A fast and parsimonious fuzzy neural network (FPFNN) for function approximation

Abstract

In this paper, a novel online self-constructing approach, named fast and parsimonious fuzzy neural network (FPFNN), which emerges the pruning strategy into the growing criteria, is proposed for a function approximator. The restrained growth not only speed up the online learning process but also build a more parsimonious fuzzy neural network while comparable performance and accuracy can be obtained since the growth criterion features characteristics of growing and pruning. The FPFNN starts with no hidden neuron and parsimoniously generates new hidden units according to the restrictive growing criteria as learning proceeds. As the second learning phase, the free parameters of hidden units, regardless of newly created or originally existing, are updated by extended Kalman filter (EKF) method. The performance of FPFNN algorithm is compared with other typical algorithms like RANEKF, MRAN, DFNN and GDFNN, etc., in function approximation. The simulation results demonstrate that the proposed FPFNN algorithm can provide more fast learning speed and more compact network structure with comparable generalization performance and accuracy.