Abstract
This paper presents a method of identification of discrete-time linear dynamical systems which is an improvement of the method proposed by G.N. Saridis and G. Stein. The random test input sequence is used in the author's method as well as in Saridis and Stein's one, and a linear regression is obtained by the whiteness of the input. The parameters of the system are estimated through the least mean square type stochastic approximation instead of the Robbins-Monro stochastic approximation by Saridis and Stein.
This method has the following advantages. 1) Few properties of both the system noise and observation noise need be known. 2) The restrictions to these noises are considerably relaxed, for example, the noises may be coloured and mutually dependent. 3) The estimates are renewed at each step. In the method of Saridis and Stein, the noises are restricted to be white and independent, and the estimates are renewed every 2n+1 steps where n is the dimension of the system's state.
In the last place the rate of convergence by this method is investigated and the results of computer similation are presented.
