Search for: All records

This paper addresses the task of sensor selection over a finite time horizon for systems modeled via discrete-time, linear state-space representations. Our method for this general linear setting accommodates both spatial and temporal noise correlations. To our best knowledge, this is the first work to do so. Scheduling policies are designed to limit sensor usage and minimize a minimum-mean-square-error-based criterion with time-varying weights to accommodate different user scenarios (e.g., prioritizing certain state elements at certain times or performing linear quadratic Gaussian control). The approach is also nonmyopic since the effects of sensor activations on all time steps are incorporated. A new but algebraically equivalent formulation of the scheduling model is introduced that readily accounts for colored noise sequences. This lends a closed-form expression for the error covariance that is explicit in all scheduling variables. Such an expression had been considered intractable for filtering in both white and colored noise regimes. This expression is leveraged to develop a well-motivated surrogate objective function that is shown to be submodular, thus enabling the use of an efficient greedy algorithm accompanied by performance guarantees with respect to the surrogate objective. Numerical examples are provided to demonstrate the effectiveness of the proposed methodology.