Failures of Adaptive Control Theory and their Resolution

Abstract

Adaptive control is a very appealing technology, at least in principle. Yet its use has been conditioned by an attitude of distrustfulness on the part of some practitioners. In this paper, we explain why such distrustfulness is warranted, by reviewing a number of adaptive control approaches which have proved deficient for some reason that has not been immediately apparent. The explanation of the deficiencies, which normally were reflected in unexpected instabilities, is our main concern. Such explanations, coupled with remedies for avoiding the deficiencies, are necessary to engender confidence in the technology. These include the unpredictable failure of the MIT rule; the bursting phenomenon, and how to prevent it; the Rohrs' counterexample, which attempted to disqualify all adaptive control algorithms; the notion that identification of a plant is only valid conceptually for a restricted range of controllers (with the implication that in adaptive control, certain controller changes suggested by adaptive control algorithms may introduce instability); and the concept of multiple model adaptive control.