Kernel-Based Markov Random Fields Learning for Wireless Sensor Networks

Distributed information inference in wireless sensor networks is of significant importance for many real-world applications in which graphical modeling of a deployed wireless sensor network is fundamental. One critical issue faced today is how to learn the graphical model parameters of a deployed sensor network as efficiently as possible, since it is usually expensive or even impossible to collect a large amount of training data in a deployed wireless sensor network given the resource constraints of tiny wireless motes. This paper attempts to address this issue. We propose a novel kernel-based approach in graphical model learning for wireless sensor networks to minimize the number of training samples of real sensor data needed. We demonstrate the proposed approach by simulations using real-world wireless sensor network data. Our results show that the proposed kernel-based learning approach can substantially reduce the number of training data needed for constructing a Markov random field model of the sensor network in comparison to the traditional learning approach without affecting the constructed model's performance in distributed information inference.