A program that taught itself Go, a robot that learnt to walk, a language model "trained with human feedback": each learnt from a single number that says better or worse. That number is a reward, and reinforcement learning is the art of turning rewards into behaviour. Run on the page: rewards and returns; bandits and the price of exploring; states, values and the Bellman equations; dynamic programming, Monte Carlo and temporal difference; traces, exploration and planning; features, neural networks, DQN, policy gradients, actor–critic and PPO; models, partial observability, games, hierarchy, offline learning and learning from human feedback; reward design, evaluation, safety and the brain — every agent trained in front of you, every curve drawn from several seeds, in English and Chinese.
Every return, value, curve and number on these pages is computed by the page or cited with its source; your progress stays in this browser. Every agent trains in your browser on small worlds made for the course; real training at scale needs hardware this course does not assume. The course names no product or company outside its history and applications. Not affiliated with or endorsed by any organisation, company, laboratory, publisher or author mentioned here.