Bridging control theory and learning systems: advances in adaptation, robustness, and automation

Syed, Usman Ahmed

Permalink

https://hdl.handle.net/2142/129385 Copy

Description

Title

Bridging control theory and learning systems: advances in adaptation, robustness, and automation

Author(s)

Syed, Usman Ahmed

Issue Date

2025-04-24

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

Hu, Bin

Doctoral Committee Chair(s)

Hu, Bin

Committee Member(s)

• Srikant, Rayadurgam
• Raginsky, Maxim
• Dullerud, Geir

Department of Study

Electrical & Computer Eng

Discipline

Electrical & Computer Engr

Degree Granting Institution

University of Illinois Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Online Optimization
• Adaptive Control
• Robustness Analysis of NNs, LLMs
• Automated Control design

Abstract

The integration of control theory and learning based systems has become increasingly vital for developing intelligent and autonomous systems that can adapt, ensure robustness, and minimize human intervention. This dissertation explores three key aspects of this integration: Adaptation, Robustness, and Automation, providing novel theoretical insights and practical advancements in each area. In the first part of the dissertation, we investigate adaptive control in the presence of unknown disturbances, drawing connections between Online Nonstochastic Control and Retrospective Cost Adaptive Control (RCAC). While Online Nonstochastic Control ensures provable near-optimal regret bounds given a stabilizing policy, RCAC is capable of stabilizing unknown unstable plants through the use of a target model. We propose a unified framework that combines the strengths of both approaches, providing a foundation for more powerful adaptive control algorithms. The second part of this dissertation addresses robustness in deep neural networks by focusing on the problem of computing tight Lipschitz bounds, which are crucial for analyzing stability, generalization, and adversarial robustness. Given that computing the exact Lipschitz constant is NP-hard, existing approaches either suffer from scalability issues due to semidefinite programming (SDP) formulations or provide overly conservative estimates. Building upon ECLipsE-Fast, a state-of-the-art scalable method, we introduce a new family of Lipschitz bounds that significantly reduces conservatism while maintaining computational efficiency. Our approach generalizes the feasible points of LipSDP at each recursive step, strictly encompassing ECLipsE-Fast as a special case, and demonstrates improved scalability and precision in empirical evaluations. In the final part of the dissertation, we explore LLM-powered automation in control design, aiming to minimize human oversight while maintaining stability and robustness guarantees. We develop a framework that automates controller synthesis for both linear and nonlinear systems, ensuring its applicability across diverse control strategies. To validate real-world feasibility, we implement a fully automated control design pipeline, employing Simulation-in-the-Loop and Hardware-in-the-Loop methodologies. Our results highlight the potential of LLM-driven automation in streamlining control design, reducing manual effort, and ensuring reliable controller deployment. Together, these contributions bridge the gap between control theory and learning systems, advancing the fields of adaptive control, neural network robustness, and automated control design. The findings presented in this dissertation pave the way for more intelligent, robust, and autonomous systems, with applications spanning robotics, autonomous vehicles, aerospace, and intelligent infrastructure.

Graduation Semester

2025-05

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/129385

Copyright and License Information

Copyright 2025 Usman Syed

Owning Collections