Abstract :
An attempt to reconcily performance and robustness in adaptive control is discussed in the framework of algorithms with moderate complexity. The basic idea is to recognize that ARMAX plants, when properly controlled, can be described by ARX models. Thus, standard RLS identification can be used. In addition, instead of relying on a single estimated plant predictor, a set of separately estimated predictors, describing the plant evolution over the control horizon, can be used. While the overall numerical complexity is bounded by both the use of RLS and a unique regressor for all predictors, the intrinsic model redundancy of the multipredictor scheme makes the related adaptive control algorithms robust with respect to plant unmodelled dynamics.
