A Convex Optimization Approach for Control of Linear Quadratic Systems with Multiplicative Noise via System Level Synthesis
This paper presents a convex optimization-based solution to the design of state-feedback controllers for solving the linear quadratic regulator (LQR) problem of uncertain discrete-time systems with multiplicative noise. To synthesize a tractable solution, the recently developed system level synthesis (SLS) framework is leveraged. It is shown that SLS shifts the controller synthesis task from the design of a robust controller to the design of the entire set-valued closed-loop system responses. To this end, the closed-loop system response is entirely characterized by probabilistic set-valued maps from the additive noise to control actions and states. A bi-level convex optimization over the achievable set-valued closed-loop responses is then developed to optimize the expected value of the LQR cost against the worst-case closed-loop system response. The solution to this robust optimization problem may be too conservative since it aims at enforcing the design constraints for all possible system realizations. To deal with this issue, the presented optimization problem is next reformulated as a chance-constrained program (CCP) in which the guarantees are not intended in a deterministic sense of satisfaction against all possible closed-loop system responses, but are instead intended in a probabilistic sense of satisfaction against all but a small fraction of the system responses. To approximately solve the CCP without the requirement of knowing the probabilistic description of the uncertainty in the system matrices, the so-called scenario optimization approach is employed, which provides probabilistic guarantees based on a finite number of system realizations and results in a convex optimization program with moderate computational complexity. Finally, numerical simulations are presented to illustrate the theoretical findings.