Hybrid neural network ensemble construction combining boosting and negative correlation learning

An ensemble of several neural networks is a convenient way to achieve better performance for a classification task. A number of methods on the basis of different techniques have been investigated for neural network ensemble (NNE) construction from early 1990s. To achieve better performance, a few hybrid NNE methods combining different individual methods are also investigated recently. This paper also presents a hybrid ensemble construction method combining boosting and negative correlation learning (NCL). The proposed method first produces a pool of predefined number of networks using standard boosting and NCL, and then genetic algorithm is used to the task of selecting an optimal subset of networks for an NNE from the pool. The proposed method builds problem-dependent adaptive NNE and shows consistently better performance with concise ensemble over the conventional methods when tested on a suite of 20 benchmark problems.