Knowledge Discovery and Data Mining KERNEL-BASED METHODS FOR NON-STATIONARY TIME-SERIES IDENTIFICATION AND PREDICTION

Identification and prediction problem of nonlinear time-series generated by discrete dynamic system is considered via Kernel Method approach. A unified approach to recurrent kernel identification algorithms design is proposed. In such a way a recurrent modification of initial Kernel Method with growing windows is considered. In order to prevent the model complexity increasing under on-line identification, the reduced order model kernel method is proposed and proper recurrent identification algorithms are designed along with conventional regularization technique. Such an approach leads to a new type of Recursive Least-Square Kernel Method identification algorithms. Finally, the recurrent version of Sliding Window Kernel Method is also developed along with suitable identification algorithms. The proposed algorithm has tracking properties and may be successfully used for on-line identification of nonlinear non-stationary time-series.