Learning in dynamical systems with guarantees: from system identification to safety verification and fast adaptation

Musavi, Negin

Permalink

https://hdl.handle.net/2142/129869 Copy

Description

Title

Learning in dynamical systems with guarantees: from system identification to safety verification and fast adaptation

Author(s)

Musavi, Negin

Issue Date

2025-07-16

Director of Research (if dissertation) or Advisor (if thesis)

Dullerud, Geir

Doctoral Committee Chair(s)

Dullerud, Geir

Committee Member(s)

• Li, Yingying
• Mitra, Sayan
• Srikant, Rayadurgam
• West, Matthew

Department of Study

Mechanical Sci & Engineering

Discipline

Mechanical Engineering

Degree Granting Institution

University of Illinois Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

system identification, safety verification, meta learning

Abstract

This thesis investigates the sample complexity of data-driven algorithms for learning in controlled dynamical systems. As more systems operate in environments where models are unknown but data are available, understanding how efficiently learning algorithms use data becomes increasingly important. The thesis addresses three central problems in this context. First, it establishes conditions under which unknown parameters in nonlinear systems can be efficiently learned via non-active exploration, showing that linearly parameterized systems with real-analytic features can be identified using least-squares and set-membership estimators. Second, it develops a framework for verifying safety and synthesizing parameters in unknown systems with hybrid state spaces, leveraging a new bandit-based method—Hybrid Hierarchical Optimistic Optimization (HyHOO)—that extends prior work in black-box optimization. Finally, the thesis explores meta-learning for optimal control, proposing methods to exploit shared structure across related control tasks for fast adaptation in new control tasks. The results contribute theoretical guarantees, practical algorithms, and experimental validations toward a deeper understanding of data efficiency in learning dynamical systems.

Graduation Semester

2025-08

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/129869

Copyright and License Information

Copyright 2025 Negin Musavi

Owning Collections