Active Sampling for the Quickest Detection of Markov Networks

Consider $n$ random variables forming a Markov field (MRF). The true model of the MRF is unknown, and it assumed to belong binary set. objective sequentially sample (one-at-a-time) such that can be detected with fewest number samples, while in parallel, decision reliability controlled. core element an optimal process rule for selecting sampling over time. Such process, at every time instant adaptively collected data, selects variable expected most informative about model, rendering overall minimized samples required reaching reliable decision. existing studies on detecting structures generally entire network same focus designing detection rules without regard data-acquisition process. This paper characterizes general MRFs, which shown asymptote large . critical insight devising information measure captures decisions’ inherent statistical dependence Furthermore, when MRFs modeled by acyclic probabilistic graphical models, take computationally simple form. Performance analysis case provided, results are interpreted several special cases: Gaussian non-asymptotic regimes, Chernoff’s controlled (active) sensing, problem cluster detection.