On the Maximum Likelihood Method of Identification

The maximum likelihood principle of estimation applied to the linear black-box identification problem gives models with theoretically attractive properties. Also, the method has been applied to industrial data (various processes in paper production) and proved able to work in practice. This paper presents further developments of the method in the case of a single output. The reliability and speed of the identification algorithm have been improved, and the method has been made easier to use. A rather sophisticated computer program, however, was needed. It employs a generalized model structure, an improved hill-climbing algorithm, and an automatic procedure for determining model orders and transport delays. Some statistics from performance tests of the program are presented.