Instructor: Daniel Hennes
Secretary: Carola Stahl
Sessions: Tuesday, 15:45 – 17:15 (0.447)
Office hours: by appointment
Communication: Announcements, such as change of schedule, will be communicated here and via the mailing list.
Slides and references to additional material will be posted as the course proceeds.
- April 10: Introduction and organizational matters
- June 12:
- June 19:
- June 26:
- Playing FPS Games with Deep Reinforcement Learning (Marc Tuscher)
- Libratus: The Superhuman AI for No-Limit Poker (Pascal Tilli)
- July 03:
- Curiosity-Driven Exploration using Auxiliary Tasks (David Holzmüller)
- Synthesizing Programs for Images using Reinforced Adversarial Learning (Marco Radic)
July 17: (Backup slot) Show & tell
- July 20: Show & tell + questions regarding seminar papers in room 0.124!
- July 25: extended deadline for seminar papers!
In this seminar, we will discuss how reinforcement learning can be combined with deep learning. Reinforcement learning is a general purpose framework for artificial intelligence. The key is learning optimal behavior through interaction with the environment: a reinforcement learning agent improves over time through a process of trial & error. Scaling reinforcement learning requires powerful representations as many complex real-world domains feature high-dimensional state and observation spaces as well as continuous action spaces. Deep learning is the state-of-the-art for many machine learning tasks such as image classification, speech recognition, and language translation. Deep learning provides powerful function approximation and representation learning. Deep neural networks learn compact low-dimensional representation (features) from data.
We will discuss current trends and methods in deep reinforcement learning, for instance:
(Crossed-out topics are already taken!)
- Atari (see DQN)
- Go: AlphaGo / AlphaGo Zero
- Chess: AlphaZero
The first lecture(s) will provide a recap of the fundamentals and an overview of recent topics. In the subsequent weeks, students will present papers in the field of deep reinforcement learning.
The seminar also includes a final project that will be based on a recent publication and demonstrate the approach/technique in simulation (e.g. using OpenAI Gym). A short report and final presentation are to be delivered at the end of the semester.
We assume familiarity with reinforcement learning and machine learning, in particular:
- Reinforcement Learning
- Definition of MDPs
- Policy and value iteration
- Q-learning / SARSA
- Machine Learning
- Classification and regression
- Fitting of linear and non-linear models
- Loss functions
- Stochastic gradient descent
- Training/test error, overfitting