Towards Transparent, Physically Consistent Machine Learning Models
Robert Babuska, Delft University of Technology / Czech Technical University in Prague, Netherlands
Dealing with “Dirty” Data: Solutions from Fuzzy Systems Research
Uzay Kaymak, Eindhoven University of Technology, Netherlands
Data Aware Agentic AI: Cognitive, Control, and Data Flows
Michael Berthold, University of Konstanz / KNIME AG, Germany
Towards Transparent, Physically Consistent Machine Learning Models
Robert Babuska
Delft University of Technology / Czech Technical University in Prague
Brief Bio
Robert Babuška received the M.Sc. (with honors) degree from the Czech Technical University in Prague in 1990, and the Ph.D. (cum laude) degree from Delft University of Technology, the Netherlands, in 1997. He holds a part-time appointment as Full Professor and Head of the Learning and Autonomous Control Group at TU Delft, and serves as Vice Director for Research and Head of the Machine Learning Group at the Czech Institute of Informatics, Robotics, and Cybernetics (CIIRC), CTU Prague. He was the founding director of both the TU Delft Robotics Institute and the ELLIS Unit Delft, which is part of the pan-European AI network ELLIS. His research interests include reinforcement learning, adaptive and learning control, and nonlinear system identification. He has applied these techniques across various fields, including process control, robotics, and aerospace.
Abstract
As machine learning is being rapidly adopted in a wide range of domains, the need for models that are both accurate and physically interpretable has never been more critical. This talk explores recent advances in symbolic regression, a rapidly evolving field that aims to derive concise, human-understandable models from data. The goal is to provide researchers and practitioners with actionable insights into building next-generation equation learners that combine the rigor of physics with the power of machine learning, even when training on sparse datasets. We begin by examining the evolution from traditional genetic programming (GP)-based symbolic regression to neural architectures that incorporate prior system knowledge and enforce physical plausibility. We then expand this paradigm to neuro-evolutionary algorithms that combine evolutionary search for neural network topologies with gradient-based fine-tuning of their parameters. Finally, we discuss transformer-based architectures that reduce the computational burden by pre-training a large, generic model that can be quickly queried for unseen data during the inference step. We give application examples from robotics where these advances are particularly impactful in providing accurate yet interpretable dynamic models, which are essential for reliable control, planning, and optimization.
Dealing with “Dirty” Data: Solutions from Fuzzy Systems Research
![]()
Brief Bio
Uzay Kaymak is Full Professor of Information Systems at Jheronimus Academy of Data Science (JADS), affiliated with Eindhoven University of Technology in the Netherlands. After his studies at Delft University of Technology, he held positions at Shell International Exploration and Production, and Erasmus University Rotterdam. His research focuses on fuzzy modeling, data-driven decision support and computational intelligence methods, where linguistic information, represented either as declarative linguistic rules derived from experts or obtained through natural language processing, is combined with numerical information extracted from data by computational and machine learning methods. Prof. Kaymak is an internationally acknowledged researcher, who is listed in “Stanford University/Elsevier's Top 2% Scientist Rankings”. In the past, he has been a visiting professor at Salford University, UK and Zhejiang University, China. Currently, he leads the Data Analytics Unit of JADS.
Abstract
As our capacity to collect data grows, there is an increasing tendency to believe that data are of high quality and sufficiently rich for developing good machine learning models. Even though automated data collection improves the situation, these conditions are often not met when data originate from humans or come from human-in-the-loop systems. In this plenary, we address implications of this for modeling, by using examples from safety critical domains (e.g. healthcare), and we will discuss several approaches that can be developed by using fuzzy sets in order to deal with the consequences of poor data quality.
Data Aware Agentic AI: Cognitive, Control, and Data Flows
Michael Berthold
University of Konstanz / KNIME AG
Brief Bio
Michael Berthold is co-founder of KNIME, the open analytics platform used by thousands of data people around the world. He is currently CEO of KNIME AG and, until recently, a professor at Konstanz University, where his research interests included bisociative data analysis and widening of mining algorithms. Previously he held positions in both academia (Carnegie Mellon, UC Berkeley) and industry (Intel, Tripos). Michael has co-authored two successful data analysis text books and is a Fellow of the IEEE as well as honorary professor at Obuda University. If time permits he still creates workflows.
Abstract
Current Agentic AI requires orchestration of prompts, tools, and proper access to data. In this talk, I will dive deeper into the building blocks of data aware agentic AI and illustrate why providing an AI with structured access to data has benefits in terms of quality but also with respect to documentation and auditability of the decision processes. Done properly, this allows AI to explain its reasoning.