neural networks will affect the performance of the ensemble.