A hybrid variable selection algorithm for multi-layer perceptron with nonnegative garrote and extremal optimization

In the paper, a new hybrid variable selection algorithm for nonlinear regression multi-layer perceptron (MLP) is proposed. The proposed algorithm applies nonnegative garrote (NNG) to compress the input weights of the MLP. The zero input weights dependent variables will be removed from the initial dataset. Next, a further variable selection is carried out by extremal optimization (EO) algorithm. The new variable selection algorithm integrates powerful global selection ability of NNG and accurate local search ability of EO. Finally, two examples of artificial data sets and an industrial application for a debutanizer column are implemented to demonstrate the performance of the new algorithm. The simulation result demonstrates that the developed algorithm presents d better model performance along with less input variable selected than other state-of-art variable selection methods.