Kernel-based model order selection for linear system identification
2013 (English)In: Proc. 11th IFAC International Workshop on Adaptation and Learning in Control and Signal Processing, ALCOSP13, 2013, 257-262 p.Conference paper (Refereed)
The mainstream approach to identication of linear discrete-time models is givenby parametric Prediction Error Methods (PEM). As a rule, the model complexity is unknownand model order selection (MOS) is a key ingredient of the estimation process. A dierentapproach to linear system identication has been recently proposed where impulse responsesare described in a Bayesian framework as zero-mean Gaussian processes. Their covariances aregiven by the so-called stable spline, TC or DC kernels that encode information on regularityand BIBO stability. In this paper, we show that these new kernel-based techniques lead alsoto a new eective MOS method for PEM. Furthermore, this paves the way to the design ofa new impulse response estimator that combines the regularized approaches and the classicalparametric PEM. Numerical experiments show that the performance of this technique is verysimilar to that of PEM equipped with an oracle which selects the best model order by knowingthe true impulse response.
Place, publisher, year, edition, pages
2013. 257-262 p.
IdentifiersURN: urn:nbn:se:liu:diva-96768DOI: 10.3182/20130703-3-FR-4038.00043ISBN: 978-390282337-3OAI: oai:DiVA.org:liu-96768DiVA: diva2:643260
11th IFAC International Workshop on Adaptation and Learning in Control and Signal Processing, ALCOSP 2013; Caen; France