Distributed Sensing and Processing: A Graphical Model Approach
Abstract
The cost of direct computation of the detection error probabilities in a sensor network -- e.g., the probability of false alarm, the probability of a miss, or the average probability of error -- is combinatorial with the number of sensors. This limits the design of the optimal detector to networks with a very small number of sensors. Our work developed a simple very accurate large-deviation method to compute these probabilities of error based on the saddle-point approximation. The saddle-point approximation can be used with networks with an arbitrary (small or large) number of sensors. We used it to resolve three major network issues: design the optimal distributed sensor network detectors in the Neyman-Pearson and Bayes criteria; establish the performance tradeoffs among different network parameters like the signal-to-noise ratio, the number of network sensors, and the number of bits quantizing the local network (soft-) decisions; and the design of the optimal sensor network topology. Our methods apply to generic parallel and distributed (web like) architectures. We showed that Ramanujan graph topologies maximize the convergence rate of distributed detection consensus algorithms, improving over three orders of magnitude over the speed of convergence of structured networks (e.g., nearest neighbor type networks) and close to two orders of magnitude over small world type network designs.
Document Details
- Document Type
- Technical Report
- Publication Date
- Nov 30, 2005
- Accession Number
- ADA455686
Entities
People
- Jose M. Moura
Organizations
- Carnegie Mellon University