Ein Angebot des JRZ ISIA
Die Reading Group hat den Charakter eines lockeren Fachseminars mit Vorträgen einer Länge zwischen 30 und 45 Minuten und anschließender Diskussion. In freundlicher Atmosphäre und unter weitestgehender Themenfreiheit werden etwa Forschungsergebnisse, wissenschaftliche oder technologische Überblicksvorträge oder die Aufbereitung einer Forschungsfrage präsentiert und diskutiert. Der Fokus liegt im Austausch, in der Diskussion. Organisatorische Fragen, Anmeldungen zu Vorträgen oder Vortragsvorschläge bitte an Stefan Huber richten.
Termine 2025
September, 10th, 2025 13.30 pm, Bernhard Heinzl | Senior Data Scientist bei SCCH - Software Competence Center Hagenberg
Physics-Informed Reinforcement Learning
Physics-informed machine learning has emerged in recent years as a new paradigm that combines data-driven methods with fundamental physical principles to address critical limitations of traditional machine learning, such as sample inefficiency, lack of generalization, or physically inconsistent predictions. By harnessing established physical laws alongside statistical modeling, physics-informed machine learning offers substantial advantages over purely data-driven approaches, especially in scenarios where system dynamics are tightly governed by physics. This synergy not only enables models to make more accurate and explainable predictions, but can also reduce the amount of data required for effective learning, broadening the applicability of machine learning to domains where training data may be limited or expensive to obtain.
The integration of physics knowledge into the learning process can be achieved through multiple mechanisms, like learning biases that embed physical laws directly into the loss functions, inductive biases that incorporate structural assumptions reflecting physical relationships, and observational biases that strategically leverage physics-consistent training data. These augmentation strategies enable models to learn more efficiently by constraining the solution space to physically plausible behaviours and providing additional guidance beyond what is available in raw data.
For reinforcement learning, physics-informed approaches demonstrate particular effectiveness in applications involving reward function augmentation, physics-based constraint implementation, and hybrid model-based planning strategies. Recent developments show significant improvements in sample efficiency and solution quality across diverse domains, including robotics, control systems, and power grid optimization.
We will give an overview of current approaches for physics-informed machine learning, explore how embedding physical knowledge into the reinforcement learning process can result in more sample-efficient and generalizable agents, and discuss several examples to illustrate how physics augmentation can benefit in classical reinforcement learning scenarios. Physics-informed reinforcement learning offers a pathway towards more reliable and trustworthy learning systems suitable for real-world deployment in physics-governed domains.
If you want to join, please send an email to Stefan Huber.