Recursive Gaussian Process-Based Adaptive Control, With Application to a Lighter-Than-Air Wind Energy System

This brief presents a nonmodel-based adaptive control technique that combines principles from machine learning and iterative design optimization with those of continuous-time, falsification-based adaptive control. At the crux of the proposed control strategy are two core elements. First, the recursive Gaussian Process (RGP) modeling is used to maintain an online characterization of the system at hand without the need to maintain a complete database of previously collected measurements (which is required in traditional GP modeling). Second, an adaptation strategy is employed that falsifies candidate controllers from a continuous candidate design space based on desired performance specifications and statistical hypothesis testing. In specific, the control parameter design space is explored by selecting points associated with high uncertainty. Through the use of statistical hypothesis testing, regions of the design space determined to be suboptimal at a user-specified level of confidence are rejected in order to converge to an optimal set of control parameters. The RGP-based adaptation is validated through simulations and laboratory-scale experiments using an airborne wind energy case study. Through these studies, the RGP-based adaptation approach is shown to be effective and is shown to exhibit favorable convergence times when compared with a mature adaptive control technique, extremum seeking (ES).