Incremental negative correlation learning with evolutionary selection of parameters

Incremental learning is attracting more and more interest in the field of machine learning due to its wide potential applications in many scientific and engineering areas. Negative correlation learning (NCL) (Liu and Yao; 1999a,b) is a successful approach to construct neural network ensembles. By encouraging the diversity of ensembles, it makes different neural networks to learn different knowledge of the incoming data so that the ensembles can learn the whole data better. Its partial learning effect can help ensembles overcome the problem of catastrophic forgetting. These features make NCL a potentially powerful approach to incremental learning. In previous researches, it has been found that Incremental NCL algorithms are very sensitive to their parameters. In this paper an approach based on evolutionary computation techniques is proposed to tackle the problem of automatic and robust parameter setting for Incremental NCL. Via typical comparative experiments, the proposed approach exhibit clearly improved performance over existing algorithms.