MDPs, value and policy methods, Q-learning and policy gradients — the machinery RLHF is built from.
10h · 8 lessons · 1 challenge
Before this: Calculus & Optimization, Probability Foundations
Reinforcement learning matters twice over: as its own field, and as the foundation of the alignment methods in the LLM track. PPO on a language model makes very little sense until policy gradients make sense on a gridworld.
The module stays hands-on — you implement tabular Q-learning, then DQN, then REINFORCE with a baseline — and ends pointed directly at RLHF.
Your exploration policy against epsilon-greedy, UCB and Thompson sampling over a thousand pulls. Regret is the score.
Worth 500 XP.
