Constrained Bayesian optimization and spatio-temporal surveillance for sensor network design in the presence of measurement errors

Abstract

The optimal placement of sensors is studied to construct a surveillance sensor network for a complicated stochastic system with random measurement errors. The problem is formulated as a joint problem of constrained black-box optimization for the fast detection of an anomaly event and spatio-temporal change-point detection for a low false alarm rate. An algorithm is proposed called Confidence-Set based Constrained Bayesian Optimization (CSCBO) that models performance measures as Gaussian Processes (GPs) and provides a flexible and easy-to-implement framework for handling noisy black-box constraints. As the decision variables of this problem are high-dimensional binary variables, the Wasserstein similarity metric is introduced as a distance measure among different solutions to capture the similarity among solutions properly. Finally, a newly proposed detection statistic for spatio-temporal surveillance is combined with CSCBO to identify the optimal sensor placement while controlling the false alarm rate. The combined procedure is applied to the Altamaha River.