Renforcement
Reinforcement learning
Description: The course presents the theoretical foundations of reinforcement learning as well as the principles of the most common algorithms. Through practical work, these elements will be extended to more complex situations, making it possible to introduce the most recent algorithms that have, for example, enabled computers to master the game of Go.
Content: Reinforcement learning is introduced using the formal framework of the Markov Decision Processes. After having shown the existence and uniqueness of a solution in the form of the value function, we will discuss the classical algorithms used to calculate this function. We will then see how approximate methods (linear approximation, monte carlo estimation, bandits, deep learning) can be used to tackle more complex contexts.
Learning outcomes: Understand the theoretical foundations of reinforcement learning. Implement these methods in a way that is adapted to the problems to be solved. Sharpen your critical thinking skills.
Teaching methods: Taking into account the context (group size), lectures will be as interactive as possible andd will aim to present the theoretical and algorithmic concepts underlying reinforcement learning. The purpose of the practical work is to really confront the methods by implementing and testing the algorithms to better understand how they work and their limitations.
Means: Courses and practical work are provided by Alain DUTECH, Hervé FREZZA-BUET and Jérémy FIX. The practical work will be based on the Python language and its scientific libraries.
Evaluation methods: The module will be evaluated by a written exam, where the idea is to test the student’s ability to use methods in a clever way, to analyze the results of an algorithm, etc.
Evaluated skills:
- Be operational, responsible, and innovative in the digital world
Course supervisor: Hervé Frezza-Buet
Geode ID: 3MD4120
