Decentralized Sensor Scheduling

Decentralized observers separately observe and estimate the state trajectory of a stochastic dynamic system. At all times, each observer chooses, based on its past information, to make its current observation using one of a finite number of costly, noisy sensors. The observers do not share observations or estimates, yet the objective is to determine causal sensor scheduling policies, and implicitly estimators, which collectively minimize a performance measure coupling the observers´ sensor csts and estimation errors. We show that the observers´ -optimal sensor scheduling policies are non-randomized, open-loop policies that can be determined by solving a deterministic optimal control problem when: the stochastic dynamic system is linear and Gaussian, the observers´ sensors are linear and perturbed by additive white Gaussian noise, and the performance measure is a quadratic function coupling the observers´ sensor costs and estimation errors.