UKF-based training algorithm for feed-forward neural networks with application to XOR classification problem

• Published in: 2012 8th International Conference on Natural Computation pp. 316 - 319
• Main Authors: Xiaozhen Zhao, Jiaxiang Yu, Fuwei Li
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.05.2012
• Subjects: Classification algorithms; Covariance matrix; extended Kalman filter; Kalman filters; Mathematical model; Neural networks; Training; unscented Kalman filter; Vectors; XOR classification problem
• ISBN: 9781457721304; 1457721309
• ISSN: 2157-9555
• DOI: 10.1109/ICNC.2012.6234549

This paper uses the recently developed unscented Kalman filter (UKF) to construct a new training algorithm for feed-forward neural networks. This UKF-based training algorithm has the merits of being more accurate and not calculating the derivatives when compared to the training algorithms based on the extended Kalman filter (EKF). Moreover, the UKF can converge more rapidly than the EKF, so the proposed UKF-based algorithm is more suitable for real-time implementation of neural training algorithms. At the end of the paper, the presented algorithm is applied to the XOR classification problem. The classification results demonstrate that the new UKF-based training algorithm performs well in solving the nonlinear XOR classification problem and has superiority over the EKF-based algorithm.