An efficient algorithm for selecting optimal configurations of AR-coefficients

R. Moddemeijer

University of Groningen, Department of Computing Science,
P.O. Box 800, NL-9700 AV Groningen, The Netherlands,
phone: +31.50.363 3940 - fax: +31.50.363 38005 - e-mail: rudy@cs.rug.nl

Abstract

There exists an essential difference between the correct Auto Regressive (AR) model and the optimal AR-model. We try to find an optimal model balancing between flexibility, using many AR-parameters, and low variance, using only a few AR-parameters. We select an optimal AR-parameter configuration consisting of zero and non-zero parameters given a maximum AR-order. This optimal configuration will be selected using a Modified Information Criterion (MIC) which is closely related to Akaike's criterion (AIC). This MIC allows an a priori selection of the probability of estimating too many parameters.

We present the method and a verification by simulations. The method is based on pivoting the Hessian matrix by Gauss-Jordan pivots. As a result we can now select an optimal parameter configuration with an a priori probability of selecting a configuration with a too large number of parameters given an a priori selected maximum AR-order.


Full paper


Presentation


Published

Twentieth Symposium on Information Theory in the Benelux, May 27-28, 1999, Haasrode, Belgium, pp 189-196, eds. A. Barbé et. al., Werkgemeenschap Informatie- en Communicatietheorie, Enschede, The Netherlands, and IEEE Benelux Chapter on Information Theory, ISBN 90-71048-14-4 BibTeX


Related work


other publications