Seminar: Offline-to-Online Data-Driven Sequential Decision Making Under Uncertainty (OORDSU) - A Computational Framework for Scientific Discovery and Control

The OORDSU framework provides a unified operational calculus for bridging high-fidelity offline scientific computation with reliable real-time embedded decision making under uncertainty. It integrates rigorous mathematical programming, stochastic optimization, parametric control, and hybrid physics-AI methods to enable trustworthy, uncertainty-aware digital twins and autonomous scientific systems.
The talk first outlines the core structure of OORDSU and then examines four active research themes, each connecting past and present results to near-future aspirations:
First, I will consider inverse problems and experimental design, focusing on structured optimization methods such as Levenberg–Marquardt schemes and probabilistic formulations of sequential decision-making (Fully Probabilistic Design), with applications to data assimilation and scientific inference.

Second, I will present recent and ongoing work on reinforcement learning model predictive control (RLMPC), with an emphasis on incorporating geometric structure and scalable optimization methods toward vision-language-action (VLA) frameworks in robotics.
Third, I will discuss high-performance computing aspects, including asynchronous and concurrent algorithms, aimed at enabling efficient offline-to-online interoperability for large-scale scientific models.
Finally, I will address control and optimization for systems with complementarity structure, motivated by applications such as chemical phase transitions and materials science, where nonsmooth and hybrid dynamics play a central role.
These directions collectively illustrate how OORDSU delivers verifiable, physics-grounded computation with rigorous theoretical guarantees, directly supporting CaSToRC’s goals in optimization, digital twins, and hybrid scientific AI.

Vyacheslav Kungurtsev received his B.S. in Mathematics from Duke University in 2007 and Ph.D. in Mathematics with a specialization in computational science from the University of California, San Diego, La Jolla, CA, USA, in 2013, under the supervision of Philip Gill, working on Sequential Quadratic Programming methods for Nonlinear Programming. Subsequently, he was a Postdoctoral Fellow at KU Leuven as part of the “optimization for engineering” project and then a Postdoctoral Researcher at Czech Technical University in Prague focused on optimization algorithms for training large machine learning models on parallel, distributed and decentralized computing hardware. In 2025 he became a tenured Researcher with Czech Technical University.

His research interests include mathematical programming and numerical optimization, with applications in machine learning and artificial intelligence, real-time process control engineering, distributed computing, and uncertainty quantification for mathematical physics.