An alternative approach for maximum likelihood identification

A maximum likelihood identification procedure, based on the likelihood function of the original measurements, is derived for a discrete linear time-varying dynamic model via the theory of the E-M algorithm. The proposed scheme yields both the smoothed state estimate and the maximum likelihood estimate of unknown parameters, and therefore should be a useful data analysis tool. Moreover, since the state estimate and parameter identification are solved separately in this procedure, the computation involved should become considerably simpler. In particular, for the case where only the statistics of the initial state, process noise and measurement noise are to be identified, the problem is decomposed into three separate problems, and no numerical optimization will be needed.