Prè-requis
Notebooks for the first 6 chapters are accessible at https://github.com/erachelson/RLclass_MVA. Please download the latest version before class.
Objectif du cours
This class aims at providing a comprehensive and modern introduction to reinforcement learning concepts and algorithms. It endeavors to provide a solid formal basis on foundational notions of reinforcement learning (MDP modeling, convergence properties of dynamic programming and stochastic gradient descent, stochastic bandits, etc.), in order to move in a principled manner towards state-of-the-art algorithms (including deep RL ones).
Organisation des séances
The schedule is designed around 3-hours sessions. It might be adjusted depending on the progression of classes.
Session 1 should cover chapters 0 to 2.
Session 2: chapter 3.
Session 3: chapter 4.
Session 4: chapter 5.
Session 5: chapter 6.
Session 6 is kept unassigned for now, to preserve the possibility to use it as a buffer to avoid rushing through previous sessions.
Session 7 and 8: stochastic bandits, monte carlo tree search and alphaGo.
Mode de validation
The final grade will be composed of three parts (coefficients TBD).
1. Between session 2 and session 6 (included), a short mandatory online 10-15 minutes quiz will be run at the beginning of class, on the contents of the previous session. These quizes will be graded and will count towards the final grade.
2. An implementation project around session 6 will also be graded.
3. An independent assignment on the last two sessions will finally be graded.
Thèmes abordés
The class is structured around a series of chapters, each covered in an independent notebook.
Chapter 0: Reinforcement Learning class introduction; key intuitions
Class rules, general definition of RL, position in the ML landscape, first elements of vocabulary.
Chapter 1: Modeling sequential decision problems with Markov Decision Processes
MDP definition, policies and value functions, definition of optimality, state distributions, horizon.
Chapter 2: Characterizing value functions: the Bellman equations
State-action value functions, dynamic programming evaluation and optimality Bellman equations, value iteration, (modified) policy iteration, asynchronous dynamic programming, linear programming.
Chapter 3: Learning value functions
Approximate value and policy iteration, AVI as a series of supervised learning problems, stochastic gradient descent for AVI, temporal difference methods, Q-learning, fitted Q-iteration. Overview of key intrinsic challenges in RL.
Chapter 4: Deep Q-Networks
Neural network architecture for value functions, DQN, improvements on DQN.
Chapter 5: Continuous actions in DQN algorithms
From DDPG to SAC.
Chapter 6: Direct policy search and policy gradient methods
Policy gradient theorem, REINFORCE, A2C, PPO, evolutionary RL.
additional chapters on bandits, exploration, MCTS and alphaGo
TBC
Emilie KAUFMANN
INRIA
Emmanuel RACHELSON
ISAE SUPAERO