Model order selection criteria: comparative study and applications
Zbigniew M. Leonowicz, Juha Karvanen, Toshihisa Tanaka, Jacek Rezmer · Repozytorium Eny Politechnika Wrocławska · 2004
A practical application of informa- tion theoretic criteria is presented in this paper. Eigenvalue decomposition of the signal correlation matrix{based AIC, MDL and MIBS criteria are in- vestigated and used for on{line estimation of time{ varying parameters of harmonic signals in power sys- tems. Determination of the model order arises in many areas of signal processing. In this paper we will focus on approaches based on eigenvalue decompo- sition of the signal correlation matrix (time{delayed in vector signal case). Wax and Kailath (1985) pre- sented a new approach for estimating the number of signals in multichannel time{series, based on statis- tical classiflcation criteria AIC (Akaike Information Criterion) and MDL (Minimal Description Length Criterion) (3). Use of such statistical criteria re- solves the problem of estimation of the signal and subspace dimension, which is necessary to obtain the correct estimates od the signal parameters, us- ing the methods considered in this work (4). New criterion (5) based on Bayesian statistics will be also investigated. II. Estimation of the order of the model A. Information theoretic criteria Wax and Kailath (10) presented a new approach for estimating the number of signals in multichannel time{series, based on statistical classiflcation crite- ria AIC and MDL. This approach does not require any subjective threshold setting. B. Approach based on \observation