Information-Driven Distributed Sensing for Efficient Bayesian Inference in Internet of Things Systems

Abstract

Distributed Bayesian inference or estimation in Internet of Things (IoT) has recently received much attention due to its broad application in the areas of object classification, target tracking and medical diagnosis etc. In many distributed IoT systems with limited resources, e.g. sensor networks and crowdsourcing systems, it is likely that only a few agents will have valuable information at any given time. Therefore, the paradigm of information-driven distributed sensing (IDDS) is essential to achieve efficient inference, where the resources are spent only on sensing and communicating valuable information. In this paper, we consider the problem of IDDS for efficient Bayesian inference with exponential family distributions. We first propose a centralized algorithm (C-IDDS) where a centralized controller exists to make sensing decisions for the sensing agents. As the centralized algorithm does not scale well in large systems, we continue to design a distributed algorithm (D-IDDS) where each individual sensing agents can make their own sensing decisions independently. Both C-IDDS and D-IDDS are online algorithms which can adapt to stochastic system conditions without any future information. Through rigorous theoretical analysis, we prove that the proposed algorithms can achieve an asymptotically optimal system-wide utility. A real testbed has been built to evaluate the performance of the proposed algorithms in real-world environments. Using the data from the real-world testbed and comparing with some baseline methods, we demonstrate the effectiveness of the proposed C-IDDS and D-IDDS algorithms.