Game Synthesis & Control
Teachers: Anca Muscholl and Guillaume Lagarde
This course is compulsory for VL and optional for AM. It is worth 6ECTS.
Part 1 (Anca Muscholl):
The first part of the course is an introduction to game theory for verification and synthesis. The synthesis of open systems or controllers is based on the principle of a reactive system, which must interact with its environment. The two entities - system and environment - are seen as 2 antagonistic players. Different types of games will be discussed: two-player games on finite arenas, games for controller synthesis, and distributed games.
References:
Lecture notes
Exercises
Research articles
Part 2 (Guillaume Lagarde):
In this second part, we will explore the theoretical foundations of reinforcement learning, an extremely powerful artificial intelligence framework that enables machines to acquire knowledge and make decisions by interacting with their environment. We will explore in detail fundamental concepts such as the one-armed bandit problem, Markov decision processes, trade-offs between exploration and exploitation, Monte-Carlo and Q-learning, etc., as well as advanced techniques such as function approximation and Deep Q-learning. By the end of this course, you'll have acquired the skills needed to understand AlphaGo, the first AI to outperform humans at the game of go.
References:
- Reinforcement Learning: An Introduction by Barto and Sutton
- Foundation of Machine Learning by Mehryar Mohri, Afshin Rostamizadeh, and Ameet Talwalkar
- DeepMind Reinforcement Learning Lecture Series 2021
Exercices
- TD1: Implementation of value iteration, monte carlo, sarsa, q_learning, learning with linear function approximation.
- TD2: Implementation Monte Carlo Tree Search
Research articles
-
Coquelin, P. A., & Munos, R. Bandit algorithms for tree search. arXiv preprint, 2007.
-
Hasselt, H. Double Q-learning. Advances in neural information processing systems, 2010.
These last two papers are a little bit more mathematical: