Defence M. Rotulo: "Results on data-driven controllers for unknown nonlinear systems"

Promotors: Prof. Claudio De Persis, Dr. Pietro Tesi

Abstract: The big data revolution is deeply changing the way we understand and analyze natural phenomena around us. In the field of control engineering, data-driven control enables researchers to explore new intelligent algorithms to model and control complex dynamical systems. Data-driven control is based on the paradigm of learning controllers of an unknown dynamical system by directly using data. The underlying idea is that information about the model can be gathered from experiments, bypassing completely the identification step, which can be impractical or too costly. This thesis presents data-driven control solutions for different families of unknown dynamical systems, with a focus on both linear and special classes of nonlinear ones. In the first part of the thesis, we consider the linear quadratic regulator problem for linear time-invariant discrete-time systems. The system is assumed to be unknown and by collecting a finite set of data we compute the optimal control law in one shot. Secondly, we present an online algorithm for learning controllers applied to switched linear systems. By collecting data on the fly, the control mechanism can capture any changes in the dynamics of the plant and adapt itself accordingly to achieve stabilization of the running dynamics. Finally, we derive data-driven methods for a more general class of nonlinear systems via nonlinearity cancellation. To this end, we make use of a "dictionary" of nonlinear terms that includes the nonlinearities of the unknown system.

Dessirtation