Active learning algorithms for autonomous platforms to characterize underwater static acoustic sources

Abstract

Integrating passive acoustics with autonomous platforms presents an opportunity to complement the spatial constraints of traditional fixed sensor methods by leveraging the capacity of autonomous underwater vehicles (AUVs) to make real-time decisions. We present algorithms that adaptively sample a survey region based on the sound field characteristics. These algorithms use an active learning strategy based on Gaussian Process (GP) regression to characterize a static sound field in a survey region. With each location sampled, the algorithms employ a GP to estimate the distribution and quantify the uncertainty of static acoustic sources within the region. The uncertainty metric is used to then choose the next sampling location. This dynamic approach not only maximizes the information gained by the AUV at every location that it samples but also ensures an efficient convergence toward the true distribution of underwater static sources in that region. These algorithms were developed in simulation and will be tested in controlled experiments.