Recursive methods for off-line identification

Abstract

Many recursive identification methods have been developed for on-line estimation of parameters in dynamical systems. Such methods can also be applied to data for off-line problems, usually by passing the data through the recursive identifier a few times. In this paper such an approach is discussed. Suppose that a recursive identification method is applied to a data string obtained by concatenation of the original finite-data record. It is proved that the recursive identification method will then converge to a (local) minimum of the off-line finite-data identification criterion. We also provide some arguments as to why recursive identification may be a competitive way of performing the minimization. When using the data more than once it is a priori not clear how to treat the initial values and the forgetting factor. The problem is investigated and comparisons are made with an off-line prediction error method. Practical experiments show that the execution time for a repeated recursive algorithm is less than or equal to that of the off-line algorithm, and that the obtained accuracies are comparable.