Reference · Glossary

Reinforcement learning

Last updated

Training by **trial and reward** — an agent takes actions, gets feedback, and learns policies that maximize long-term score.

When to use

Game-playing bots, robotics control, ad bidding, and alignment techniques like RLHF that rank preferred answers.

When not to

Simple prediction from labeled data — that's usually supervised learning, not RL.

Example

After initial training, humans rank two model replies; RLHF nudges the model toward answers people preferred over time.