Center for Applied Mathematical Sciences Visit USC
USC Donrsife Homepage
Monday
22
Oct 12
3:30 PM - 4:30 PM KAP 414
Data-Efficient Quickest Change Detection
Venugopal Veeravalli University of Illinois-Urbana

In the classical quickest change detection problem, there is a sequence of observations whose distribution changes at an unknown time, and the goal is to detect the change as quickly as possible, subject to a false alarm constraint. In many engineering applications of quickest change detection, there may be a cost associated with acquiring observations. We therefore consider the quickest change detection problem with an additional constraint on the cost of the observations used in the detection process, i.e., we seek algorithms that are data-efficient or energy-efficient. The objective is to select an observation control policy along with the stopping time at which the change is declared, so as to minimize the average detection delay, subject to both a false alarm constraint as well as a constraint on the average number of observations used before the change point. We consider both Bayesian and minimax formulations of the problem, and extensions to the setting where the observations are available at a set of distributed sensors. In all of these cases, we develop algorithms that are first-order asymptotically optimal as the false alarm rate goes to zero. We also demonstrate through numerical studies that our algorithms can be considerably more efficient than algorithms based on fractional sampling, where the observations to be skipped are determined a priori in order to meet the observation constraint.

(This is joint work with Taposh Banerjee.)

Previous colloquium: Bayes and Empirical Bayes Information (Learning from the experience of others) Next colloquium: Stability theory of polytropic gaseous stars