Graphical Models for Distributed Inference in Wireless Sensor Networks

Distributed information fusion is an active area of research; however, fusion in the large, dynamic, unpredictable, and power-scarce setting of wireless sensor network (WSN) requires more than just distributed fusion algorithms. Some of the important questions that need to be addressed are dynamic mapping of inference responsibilities to sensor nodes in a distributed manner and cost-effective, fault-tolerant exchange of fusion information. Graphical models provide compact representation of joint probability distributions; they help in drawing inference more efficiently. We propose two graphical models called inference cost network (ICN) and dynamic inference cost network (DICN) that generalize Bayesian networks and dynamic Bayesian networks for truly distributed implementation. We also propose distributed inference algorithms for ICN/DICN that address the unique requirements of information fusion in WSN. We prove correctness of the model and the algorithms and demonstrate that the algorithms are lightweight in terms of computational complexity.