Improving RELS local convergence via long-range predictive identifiers

A recursive autoregressive moving-average (ARMA) identifier based on multistep prediction error minimization is considered in order to sidestep the positive realness conditions of recursive extended least squares (RELS). For a single-step prediction horizon, the algorithm reduces to RELS while, as shown by local convergence analysis, it tends to be a local approximation of the recursive maximum likelihood (RML) algorithm as the prediction horizon increases.<>