Deep Reinforcement Learning

General Information

Suggested prerequisites: Mathematics I to IV, Neurocomputing, basic knowledge in Python.

Exam: written examination (90 minutes), 5 ECTS.

Contact: julien dot vitay at informatik dot tu-chemnitz dot de.

Language: English. The exam can of course be done in German.

Registration and materials: https://bildungsportal.sachsen.de/opal/auth/RepositoryEntry/21637267457

Content

The course will dive into the field of deep reinforcement learning. It starts with the basics of reinforcement learning (Sutton and Barto, 2017) before explaining modern model-free (DQN, DDPG, PPO) and model-based (I2A, World models, AlphaGo) architectures making use of deep neural networks for function approximation. More "exotic" forms of RL are then presented (successor representations, hierarchical RL, inverse RL, etc).

The different algorithms presented during the lectures will be studied in more details during the exercises, through implementations in Python.

The plan of the course is:

1. Reinforcement Learning (MDP, dynamic programming, Monte-Carlo methods, temporal difference)
2. Value-based deep RL (DQN)
3. Policy gradient methods (A3C, DDPG, TRPO, PPO)
4. Model-based RL (Dyna Q, AlphaGo, I2A)
5. Successor representations
6. Hierarchical RL, Inverse RL, Multi-agent RL

Literature

• "Reinforcement Learning : An Introduction", Sutton & Barto