Critical Review of 30 Years of Adaptive Control

Monday, November 8, 2010: 2:35 PM
250 B Room (Salt Palace Convention Center)
B. Erik Ydstie, Department of Chemical Engineering, Carnegie Mellon University, Pittsburgh, PA

Adaptive control experienced its heyday during the 1980ies when a large number of researchers in the fields of chemical and electrical engineering were engaged in developing algorithms and theories for stable and robust adaptive control of a wide range of systems. Successful applications were made to a range of chemical processes and significant breakthroughs were made in adaptive control theory. Many of these ideas were reviewed in Adaptive control strategies for process control: A survey, by Seborg, Edgar and Shah, a paper which provided an exhaustive review of the state of art. It might have been expected that more research and many more applications would follow since adaptive control appeared to meet industry need and the theory and algorithms were nearing a stage of maturity. This turned out not be the case. Few industrial implementations followed (to the disappointment of some) and the research field fizzled out already in the early 1990ies as many researchers moved to other fields. Few new ideas have been presented since then and sessions on the topic of adaptive control disappeared from the AIChE and other meetings. The purpose of the current paper is to briefly describe the history of adaptive control. Focus is placed on the so-called linear certainty equivalence approach where identification is used to estimate a model and the model is used to design a control. This approach was espoused by Tom Edgar and many other researchers since it provided a very rich framework for developing new algorithms and theories. CE adaptive control has also been applied to many industrial problems. The strengths and weaknesses of the CE approach will be reviewed and critiqued. It will be shown why this approach fails to provide a compelling paradigm for robust adaptive control of industrial processes and why modifications must be introduced. The main reasons are that data provided in closed loop cannot be sufficiently rich and that uncontrollable models may result. A third problem, which was overlooked was that model changes should be linked to the tuning procedure in order to achieve robust performance. Taking all these issues into account required more computational resources and signal processing than was thought to be available on-line at that point in time. Two decades have now passed since the review provided by Seborg, Edgar and Shah. Computers are much faster, algorithms are better and adaptive control theory has advanced. This may be an opportune moment for our community to revisit this problem.

Extended Abstract: File Not Uploaded
See more of this Session: In Honor of Tom Edgar's 65th Birthday I
See more of this Group/Topical: Computing and Systems Technology Division