Objectif du cours

Objective of the course:
Be it on Atari Games, Go, Chess, Starcraft II or Dota, Deep Reinforcement Learning (DRL) has
opened up Reinforcement Learning to a variety of large scale applications. While it could
formally appear as a straightforward extension of reinforcement learning to deep learning based
function approximations, DRL often involves more than simply plugging the newest deep
learning architecture into the best theoretical reinforcement learning method. In this course, we
will journey through the recent history of DRL, from the now seminal Neural fitted-Q, to the
most popular Deep Q-Network (DQN) and Asynchronous Actor Critic, to the latest MuZero. At
each milestone, we will emphasize both the theoretical and practical considerations underlying
the improvements in each algorithm. This will range from addressing overestimation biases, to
fixing off-policyness in multi-step return estimates, to properly regularizing policy optimization
steps. Alongside the lectures, the practical session will revolve around implementing and testing
DRL algorithms in JAX and Haiku on simple environments. At the end of the course, students
should have a good understanding of the broad principles underlying the conception of DRL
algorithms, as well as hands-on experience on how to implement them.