Online kernel learning for time series prediction

• Main Author: Gao, Ning
• Other Authors: Song Qing
• Format: Final Year Project
• Language: English
• Published: 2015
• Subjects: DRNTU::Engineering::Electrical and electronic engineering
• Online Access: http://hdl.handle.net/10356/65632
• Institution: Nanyang Technological University

This report is based on online kernel learning theory which for time series prediction study. Robust recurrent kernel online learning (RRKOL) algorithm is used to do simulation of prediction in different scenarios for comparing different prediction methods of multiple steps prediction in terms of accuracy, convergence speed, time and consistency. Robust recurrent kernel online learning (RRKOL) algorithm based on celebrated real-time recurrent learning (RTRL) approach was discussed. By using MATLAB programming language new algorithm was implemented, and it is able to make multiple steps prediction. Furthermore, RRKOL method can be adopted in commercial use such as make weather prediction as well as car running condition prediction. Since the RRKOL algorithm guarantees weight convergence with regularized risk management through the use of adaptive recurrent hyper parameters for superior generalization performance, among the other existing kernel learning algorithms, the prediction results of RRKOL is more accurate and more consistent (Xu, Song, Haijin, & Wang, 2012). Moreover, because RRKOL combines the advantages of both kernel and recurrent learning methods, it also contains a recurrent learning feedback term which the risk of cost can be minimized. Additionally in this report, the detailed structure update error such that the weight convergence and robust stability proof can be integrated with a kernel sparsification scheme based on a solid theoretical ground was discussed. Furthermore, it’s the first time by using RRKOL algorithm to make multiple steps prediction as a result it will automatically weights the regularized term in the recurrent loss function. In that way, the estimation error can be minimized and the generalization performance via sparsification can also be improved.