The Need-to-Know Filter for Approximate Inference

Bayesian Agents and Object Oriented Bayesian networks (OOBN) and Dynamic Bayesian Networks (DBN) are new Bayesian Network (BN) representations enabling on-the-fly BN construction that is especially useful in the situational awareness of high-level sensorfusion and for missile defense systems that detect, identify and assess anomalies and threats. Yet a large BN can be computation-ally difficult to solve using exact solution techniques.Human organizations use approximation in integrating large quantities of information. In sensor fusion, and in intelligence organizations filters limit the number of contacts that get through to analysts. Without such filters the analyst is oftenoverloaded at critical moments. Only reports for which there is a need-to-know (NTK) are included in the fusion analysis.In this SBIR project a similar NTK filter is applied to propagation algorithms used for Bayesian networks and systems of Bayesian agents. The project develops a new approximation algorithm: Information flows across an interface only if the message containsnovel information that is of value to the receiver. One measure of novelty is the relative entropy measure of change in the distributions of the in-terface prior and posterior to the flow. Several variants of the NTK filter are ex-plored including biddingfor bandwidth needed to transmit a message. Agents consult potential information servers to determine the novelty of information that could be available if the server were incorporated into the situational awareness. The winning bid identifies the nextbest piece of information.The feasibility, performance and usefulness of the filter are demonstrated with an example from remote sensor fusion. Bayesian Networks and Agent Systems constitute a technology that is especially in tune with high-level fusion, where diverse information sources, both sensors and human intelligence, must be reconciled, and lead to threat assessment. In the sensortracking market, Bayesian methods have a lengthy history. The net-work representation brings considerable capability in learning, modeling, and reasoning in large complex systems, but the representation power is far ahead the ability to obtain solutions inlarge networks. This SBIR R&D, if successful, will provide tools and products of use to the MDA and the greater remote sens-ing, surveillance, and intelligence communities in both government and commer-cial markets.