Probabilistic Asynchronous Inference in Wireless Networks with Spectral Clustering

This work builds upon the concept of asynchronous, in-network processing with probabilistic graphical model (PGM)-based inference algorithms. Specifically, this work studies the problem of efficient mapping of the PGM to wireless sensor network (WSN) terminals, exchanging messages in a (probabilistic) asynchronous manner. The PGM considered stems from Gaussian belief propagation (GBP), which is versatile and powerful, able to describe many known algorithms. It is shown that node clustering methods based on spectral clustering outperform autonomous clustering and k-means, in most cases, in terms of convergence rate, for given probabilities of WSN terminals being active. Interestingly, it is also found that fast convergence rate can be achieved, which is independent of the mapping method applied. Finally, optimization of WSN communication energy consumption is also addressed with spectral clustering.