Distributed belief propagation and its generalizations for location-aware networks

Abstract

This thesis investigates the use of generalized belief propagation (GBP) and belief propagation (BP) algorithms for distributed inference. The concept of a network region graph is introduced, along with several approximation structures that can be distributed across a network. In this formulation, clustered region graphs are introduced to create a network "backbone" across which the computation for inference is distributed. This thesis shows that clustered region graphs have good structural properties for GBP algorithms. We propose the use of network region graphs and GBP for location-aware networks. In particular, a method for representing GBP messages non-parametrically is developed. As an special case, we apply BP algorithms to mobile networks without infrastructure, and we propose heuristics to optimize degree of network cooperation. Numerical results show a five times performance increase in terms of outage probability, when compared to conventional algorithms.