Thesis Work: Offline Reinforcement Learning with Physics-Informed Data-Driven Models

Details

• Period: 5 months (January/February - June/July)
• Number of credits: 30 ECTS
• Number of students for this thesis work: 1
• Location: ABB Research Center (Vasteras)

Your role and responsibilities

Advanced control solutions like Reinforcement Learning (RL) often rely on simulators that may not fully capture the real-world process due to noise, disturbances, or modeling limitations.

This thesis explores model-based offline RL, where the model is built using both physics knowledge and data. The work will investigate how we can refine physics-based simulators with data or embed physics knowledge using techniques from the area of Physics-Informed Machine Learning.

Goals:

• Review state-of-the-art model-based Reinforcement Learning approaches
• Investigate techniques in system identification, such as Physics-Informed Neural Networks and their applicability in real-world scenarios
• Develop and validate hybrid models using simulations or lab experiments

• Master's student in Computer Science, Industrial Engineering, or a related field
• Background in Machine Learning, Control and Systems Engineering, or similar disciplines
• Motivation to solve real-world problems using state-of-the-art methods
• Good programming skills (Python)
• Self-driven and solution-oriented

Supervisor Iga Pawlak, iga.pawlak@se.abb.com, will answer all your questions about the thesis topic and expectations. Recruiting Manager Linus Thrybom, +46 730 80 99 06, will answer your questions regarding hiring.

Positions are filled continuously. Please apply with your CV, academic transcripts, and a cover letter in English. We look forward to receiving your application!

Join us. Be part of the team where progress happens, industries transform, and your work shapes the world. Run What Runs the World.