Combining Two Negatively Correlated Learning Signals in a Committee Machine

Negative correlation learning is to let a set of individual learners communicatively learn the same data set without data sampling although data sampling techniques could be used in negative correlation learning. In this paper, weak learning signals were introduced by defining the new target values in negative correlation learning besides the original strong learning signals. The purposes of using weak learning signals are to create more different learning models for a committee machine, and speed up learning on the hard data. However, neither strong learning signals nor weak learning signals would perform well alone in ensemble learning. Negative correlation learning with mixed strong and weak learning signals was therefore implemented and tested in this paper. The results suggest that the modified negative correlation learning could well control the complexity of a committee machine in order to solve the learning tasks with robust performance.