Reinforcement Learning

In supervised learning somebody hands you the right answer for every input. In reinforcement learning nobody does. An agent takes an action, the environment responds with a new state and a reward, and over many episodes the agent works out which sequences of actions pay off.

The hard part is credit assignment: a reward arriving now might be the result of a decision made two hundred steps ago. Techniques like temporal-difference learning and policy gradients exist to spread that credit backwards sensibly.

It is the natural fit for control problems — robotics, game playing, resource scheduling — and it underpins the fine-tuning stage of modern chat models.