Comparison of Probabilistic Chain Graphical Model-Based and Gaussian Process-Based Observation Selections for Wireless Sensor Scheduling

The constrained power source given by batteries has become one of the biggest hurdles for wireless sensor networks to prevail. A common technique to reduce energy consumption is to put sensors to sleep between duties. It leads to a tradeoff between making a fewer number of observations for saving energy while obtaining sufficient and more valuable sensing information. In this paper, we employ two model-based approaches for tackling the sensor scheduling problem. The first approach is to apply our corrected VoIDP algorithm on a chain graphical model for selecting a subset of observations that minimizes the overall uncertainty. The second approach is to find a selection of observations based on Gaussian process model that maximizes the entropy and the mutual information criteria, respectively. We compare their performances in terms of predictive accuracies for the unobserved time points based on their selections of observations. Experimental results show that the Gaussian process model-based method achieves higher predictive accuracy if sensor data are accurate. However, when observations have errors, its performance degrades quickly. In contrast, the graphical model-based approach is more robust and error tolerant.