A cancelation method for ARMA identification with overestimated orders

Identification of an unknown ARMA model consists in estimating its coefficients and its orders. Several algorithms arc often used to solve this problem. The ELS algorithm need to satisfy the SPR condition to converge. The RML algorithm has no known Lyapounov function to prove that it converges. The averaging method apply to the RML algorithm gives an algorithm which minimizes a cost function based on the prediction error. In the case of overestimated orders, this function has several global minima. We propose a particular version of this algorithm, which is stable and converges to one of these global minima. We design also a method which gives the correct orders and the coefficients estimated.