Rethinking adaptive control for the 1990s

A new philosophy of identification for use in control design is proposed. The current philosophy consists of attempts to find the single best plant model at each time. An example demonstrates the serious problems that may result from using a single model in a control law during the transient phase of identification when this model is not likely to be correct. The new philosophy proposes maintaining a set of possible plant models and eliminating models which are inconsistent with incoming data. The model set should then shrink so that control laws which provide improved performance for the entire model set become possible