Upcoming Event: PhD Dissertation Defense

Learning and Adaptation in Dynamical Environments via Generalized Hamiltonian Dynamics

Yi Wang, CSEM Student, Oden Institute

Abstract

This dissertation develops a differential policy optimization framework for learning to optimize and explore across acquired tasks. Motivated by finite-horizon stochastic optimal control in dynamic and online environments, it formulates goal-oriented learning and optimal experimental design as an identification problem for (latent) port-Hamiltonian dynamics. This model class provides a physically informed structural prior: it represents motion, dissipation, and energy exchange with the environment, while offering a principled phase-space description of controlled adaptation. The dissertation first formulates the inverse problem of recovering such templates in differential form. It then studies the practical identifiability and learnability of port-Hamiltonian systems across regimes of increasing difficulty: low-dimensional physical systems with full-state observations, high-dimensional non-physical systems with partial observations, and operation-induced systems whose dynamics are represented in lower-dimensional latent spaces. Across these regimes, machine learning methods are developed to recover structure-preserving and control-relevant models for joint data assimilation and inference, exploration under safety constraints and non-convex numerical optimization.

Biography

Yi Wang is a PhD candidate at the Oden Institute at The University of Texas at Austin. His research focuses on scientific machine learning, numerical optimization, stochastic optimal control, and physical simulation. He is particularly interested in combining physical modeling and machine learning to solve inverse problems and control complex dynamical systems.