Inference Degradation in Information Fusion: A Bayesian Network Case

Abstract

Dynamic and active information fusion processes select the best sensor based on expected utility calculation in order to integrate the evidences acquires both accurately and timely. However, inference degradation happens when the same/similar sensors are selected repeatedly over time if the selection strategy is not well designed that considers the history of sensor engagement. This phenomenon decreases fusion accuracy and efficiency, in direct conflict to the objective of information integration with multiple sensors. This chapter tries to provide a mathematical scrutiny of this problem in the myopia planning popularly utilized in active information fusion. In evaluation it first introduces the common active information fusion context using security surveillance applications. It then examines the generic dynamic Bayesian network model for a mental state recognition task and analyzes experimentation results for the inference degradation. It also discusses the candidate solutions with some preliminary results. The inference degradation problem is not limited to the discussed task and may emerge in variants of sensor planning strategies even with more global optimization approach. This study provides common guidelines in information integration applications for information awareness and intelligent decision.