Reinforcement learning
Also called Reinforcement Learning
Learning by trial and error, guided by rewards for good actions and penalties for bad ones.
Think of it like
Like training a dog with treats and “no” — behavior that earns reward gets repeated, behavior that doesn’t fades.
Example
An agent learns to play a game by trying moves and getting points, gradually favoring the actions that win.
How it actually works
Instead of labeled answers, RL gives an agent rewards for the outcomes of its actions, and it learns a policy that maximizes cumulative reward over time. The hard parts are delayed rewards (which action deserves credit?) and exploration versus exploitation. In language models, RL powers alignment: RLHF optimizes the model against a reward model trained on human preferences.
For product teams
The paradigm behind alignment tuning and agents — useful when success is a signal, not a labeled answer.
For engineers
Learn a policy maximizing cumulative reward via trial and error; credit assignment and exploration are core challenges.
Related
- RLHF — The basis of RLHF.
- Supervised learning — Contrast with supervised learning.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome