Collaborative Research: RI: Gaussian Processes that Embrace Heterogeneity for Multi-Sensor Sampling Mission

Scientists, engineers, and decision makers rely on maps built from sensor data to understand the world, predict its properties, and make informed decisions. Creating accurate maps can be difficult, time consuming, and expensive because collecting data samples often requires visiting specific physical locations to take sensor measurements or gather specimens. This project studies how autonomous systems can gather data more efficiently by deciding where, when, and how to collect new measurements. The sample gathering process can be optimized using a method known as active learning. Active learning considers the best locations from which to gather new data samples, allowing high quality maps to be created with fewer overall samples. Gaussian processes (GPs) are a popular class of machine learning methods that have found widespread adoption across many different fields. One advantage of GPs is that the resulting maps come with an estimate of how uncertain the maps are. However, current GP theory and practice assumes that all sensors, agents, and data sources behave similarly (have homogeneity) which is rarely true in real-world applications. This project is focused on filling this gap by developing new GP theory and algorithms that can use sensors, agents, and data sources that differ from each other (have heterogeneity) in ways that are common in real-world tasks. In particular, this project aims to advance GPs and GP-based active learning algorithms to leverage and reason about three forms of heterogeneity: (1) heterogeneity of sensing agents, (2) heterogeneity of sensor operation, and (3) heterogeneity of the phenomena being sensed. The first thrust will investigate how autonomous systems with different movement capabilities, costs, and motion constraints can collaborate to gather informative data. The second thrust will research how sensing systems with different sampling rates, spatial footprints, and measurement accuracies can be integrated into the learning process . The third thrust will study how to learn and reason about multiple phenomena that evolve in different spatial and temporal scales, correlated with one another. By enabling GPs and GP-based active learning algorithms to leverage and reason about real-world forms of heterogeneity, this effort will extend the usefulness and efficiency of accurate exploration using GPs in existing domains and enable active learning in new application domains. More broadly, addressing heterogeneity is essential to enable the use of GP-based active learning with applications spanning autonomy, robotics and control, as well as machine learning and artificial intelligence. 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.