システム同定における数値計算上のパラメータ推定精度

抄録

In identification of linear discrete timeinvariant systems, there are some systems whose parameters are difficult to be estimated by least square method (LS) under the condition that unbiased consistent estimators are given. The problem of parameter estimation by LS reduces to numerical analysis of simultaneous linear equation. In this paper, it is shown that these systems have large condition number of data matrix which is the coefficient matrix of the simultaneous linear equation to be solved. Further, in these systems, it is pointed out that the estimated parameters by recursive least square method (RLS) converge to the true values of system parameters, but numerical unstable phenomena often appear in RLS algorithm.