CAREER: Leveraging Signal Structure for Cost-Sensitive Adaptive Sampling

点击翻译按钮获取中文摘要

A persistent challenge to scientists and engineers is the ability to rapidly sense the environment. Whether monitoring for harmful air pollutants, wildfires, or low oxygen levels in lakes, an essential task is to determine all locations where a factor of interest reaches a critical level. To solve this problem, practitioners are increasingly utilizing mobile sensors such as those deployed on unmanned vehicles. While these provide a safe means of exploring large spatial regions and hazardous environments, they also require extensive planning to ensure the most relevant measurements are collected and to manage battery life. This project will overcome these issues through the design of adaptive sampling algorithms, which automatically guide sampling vehicles to the most important regions while accounting for the realistic costs associated with the measurement process. The resulting approaches will provide general-purpose solutions to environmental sampling that can be easily utilized by practitioners while also accounting for the realistic challenges that typically prevent adaptive methods from translating into practice. Furthermore, this research will support education and diversity through the development of curriculum for a high school course on unmanned aerial vehicles as well as a citizen science campaign that leverages adaptive sampling to benefit one of the nation's largest urban parks.The objective of this project is to design and analyze cost-sensitive adaptive sampling algorithms for the problem of level set estimation. Novel adaptive sampling techniques will be developed under three forms of level set structure: (1) boundary smoothness, where no domain or side knowledge is available, (2) known similarity structure that indicates which locations should have similar measurement values, and (3) unknown cluster structure, where the signal of interest is constant within each cluster of locations, and the goal is to simultaneously learn the cluster structure and measurement values. For each case, principled algorithms will be derived based on recent developments from the fields of active learning, multi-armed bandits, and reinforcement learning, with the goal of providing improved empirical performance, rigorous theoretical characterization, and incorporating realistic costs such as the distance traveled while sensing. Algorithms will be evaluated on real-world datasets including those measuring air quality and geothermal energy prospects.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.