Study Reinforcement Learning
activeLearn reinforcement learning by running it in your browser. 34 lectures go from a gridworld to how language models are trained, and nine exercises have tests that pass only when your code is right.
Reinforcement LearningRLHFInteractive Learning
Study Reinforcement Learning
rl.yad.codes is a site for learning reinforcement learning by playing with it. Its home page opens with a Q-learning agent learning a small grid in your browser: it starts out knowing nothing, falls into the pits, and by fifty episodes heads straight for the goal.
It’s built from a study repo that started on 31 December 2016 as a list of talks, books and courses.
What’s there
- 34 lectures in eight tracks, from MDPs and policy gradients to RLHF, DPO and GRPO. They’re drafts, each marked unreviewed until a person has checked it.
- A figure to try on every lecture with an exercise. Five of the nine have a setting to change, such as γ on value iteration, β on DPO or the group size on GRPO.
- Nine exercises, from value iteration on a gridworld to GRPO, with tests that pass only when your code is right.
- Hand-written notes from Berkeley’s CS 294 in 2017, and the original list of talks, books and courses.
Links
- rl.yad.codes
- Why I Keep Coming Back to Reinforcement Learning, the essay with the slot machines
- The study repo on GitHub