Online learning and control using data-driven methods

Abstract

Online learning and control refers to identifying a dynamical system on the fly and using thelearned system for model-based control design. In order to be effective, the learning method needs to be both computationally tractable and accurate. To ensure efficiency, an online version of dynamic mode decomposition (DMD) is used to learn the dynamics. To make the set up closer to the lab setting, scenarios with sensor and process noise are considered, and it is shown that the introduction of either sensor or process noise result in bias in the model identified by DMD. Convergence of DMD estimation is also explored in chapter three. For controlling the dynamics, optimal feedback control is employed to minimize a quadratic cost function, and an algorithm adjusted for sensitivity is applied to online learning and control of a linearized inverted pendulum. Finally, an experiment is set up to evaluate these methods in a laboratory setting.