Scheduled sampling for robust sensing

Abstract

We consider the problem of optimizing the sensing strategy of a monitoring system in the presence of faulty sensors. We develop ORSg, an efficient data-driven algorithm for computing sampling strategies that nearly maximize the submodular utility of sensing with only a fraction of active and fault-prone sensors. Our approach combines techniques from information theory, game theory and submodular optimization. We empirically evaluate our algorithm with a real-world sensing scenario.