Stochastic Systems Group
Home Research Group Members Programs  
Demos Calendar Publications Mission Statement Alumni

SSG Seminar Abstract

Distributed Estimation with Unreliable Communications

David Castanon
Boston University

Consider the problem of tracking a moving object with a network of sensors.  These sensors are connected to a fusion center by means of communication links that are unreliable. In such networks, retransmission of lost information is expensive and often redundant, as new measurements are available. Modeling the motion as a linear Gaussian dynamical system, and sensor observations as linear measurements with additive Gaussian errors, the optimal estimation algorithm is an event-driven Kalman filter that performs measurement updates whenever measurements are received. Such a model was studied recently by several authors, focusing on the problem of a single sensor communicating raw measurements to an estimator through an unreliable network. The main results provide necessary and sufficient conditions for boundedness of the expected error covariance in terms of the probability of successful communication and the unstable eigenvalues of the system matrix. For typical target tracking dynamics, these results are limited, as the conditions for bounded expected error covariance are always satisfied!

This talk is focused on a different architecture, where multiple sensors are communicating with a central fusion center, and where messages may be dropped due to random interference in channels. In contrast to the previous results, we consider architectures where local processing of the measurements is possible, and where sensors can communicate processed statistics over the channel. We develop communication protocols based on extensions of previous results on optimal distributed estimation that allow the fusion center to compute optimal estimates in the presence of message losses. We compare the performance of our distributed estimation algorithms with alternative algorithms based on communication of observations using several simple examples.

Problems with this site should be emailed to