A modified EM algorithm for parameter estimation in linear models with time-dependent autoregressive and t-distributed errors

Research output: Chapter in book/report/conference proceedingConference contributionResearchpeer review

Authors

Research Organisations

External Research Organisations

  • University of Bonn
View graph of relations

Details

Original languageEnglish
Title of host publicationITISE 2017
Pages1132-1145
ISBN (electronic)978-84-17293-01-7
Publication statusPublished - 2017

Abstract

We derive an expectation conditional maximization either (ECME) algorithm for estimating jointly the parameters of a linear regression model, of a time-variable autoregressive (AR) model with respect to the random deviations, and of a scaled t-distribution with respect to the white noise components. This algorithm is shown to take the form of iteratively reweighted least squares in the estimation of the parameters both of the regression and time-variability model. The fact that the degree of freedom of that distribution is also estimated turns the algorithm into a partially adaptive estimator. As low degrees of freedom correspond to heavy-tailed distributions, the estimator can be expected to be robust against outliers. It is shown that the initial stabilization phase of an accelerometer on a shaker table can be modeled parsimoniously and robustly by a Fourier series with AR errors for which the time-variability model is defined by cubic polynomials.

Keywords

    Linear regression model, time-dependent AR process, partially adaptive estimation, robust parameter estimation, EM algorithm, iteratively reweighted least squares, scaled t-distribution

Cite this

A modified EM algorithm for parameter estimation in linear models with time-dependent autoregressive and t-distributed errors. / Kargoll, Boris; Omidalizarandi, Mohammad; Alkhatib, Hamza et al.
ITISE 2017. 2017. p. 1132-1145.

Research output: Chapter in book/report/conference proceedingConference contributionResearchpeer review

Download
@inproceedings{59090c73b64c4f138f86baa8f6551a41,
title = "A modified EM algorithm for parameter estimation in linear models with time-dependent autoregressive and t-distributed errors",
abstract = "We derive an expectation conditional maximization either (ECME) algorithm for estimating jointly the parameters of a linear regression model, of a time-variable autoregressive (AR) model with respect to the random deviations, and of a scaled t-distribution with respect to the white noise components. This algorithm is shown to take the form of iteratively reweighted least squares in the estimation of the parameters both of the regression and time-variability model. The fact that the degree of freedom of that distribution is also estimated turns the algorithm into a partially adaptive estimator. As low degrees of freedom correspond to heavy-tailed distributions, the estimator can be expected to be robust against outliers. It is shown that the initial stabilization phase of an accelerometer on a shaker table can be modeled parsimoniously and robustly by a Fourier series with AR errors for which the time-variability model is defined by cubic polynomials.",
keywords = "Linear regression model, time-dependent AR process, partially adaptive estimation, robust parameter estimation, EM algorithm, iteratively reweighted least squares, scaled t-distribution",
author = "Boris Kargoll and Mohammad Omidalizarandi and Hamza Alkhatib and Wolf-Dieter Schuh",
year = "2017",
language = "English",
pages = "1132--1145",
booktitle = "ITISE 2017",

}

Download

TY - GEN

T1 - A modified EM algorithm for parameter estimation in linear models with time-dependent autoregressive and t-distributed errors

AU - Kargoll, Boris

AU - Omidalizarandi, Mohammad

AU - Alkhatib, Hamza

AU - Schuh, Wolf-Dieter

PY - 2017

Y1 - 2017

N2 - We derive an expectation conditional maximization either (ECME) algorithm for estimating jointly the parameters of a linear regression model, of a time-variable autoregressive (AR) model with respect to the random deviations, and of a scaled t-distribution with respect to the white noise components. This algorithm is shown to take the form of iteratively reweighted least squares in the estimation of the parameters both of the regression and time-variability model. The fact that the degree of freedom of that distribution is also estimated turns the algorithm into a partially adaptive estimator. As low degrees of freedom correspond to heavy-tailed distributions, the estimator can be expected to be robust against outliers. It is shown that the initial stabilization phase of an accelerometer on a shaker table can be modeled parsimoniously and robustly by a Fourier series with AR errors for which the time-variability model is defined by cubic polynomials.

AB - We derive an expectation conditional maximization either (ECME) algorithm for estimating jointly the parameters of a linear regression model, of a time-variable autoregressive (AR) model with respect to the random deviations, and of a scaled t-distribution with respect to the white noise components. This algorithm is shown to take the form of iteratively reweighted least squares in the estimation of the parameters both of the regression and time-variability model. The fact that the degree of freedom of that distribution is also estimated turns the algorithm into a partially adaptive estimator. As low degrees of freedom correspond to heavy-tailed distributions, the estimator can be expected to be robust against outliers. It is shown that the initial stabilization phase of an accelerometer on a shaker table can be modeled parsimoniously and robustly by a Fourier series with AR errors for which the time-variability model is defined by cubic polynomials.

KW - Linear regression model, time-dependent AR process, partially adaptive estimation, robust parameter estimation, EM algorithm, iteratively reweighted least squares, scaled t-distribution

M3 - Conference contribution

SP - 1132

EP - 1145

BT - ITISE 2017

ER -

By the same author(s)