Data-Driven Kalman Filter using Maximum Likelihood Optimization
This paper investigates the state estimation problem for unknown linear systems with process and measurement noise. A novel data-driven Kalman filter (DDKF) that combines model identification with state estimation is developed using pre-collected input-output data and uncertain initial state information of the unknown system. Specifically, the state estimation problem is first formulated as a non-convex maximum likelihood (ML) optimization problem. Then, to reduce the computational complexity, the optimization problem is broken down into a series of sub-problems in a recursive manner. Based on the optimal solutions to the sub-problems, a closed-form DDKF is designed for the unknown system, which can estimate the state of a physically meaningful state-space realization, rather than these up to an unknown similarity transformation. The performance gap between the DDKF and the traditional Kalman filter with accurate system matrices is quantified through a sample complexity bound. In particular, when the number of the pre-collected trajectories tends to infinity, this gap converges to zero. Moreover, the DDKF is used to facilitate data-driven control design. A data-driven linear quadratic Gaussian controller is defined and its closed-loop performance is characterized. Finally, the effectiveness of the theoretical results is illustrated by numerical simulations.
PDF Abstract