Selection of sufficient stochastic data-models applied to modeling

R. Moddemeijer and L. Spaanenburg

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

The constructive derivation of the sufficient model for analysis and design of an electronic circuit assumes the availability of a rich set of alternatives. Rather than to iterate the selection of a single model by trial-and-error, we advocate here the structured elimination through the application of log-likelihood statistics. This method was originally developed for auto-regressive (AR) order estimation. The generic architectural model is validated once to establish the desired balance between model bias caused by insufficient modeling and additional variance caused by superfluous parameters. The result is one or more functionally indistinguishable model implementations. From restrictions on the scarce availability of realization resources, the most suitable model for design and analysis can finally be picked. This top-down style of model derivation is illustrated in the analysis of an analogue integrated circuit. It shows how already from a limited set of experimental data a sufficient model can be pinpointed. This allows to reduce the long and painstaking model derivation for new devices and macro-models to a matter of CPU minutes.


Full paper , pdf-format


Poster presentation


Published

Proceedings of the 10th workshop on Circuits, Systems and Signal Processing, (IEEE/ ProRISC ), 25 - 26 November 1999, Mierlo (NL), pp. 307-316, eds. Jean Pierre Veen, STW Technology Foundation, Utrecht (NL), ISBN: 90-73461-18-9, BibTeX , Note: proceeding are published on CD-rom only