Reinforcement Learning
An agent learns by trial and error, receiving rewards or penalties for its actions.
- No labeled data; signal comes from environment rewards.
- Agent observes state, picks action, gets reward, repeats.
- Powers AlphaGo, robotics, and the RLHF stage of modern LLMs.
When to use: When you can simulate or interact with an environment and define a reward, but cannot label correct actions in advance.
Example: A trading bot that buys/sells/holds each tick and gets reward equal to net profit learns a policy over millions of simulated days without labeled trades.