Time-series prediction using self-organizing fuzzy neural networks

Abstract

A novel online self-constructing fuzzy neural network is proposed for time-series prediction. The proposed approach not only speeds up the learning process but also builds a more parsimonious fuzzy neural network while comparable performance and accuracy can be achieved since the new growing criteria feature characteristics of growing and pruning. The learning scheme starts with no hidden neurons and parsimoniously generates new hidden units according to the proposed growing criteria as learning proceeds. In the parameter learning phase, all free parameters of hidden units are updated by the extended Kalman filter (EKF) method. Simulation results demonstrate that the proposed approach can provide faster learning speed and more compact network structure with comparable generalization performance and accuracy.