Principal Investigator Alan Willsky
This research addresses a fundamental class of statistical decision problems (e.g., large-scale hypothesis-testing) under the assumptions that measurements are distributed over multiple sensors and there exist explicit costs or constraints on the available algorithmic resources (e.g., computation, communication, memory). Such assumptions apply in a variety of forward-looking engineering developments, including (1) wireless sensor networks, in which every symbol exchanged is a power drain and brings the communicating nodes closer to expiration, and (2) real-time network security, in which extremely fast sensing rates and quality-of-service objectives at each node can tolerate only minimal computation/memory per cycle. Our work extends the models and methods at the intersection of team decision theory and graph-based probabilistic inference, culminating in a new class of efficient message-passing algorithms to be executed "offline" (i.e., before processing measurements) by which the nodes "self-organize" (i.e., iteratively couple the parameters of their local processing rules) in order to mitigate the loss in performance resulting from the explicit "online" resource constraints (i.e., the constraints when making decisions from measurements).