Topic: Message-Passing Algorithms for Optimizing Decentralized Detection Networks
By: O. Patrick Kreidl (MIT), Alan Willsky (MIT)
A promising feature of emerging wireless sensor networks is the opportunity for each node to process data about \"locally\" sensed activity and then communicate relevant information output, altogether in a manner that supports \"globally\" effective decision-making. We consider a global objective of solving a Bayesian detection problem, using a decentralized model that is most appropriate for sensor network applications where communication resource is scarce (e.g., voluminous local data exceeds stipulated link capacities, active single-hop connections are sparse) but the per-node computation and memory resource is in relative abundance. Under certain assumptions, the decentralized detection problem is known to admit a set of necessary optimality conditions, leading to a system of multi-variable nonlinear equations for which an algorithmic solution relies on an iterative application of the single-sensor computation. We identify additional assumptions that allow this algorithmic computation to exhibit a ``message-passing\'\' property, greatly facilitating its distributed implementation: the computation can be executed in a manner where each sensor node need only be aware of information that is either (i) locally available at initialization or (ii) provided by messages from only its immediate neighbors in the network. Convergence issues, message-scheduling protocols and drawing connections to existing message-passing algorithms (e.g., belief propagation) are all topics for further investigation.
Questions and comments to the webmaster. This page is copyrighted by the Massachusetts Institute of Technology.