Simultaneous Input Selection and Hidden Nodes Optimization for Sunspot Time Series Prediction

Artificial Neural Network (ANN) is a popular computational intelligence technique in many applications. However, the performance depends on appropriate structure design and considered as a difficult task. Until now, there is no standard procedure in designing an optimal ANN. Two important issues in designing ANN are the selection of significant input subset and hidden nodes size. In this paper, binary coded Genetic Algorithm (GA) has been used to simultaneously select the significant input to ANN and automatically determine the optimal number of hidden nodes. The goal is to obtain optimal ANN design in term of prediction accuracy and complexity. The performances with and without input selection are compared. The proposed method is applied on the prediction of benchmark sunspot time series.