MechE PhD Prospectus Defense: Xinhuan (Leo) Sang

TITLE: Gaussian Processes with Gradient Information for Aerodynamic Modeling Surrogates, Model-free Collision Avoidance through Reinforcement Learning, and Human Activity Recognition

ABSTRACT: This dissertation investigates novel applications of Gaussian Processes (GPs) with gradient information on large datasets. Unlike traditional Gaussian processes (GPs), GPs with gradient information leverage derivatives as additional input information to achieve improved predictions (especially for extrapolation beyond the existing dataset). We explore the application of this method in three projects: the dynamics and acoustic modeling of a quadrotor, reinforcement learning for real-time control based on model-free Control Lyapunov Functions (CLF) and Control Barrier Functions (CBF), and the prediction of human gait and activity using IMU sensor data. In the first project, our aim is to create dynamic and acoustic models for quadrotors that can operate in real-time on personal-level computers by distilling high-precision large datasets generated by Computational Fluid Dynamics (CFD) on supercomputers into more compact GP models. By efficiently leveraging CFD's capability to generate gradient information, GPs with gradient information produce predictions with accuracy surpassing traditional GPs. We achieved rapid, high-accuracy predictions by focusing on local datasets proximal to the prediction points. Our next step is to improve the algorithm to utilize the entire dataset for predictions. Our next step will be using approximation methods to train GPs with gradient information on the whole dataset.

COMMITTEE: ADVISOR/CHAIR Professor Roberto Tron, ME/SE; Professor Sheryl Grace, ME; Professor Lou Awad, Sargent College, Physical Therapy