Search

In Italy, a committee comprising three members grants tenured positions at the lowest level, namely research assistant professor. These committees draw members from two sources: The institution that makes the [...]

This paper describes a recursive identification method from input-output data, for linear unvariant stochastic multivariable process. This method is based on simultaneous estimation of the model parameters, including the optimal filter gain. The originality lies in the choice of the canonical form retained, and of the related operatorial model. This one lead to a diagonal variance-covariance matrix of the innovation. Moreover, parameters are recursively adjusted, using an estimation (and not a prediction) of the output errors. As a consequence, it can be shown that the estimator exhibit global stability properties.

Problems of classification or segmentation of random signals require to define a distance measure between stationary signals, or, more precisely, between their parametric models. When a signal is represented by a fixed-order ARMA process, one uses an identification method to calculate this model, and to deduce the probability law of a N-point data sequence, or the power spectrum. Most of the known solutions consist then in defining a distance measure either between probability laws (Kullback divergence, Bhattacharyya distance), or between spectra (Itakura-Saito distance, Root-Mean-Square distance for which cepstral distance is a good approximation). But in this approach, one supposes the model to be measured, that is to say to be an exact representation of the signal, while it is only estimated conditionally to a finite data sequence, and the probability law of this estimation depends on the choosed identification method. In this page, we use this important feature to develop a distance between ARMA models identified by Maximum-Likelihood method, taking into account the variance of the estimates. Then, we compare this distance with cepstral one which is known to give good results in classification.