Inference of candidate loop performance and data filtering for switching supervisory control

The paper studies the problem of inferring the performance of a linear feed back-loop consisting of an uncertain plant and a candidate controller from data taken from the same plant possibly driven by a different controller. I n such a context, a convenient tool to work with is a quantity called norma lized discrepancy. This is a quadratic measure of mismatch between the loop made up by the unknown plant in feedback with the candidate controller and the nominal "tuned-loop" related to the same candidate controller. It is s hown that discrepancy can be in principle obtained by resorting to the conc ept of a virtual reference, and conveniently computed in real-time by suita bly filtering an output prediction error. The latter result is of relevant practical value for on-line implementation and of paramount importance in s witching supervisory control of uncertain plants, particularly in the case of a coarse candidate model distribution. (C) 2001 Elsevier Science Ltd. AL l rights reserved.