Developing methods for robust distributed data fusion

Developing methods for robust distributed data fusion



Distributed data fusion is a key enabling technology for distributed tracking and sensing. Such networks consist of a set of nodes - some equipped with sensors to collect data, others to fuse data and communicate it to other nodes in the network. By communicating fused estimates rather than raw sensor data, substantial reductions in required bandwidth can be obtained together with robust, scalable and modular operations. However, a critical limitation of Bayesian data fusion algorithms is that the probability density function - or at least the correlations - between different estimates must be known. If naive assumptions of independence are used, highly inaccurate estimates can result. However, developing a full description of the probability distribution requires all nodes in the network to have full knowlege of the entire state of the network, thus undermining all practical advantages of a distributed fusion system.
One means of overcoming these difficulties is to develop a robust data fusion system which trades optimality for robustness. In this talk I shall describe a technique which replaces the normal product in Bayes rule by an exponential mixture of the densities. Preliminary results suggest that the algorithm works extremely well, but the full theoretical reason for its robustness is still unclear. I shall discuss related results in alpha divergence and robust two-class detection problems.