WNN Learning Algorithm Based on Unscented Kalman Particle Filter
Download as PDF
Niu Junfeng, Wu Yarong, Li Huaping
In order to improve the nonlinear modeling capability of Wavelet Neural Network (WNN), a learning algorithm of WNN based on modified Unscented Kalman Particle Filter (UPF) is proposed. In the algorithm, first a minimal skew strategy is introduced to reduce the number of Sigma sampling points of Unscented Transform (UT), improving Unscented Kalman Filter (UKF), and then the improved UKF is used to select the importance density function of Particle Filter (PF), forming a new UPF (SUPF), finally, SUPF is taken as learning algorithm of WNN for training and test. The simulation results indicate that, for the learning problem of WNN, the model precision of UPF based on new sampling strategy is approximately close to that of simple UPF, but the former has faster training rate and higher learning efficiency, which validate its feasibility and effectiveness.
Wavelet Neural Network, Kalman Filter, Unscented Transform, Particle Filter.