MAXIMUM-ENTROPY SAMPLING FOR SENSOR NETWORKS- GRAPHICAL MODELS, ALGORITHMICS, AND GENERALIZATIONS
Abstract
Sensor networks are the essential backbone of the growing infrastructure for collection of data and ultimately the use of prescriptive analytics in decision making. With ever-decreasing costs of individual sensors, the opportunity to collect massive amounts of data (e.g., telemetry, environmental, meteorological, geostatistical) is fully upon us. But inevitably, there are physical limits and overall budget limits that must be accounted for. Moreover, communication networks, which support data collection networks, have limited bandwidth, storage capacities, buffer capacities, and processing throughput. With all of these constraints in mind, it is an enormous and important challenge to identify the best locations for monitoring, so that we can quickly make optimal decisions for systems that may rapidly evolve. The problem of determining the optimal set of locations for sensors, in a given system, is a very challenging problem with respect to both modeling and computation. This is due to correlations that exist between data observed by different sensors. Because collected data can have many uses, some not even anticipated at the moment of collection, the accepted practice is to use information theory to compare sets of sensor locations, rather than a narrowly-defined parametric statistical model. The requirements of an information-theoretic model are rather weak, and so the data collected is robustly determined. Our technical approach is based on advanced techniques of mathematical optimization and data analysis. In that context, there is no standard algorithm for handling our problem. So we are developing new techniques to account for the complex nature of our formulation. One key structure that we are exploiting is the idea of sparse graphical models, which account for weak coupling between distant pairs of sensors. We are also taking the best general-purpose optimization algorithms and refining them for our setting. Finally, we are extracting from our ideas, general principles that can be applied to other experimental-design problems. We anticipate being able to solve larger and more realistic instances of sensor-network optimization problems than the current state-of-the-art, with the potential for impact on all uses of sensor networks and beyond.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Mar 07, 2023
- Source ID
- FA95502210172
Entities
People
- Jon Lee
Organizations
- Air Force Office of Scientific Research
- Board of Regents of the University of Michigan
- United States Air Force