Generalized minimum variance control law revisited

The generalized minimum variance control law is employed in many of the most successful self-tuning control schemes. However, the control law was derived by using a single-stage cost function where a conditional expectation operator was employed. This has limited our understanding of the operation of the control law and resulted in certain operational disadvantages. By rederiving the controller using a normal LQG approach but with a cost function of the same basic form, some of these deficiencies are mitigated. Moreover, the reason that the controller has poor stability characteristics on non-minimum phase systems becomes apparent and it is shown that the magnitude of the error weighting term must be severely restricted in some problems if stability is to be maintained. Thus it may not be possible to limit error variance to the desired extent.