1986 年 22 巻 7 号 p. 739-744
This paper is concerned with the improvement of convergence rate of RML (Recursive Maximum Likelihood) method in ARMA parameter estimation, A recursive parameter estimator is developed by using Gauss-Newton method. Main difference the proposed method with RML one is the estimation of the expectation of Hessian in Gauss-Newton method. Proposed method estimates this quantity so as to minimize the estimation error of Fisher's information matrix which is considered as the information processing ability of RML estimator. Resultant algorithm is constructed by linking RML estimator with extended Kalman filter. The feature of proposed method is that the stability monitoring is not required in intermediate computation.
It is shown in numerical experiment that the convergence rate of the parameter estimates is improved by using proposed method.