Online Actuator Selection and Controller Design for Linear Quadratic Regulation with Unknown System Model

Abstract

We study the simultaneous actuator selection and controller design problem for linear quadratic regulation over a finite horizon, when the system matrices are unknown a priori. We propose an online actuator selection algorithm to solve the problem which specifies both a set of actuators to be utilized and the control policy corresponding to the set of selected actuators. Specifically, our algorithm is a model based learning algorithm which maintains an estimate of the system matrices using the system trajectories. The algorithm then leverages an algorithm for the multiarmed bandit problem to determine the set of actuators under an actuator selection budget constraint and also identifies the corresponding control policy that minimizes a quadratic cost based on the estimated system matrices. We show that the proposed online actuator selection algorithm yields a sublinear regret.

Self-Tuning Network Control Architectures

“…Some works have studied time-varying actuator scheduling [11]- [15], but the architecture, while time-varying, remains in open-loop and does not adapt to changing network state or dynamics. Very recent work has considered selecting actuators for uncertain systems based on data measured from the system in limited settings with linear systems and specific controllability metrics [16], [17].…”