Short or Long Memory Estimators

Abstract This paper contributes to the question of how to choose estimator memory length in system identification. Traditional approaches to this problem have been based on stationary models for parameter time variations. This leads to a fixed trade-off between the “size” of parameter variations and the “size” of noise. However, in practice, one often experiences non-stationary parameter behaviour. The latter problem leads to the desirability of some form of time varying estimator memory. In particular, a short memory is desirable when rapid parameter changes occur. However, a long memory is desirable when infrequent parameters change occur so as to give maximal noise discrimination. The current paper discusses these issues at a conceptual rather than theoretical level. We illustrate the ideas by reference to parameter estimation in zinc galvanizing lines.