Adaptively fusing neural network predictors toward higher accuracy: A case study

In order to provide function approximation solutions with high accuracy, we employ a multi-learner system that combines a group of component neural networks (CNNs) with an adaptive weighted fusion (AWF) method. In the AWF, the optimization of the normalized weights is obtained with the constrained quadratic programming. Depending on the prediction errors of the CNNs from one input sample to another, the AWF can adaptively adjust the weights which are assigned to the CNNs. The results of the function approximation experiments on six benchmark data sets demonstrate that the AWF method can effectively help the multi-learner system achieve higher accuracy (measured in terms of mean-squared error) of prediction, in comparison with the popular the Bagging algorithm.