Real-Time Communication Control in Decentralized Autonomous Sensor Networks

Abstract

Minimizing the amount of communication required by a sensor network is crucial to minimizing both energy and time consumption, as well to operating covertly and robustly in communication-contested environments. This paper presents a novel intermittent communication control approach applicable to sensor networks deployed to sense and model spatio-temporal processes by nonparametric models such as Gaussian processes (GPs). The approach relies on a novel and efficient approximation of the GP average generalization error (AGE), as well as on novel GP sensor control and regression methods presented in this paper. This novel AGE approximation allows each sensor to characterize the nominal prediction performance of the learned GP model in the absence of communications. As a result, individual sensors can update the GP hyperparameters based solely on local measurements and decide to communicate only if and when their estimate of the nominal prediction performance falls below an acceptable threshold.