Abstract
Learning and synthesizing stabilizing controllers for unknown nonlinearsystems is a challenging problem for real-world and industrial applications.Koopman operator theory allow one to analyze nonlinear systems through the lensof linear systems and nonlinear control systems through the lens of bilinearcontrol systems. The key idea of these methods, lies in the transformation ofthe coordinates of the nonlinear system into the Koopman observables, which arecoordinates that allow the representation of the original system (controlsystem) as a higher dimensional linear (bilinear control) system. However, fornonlinear control systems, the bilinear control model obtained by applyingKoopman operator based learning methods is not necessarily stabilizable andtherefore, the existence of a stabilizing feedback control is not guaranteedwhich is crucial for many real world applications. Simultaneous identificationof these stabilizable Koopman based bilinear control systems as well as theassociated Koopman observables is still an open problem. In this paper, wepropose a framework to identify and construct these stabilizable bilinearmodels and its associated observables from data by simultaneously learning abilinear Koopman embedding for the underlying unknown nonlinear control systemas well as a Control Lyapunov Function (CLF) for the Koopman based bilinearmodel using a learner and falsifier. Our proposed approach thereby providesprovable guarantees of global asymptotic stability for the nonlinear controlsystems with unknown dynamics. Numerical simulations are provided to validatethe efficacy of our proposed class of stabilizing feedback controllers forunknown nonlinear systems.