Learning Targets for Building Cooperation Awareness in Ensemble Learning

Abstract

For an ensemble learning system by a set of individual learning models, not only should each individual model be able to learn well the data, but also be aware of what other individuals have learned. With such awareness, each individual would be able to adjust its own learning so that all the individuals could cooperatively and efficiently solve the whole data better. In this paper, all the individual learners are trained simultaneously by a modified negative correlation learning with opposition learning. The idea is that some individuals might choose to learn to be more different to other individuals on some data points once the whole ensemble have well learned these data points. Experimental results have been presented to show how such opposition learning could build awareness among individual learners so that they could be helpful in designing a robust ensemble learning system.