Learning to Control Using a Convex Combination of Controllers

Athindran Ramesh Kumar, Peter J. Ramadge

Abstract: We explore the idea of online learning and control in concert with a form of robust control. The objective is to avoid excessive transients during the learning phase. We examine a controller formed as a convex combination of an aggressive and conservative controller. The stability of the convex combination is achieved using an adaptation law obtained from the appropriate derivatives of the spectral radius. We provide numerical results on a sinusoidal reference tracking experiment to verify the efficacy of our technique.