GoGE Workshop on Optimization, Decision and AI

Abstract 

 At the root of the recent wave of results in the field of data-driven control are techniques aimed at eliminating the uncertainty induced by noisy data. This talk will show how a classical result from the '60s allows the designer to  deal with errors in the measured variables with relatively little effort.

Biography

Claudio De Persis is a professor with the Engineering and Technology Institute Groningen, University of Groningen, the Netherlands, since 2011. He received the Laurea and PhD degree in electronic and systems engineering in 1996 and 2000, both from the University of Rome “La Sapienza”, Italy. Before joining the University of Groningen, he held postdoctoral and faculty positions at Washington University in St. Louis, Yale University, the University of Rome "La Sapienza" and Twente University.

Abstract 

 We revisit the most basic structural scheme in feedback control, to regulate system dynamics via time-varying state feedback. This scheme is especially pertinent in coordinated motion, in repositioning a collection of agents obeying identical dynamics to a specified terminal configuration, via a control protocol that consists of a common time-varying feedback gain matrix that is broadcast and available to all to implement in conjunction with their individual state. We follow up on some early considerations by Roger Brockett in an influential publication on the controllability of the Liouville equation, and establish that linear controllability is sufficient for strong controllability of the nonlinear dynamics. This is the first known class of non-trivial nonlinear control problems with an explicit condition that guarantees strong controllability. The problem was considered by Roger Brockett but remained unresolved until recently, as the original treatment was erroneous. The key conceptual step in our work, establishing strong controllability, is based on insights gained from the theory of optimal mass transport. We also discuss a suggestion in Brockett’s work, to construct control protocols that are continuous in the problem data and show that such protocols are impossible due to an inherent topological obstruction in steering state transition matrices between terminal specifications. We conclude with a discussion on the controllability in the space of orientation-preserving diffeomorphisms.

The presentation is on joint work with Dr. Mahmoud Abdelgalil (maabdelg@uci.edu) and is based on Collective Steering in Finite Time: Controllability on GL+(n,R), IEEE Trans. on Aut. Control, (to appear) 10.1109/TAC.2025.3574186  

Biography

Tryphon T. Georgiou (Life Fellow, IEEE) received the Diploma in mechanical and electrical engineering from the National Technical University of Athens, Athens, Greece, in 1979, and the Ph.D. degree in electrical engineering from the University of Florida, Gainesville, FL, USA, in 1983. He is currently a Distinguished Professor of mechanical and aerospace engineering with the University of California, Irvine, CA, USA, and a Professor Emeritus with the University of Minnesota, Minneapolis, MN, USA. Dr. Georgiou is a Fellow of the Society of Industrial and Applied Mathematics, International Federation of Automatic Control, and American Association for the Advancement of Science, and a Foreign Member of the Royal Swedish Academy of Engineering Sciences.

Abstract 

 Ensuring safety in autonomous systems operating under uncertainty is a central challenge for realizing reliable, large-scale autonomy. Classical model-based safety frameworks—such as Hamilton–Jacobi (HJ) reachability and Control Barrier Functions (CBFs)—provide rigorous guarantees but rely heavily on accurate system models. To overcome this limitation, recent research has sought data-driven extensions that incorporate empirical information from real-world operation. This talk presents a unified perspective contrasting two broad categories of such frameworks: indirect and direct data-driven safety approaches.

In indirect frameworks, data are used to model or bound the uncertainty in dynamics, which are then embedded into existing model-based frameworks like HJ reachability or CBFs. These approaches enable scalable learning of safety guarantees and have been successfully applied to flight envelope protection for emerging electric vertical takeoff and landing (eVTOL) vehicles. However, they remain reliant on intermediate model-learning steps and domain-specific assumptions.

In contrast, direct data-driven frameworks, exemplified by the proposed Data-Driven Hamiltonian (DDH), bypass explicit model identification and instead infer safety certificates directly from trajectory data. By approximating the Hamiltonian with observed state–velocity pairs, the DDH method constructs safe sets and safety filters through a purely data-driven formulation without explicit dynamics model and guarantees conservative (inner) approximations of the true safe set.

Through this comparison, the talk highlights a conceptual shift from learning models for safety analysis to learning safety itself, outlining how direct data-driven frameworks can generalize safety guarantees across diverse dynamical systems.

Biography

Jason Jangho Choi is an assistant professor at the UCLA Electrical and Computer Engineering Department, and the principal investigator of the Safety and Collective Intelligence (SCI) Autonomy Lab. He finished his Ph.D. study in Mechanical Engineering at UC Berkeley in 2025. He received his B.Sc. in Mechanical Engineering from Seoul National University, Korea, in 2019. His research interests center on safety assurance for learning-enabled decision-making in dynamical systems. More broadly, his research lies at the intersection of learning and control, such as nonlinear systems and optimal control theory, and reinforcement learning. He is recognized as Robotics: Science and Systems Pioneers 2024

Abstract

 Sharpness-Aware Minimization (SAM) is a widely used method that steers training toward flatter minimizers, which typically generalize better. In this work, however, we show that SAM can converge to hallucinated minimizers--points that are not minimizers of the original objective. We theoretically prove the existence of such hallucinated minimizers and establish conditions for local convergence to them. We further provide empirical evidence demonstrating that SAM can indeed converge to these points in practice. Finally, we present a simple yet effective remedy for avoiding hallucinated minimizers.

Biography

 Chanwoong Park is currently pursuing the Ph.D. degree with the Learning and Decision Systems (LDS) Laboratory, advised by Prof. Insoon Yang, in the Department of Electrical and Computer Engineering at Seoul National University, Seoul, Korea. He received the B.S. degree in Electrical and Computer Engineering from Seoul National University in 2020. His primary research area is optimization theory, with a recent focus on optimization for nonconvex problems such as deep learning.

Abstract

 This talk aims to provide an efficient implementation of encrypted linear dynamic controllers that perform recursive multiplications on a Ring-Learning With Errors (Ring-LWE) based cryptosystem. By adopting a system-theoretical approach, it is able to significantly reduce both time and space complexities, particularly the number of homomorphic operations required for recursive multiplications. Rather than encrypting the entire state matrix of a given controller, the state matrix is transformed into its rational canonical form, whose sparse and circulant structure enables that encryption and computation are required only on its nontrivial columns. Simulation results demonstrate that the proposed design enables a remarkably fast implementation of encrypted controllers.

Biography

Donghyeon Song received his B.S. degree in Electrical and Computer Engineering from Seoul National University, Seoul, Korea, in 2022, where he is currently working toward the Ph.D. degree in Electrical and Computer Engineering. His research interests include disturbance observer, secure control systems, and hybrid dynamic systems.

Abstract

 This talk examines the persistency of excitation condition for inputs in Willems’ Fundamental Lemma, which provides a sufficient condition for data to be informative in data-driven control. The conservativeness of this condition is analyzed, and a relaxed input requirement is proposed. By interpreting the maximum order of persistency of excitation from a signal generator perspective, it is shown that for some persistently exciting input satisfying certain conditions, an equivalent signal generator can be designed to reproduce it. Simulation results demonstrate the validity of the proposed relaxed condition.

Biography

 Yunjeong Yang received her B.S. degree in Electrical and Computer Engineering from Seoul National University, Seoul, Korea, in 2025, where she is currently a M.S. student in Electrical and Computer Engineering (combined M.S./Ph.D.). Her research interests include data-driven control and multi-agent systems.