CAREER: New Foundations for Multi-Fidelity Prediction, Estimation, and Learning Under Uncertainty in Dynamical Systems

Regents Of The University Of Michigan - Ann Arbor, Ann Arbor MI

Investigators

Abstract

This Faculty Early Career Development (CAREER) grant will fund research that enables autonomous systems to estimate the effects of prediction uncertainty on planning and control decisions, with application to autonomous flight of soaring aircraft and in urban environments, thereby promoting the progress of science and advancing the national prosperity. Autonomous soaring aircraft offer new fuel-efficient approaches for travel, reconnaissance, and observation, including in hard-to-reach areas of the atmosphere. Their built-in simulators make assumptions about the presence of atmospheric boundary layers and updrafts to optimally extract energy from the prevailing winds. Computational models are also integrated in the planning and control architecture of unmanned aerial vehicles as they predict optimal paths through urban infrastructure based on estimates of the surrounding flow fields. Without assessing the uncertainty in their predictions, such simulators may result in suboptimal or catastrophic decisions, as opportunities for optimal performance are missed or safety constraints are violated. This project addresses this challenge by developing new, fast, and automated algorithms for rigorously quantifying uncertainty and updating computational models accordingly, and by validating these algorithms using experimental aircraft in controlled but complex wind conditions. The research is integrated with educational efforts aiming to bring a computational perspective on modeling, data science, and statistics to engineering students and the public through a series of workshops built around relevant case studies, a new data science class for an aerospace engineering curriculum, and a partnership with the Ann Arbor Hands-On Museum to develop exhibits accessible to K-8 students and their parents. This research aims to develop the foundations of automated approaches for deriving problem-specific multi-fidelity uncertainty quantification techniques. Such techniques aim to fuse information from simulation and data sources of varying fidelity and cost to achieve accurate predictions at a significantly lower computational cost than that required by the highest fidelity model. Current realizations are based on heuristics that are not adapted to the specific relationships between data sources of a given problem. To overcome this limitation, the research will derive new statistical estimators through analysis of Bayesian posteriors and maximum entropy distributions endowed with only the information available rather than heuristics; use these estimators to develop new filtering, state estimation, and Bayesian inference techniques; and demonstrate how these techniques may be applied to challenging nonlinear, chaotic, and non-Gaussian dynamical systems arising in the context of planning and control of soaring aircraft in complex wind fields. 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.