Beschreibung

Many of today’s AI systems, from robotic agents to LLMs, learn by trial and error: they take actions, see outcomes, and improve. That’s reinforcement learning (RL). This course gives you the foundations of Deep RL: sequential decision-making in complex settings. For research, industry, or understanding how modern AI is trained, you’ll get the concepts and tools to build and reason about RL agents.

Learning Goals:

  • Understand and apply core RL paradigms
  • Implement Deep RL algorithms from scratch
  • Design and formalize RL problems
  • Train and evaluate RL agents
  • Apply RL to specialized domains

What you get

This course gives you knowledge and hands-on experience with modern deep reinforcement learning methods. You will move from basic RL (REINFORCE, Q-learning) in the first weeks to advanced topics such as Deep Q-Networks, policy gradient methods (DDPG, PPO), model-based RL (Dreamer) and RL from human feedback (RLHF). You will apply what you learn in a team project and present your results at the end of the semester.

This course combines lectures, guided practical work with TorchRL (PyTorch’s RL library), and a team project. Parts of the content are inspired by standard RL/DRL curricula and resources.

How it works

The course runs in weekly sessions (see the schedule for exact dates, from April to June 2026). The format is hybrid: you can attend online or in person in Kiel.

During the week you are expected to go through the assigned material and practical content (roughly 5–10 hours per week). In the weekly session we will discuss the theory, go through exercises, and answer questions. There will be homework submissions during the semester; these are required to pass the course. In the second part of the semester you will work in a team on a deep reinforcement learning project and present it at the end of the course. All required software and course materials are available for free. For the practical parts you will need a working Python environment with PyTorch and TorchRL; no paid services are required.

Before the first session there will be a course kick-off: you can meet the instructor and ask any questions about the course and how to participate. Attendance at the kick-off is not mandatory but recommended. You can register here: [registration link].

What you should bring

You have basic programming experience in Python and some familiarity with machine learning or optimization (e.g. gradient descent, neural networks). Core concepts such as MDPs and value functions will be introduced from the ground up; experience with PyTorch is helpful. You should be motivated to engage with the material and contribute to your team project—this course is a good opportunity to learn and discuss state-of-the-art deep reinforcement learning methods.

Workload

You should plan to spend about 10 hours per week for the course including our weekly course session. If you are really good and have solid foundations you might go down to 6-7 hours. This is the absolute minimum you should expect. Google and peers do not pay entry salaries over a 100k because Machine Learning is so particularly easy. You have to spend hard work in the course but I promise it is worth it.

The Formalities

In order to receive a certificate of attendance (Leistungszertifikat) for this course, active participation is expected, and no more than two classes may be missed. The active participation is proven via the homework submisson and the final presentation of your project by you and your team, and the delivery of a well documented project source code. The same conditions apply in order to receive ECTS.

In the online sessions it is necessary that you always provide your full name in Zoom so that your presence is registered on the EDU platform. No mere certificate of attendance will be issued for this course.

Further details may be given in the course.