GA-based input features and learning parameters selection method for decorrelated neural network ensembel model

Using more features than needed as inputs decreases prediction performance and interpretation ability of the learning model. Ensemble learning-based soft measuring model has better generalization performance than that based on single model. Negative correlation learning and random vector functional link networks based decorrelated neural network ensembles (DNNE) can overcome some shortcomings of error back-propagation neural networks (BPNNs) in term of effective and efficient. However, its performance is sensitive to some learning parameters. Thus, genetic algorithm (GA) is used to select input features and leaning parameters of DNNE model jointly. Six benchmark datasets are used to validate the proposed method.