Convergence and stability properties of an adaptive regulator with variable forgetting factor

Abstract A modified version of Fortescue's adaptive regulator is described, based on a minimum-variance estimator with variable forgetting factor and a d-step ahead control law. As in the Fortescue algorithm, the forgetting factor is chosen at each step to keep a measure of the information content constant, but the formula obtained is slightly different and avoids the need for an arbitrary lower bound. It also makes possible a proof of convergence for a deterministic linear system, without additional assumptions. A similar analysis is given for an incremental version of the regulator.