PASCAL - Pattern Analysis, Statistical Modelling and Computational Learning

Consistency of the unlimited BIC Context Tree Estimator
Aurelien Garivier
IEEE-IT Volume 52, 2006.

Abstract

Motivated by the Minimum Description Length principle , the BIC criterion has encountered a certain interest in the last years on model selection issues. In a recent article, Imre Csiszar and Zsolt Talata consider the case of Context Tree sources, a flexible generalization of Markov models widely used in data processing. They prove in particular that the BIC Context Tree estimator is consistent when its size is bounded a priori as o(\log n). We show here that this limitation is not necessary for finite context sources, and provide an algorithm to compute it efficiently.

Postscript - Requires a viewer, such as GhostView
EPrint Type:Article
Project Keyword:Project Keyword UNSPECIFIED
Subjects:Theory & Algorithms
ID Code:1773
Deposited By:Aurelien Garivier
Deposited On:28 November 2005