📝 SummarXiv

Abstract

Technological advancements in miniaturization and wireless communications are yielding more affordable and versatile sensors and, in turn, more applications in which a network of sensors can be actively managed to best support overall decision-making objectives. We propose modeling the opportunity for sensor management within multi-stage stochastic control problems with imperfect state information. Such formulations inherently assume the state of the modeled environment cannot be accessed directly but instead the controller can observe only noisy measurements of the state and, therefore, at each decision stage some form of state estimation is required before a control is actuated. The notion of sensor management arises when the modeled controls not only affect the subsequent evolution of the state but can also affect the nature of future measurements and, hence, the quality of state estimates that drive future control decisions. In principle, the optimal strategy for any appropriately modeled multi-stage stochastic control problem with imperfect state information (with or without opportunity for sensor management) is the solution to a dynamic program; in practice, the computational requirements are typically prohibitive yet dynamic programming methods are still useful to guide the development of effective suboptimal strategies. In this spirit, we model the opportunity for sensor management within small-scale examples of two well-studied dynamic programming formulations, namely (1) the finite-state/finite-action Partially-Observable Markov Decision Process (PO-MDP) and (2) the Linear-Quadratic-Gaussian Regulator (LQGR). These examples admit solvable dynamic programs and confirm how the interplay between sensing and acting is a natural by-product of a dynamic programming solution.