After successful completion of the course, students are able to not only understand, to explain and to apply the theory and the methods of reinforcement learning including the latest developments, but also to implement the most important algorithms.
[This year (summer semester 2024) with updated, extended lecture notes!]
Reinforcement learning is a field of artificial intelligence / machine learning and is concerned with the development of strategies that an agent uses to maximize its reward in a random environment in a model-free manner.
Applications include robotics (OpenAI gym), computer vision, games (such as Go, chess, Atari 2600, or Dota 2) at the human level or above and many more. Furthermore, RL is the final and instrumental training step in large language models (LLM) such as ChatGPT and Gemini.
Theory and algorithms of reinforcement learning:
- Introduction
- Bandit problems
- Markov decision problems
- Bellman equations
- Hamilton-Jacobi-Bellman equation
- Dynamic programming
- Monte-Carlo learning
- Temporal-difference learning
- Tabular methods
- Function approximation and deep learning
- On-policy vs. off-policy
- Eligibility traces
- Policy gradients and actor-critic
- RL with human feedback: InstructGPT and ChatGPT
- Applications
In the tutorial, the theory will be repeated and extended and the algorithms will be implemented.
The theoretical aspects will be explained in the lectures in a self-contained manner. The usual knowledge of linear algebra, calculus, and probability theory is required.
