The reduction of order of high-order models obtained from an identification experiment is discussed from a statistical point of view. The asymptotic maximum likelihood (ML) approach is defined to reduce the order of an estimated model. This approach considers the maximum likelihood criterion given the asymptotic statistics (both in the number of data and the order) of the estimated model, and corresponds to frequency weighted L2-norm model reduction. By using the insight from the asymptotic ML approach, an identification algorithm is proposed based on a highorder ARX estimate and model reduction via a frequency weighted balanced realization. The advantage of this algorithm is that iterative minimization methods are not required to find the estimate.
IntroductionThe problem of reducing the order of a high-order model obtained from a system identification experiment is considered, i.e. to perform model reduction by taking into account the fact that we have an estimated model. The extra complication in the case of an estimated model is that the model is a random variable, the source of randomness being the identification experiment. The statistical properties of estimated models are complicated, but as shown in § 2 it is possible to give simple expressions for the asymptotic properties, e.g. asymptotic distributions, as the number of observations and the order of the estimated model tend to infinity. Based on these