RL (Reinforcement Learning) — Glossary

A training method where an AI learns by taking actions and receiving feedback — rewards for good choices, penalties for bad ones — until it figures out how to achieve a goal.

Reinforcement learning trains an AI through trial and error: the system takes actions, receives rewards or penalties, and gradually learns which choices lead to better outcomes — discovering strategy from feedback rather than from labeled examples.

AlphaGo learned to beat Go world champions largely by playing against itself; the same approach trains robots to walk and trading systems to optimize returns. In language models, RLHF (reinforcement learning from human feedback) rewards the responses humans rank higher, aligning chatbots toward helpful, harmless, and honest behavior.

The trade-offs: RL is slow and data-hungry, since the model needs many attempts to learn, and it risks reward hacking — maximizing the score without actually solving the problem. It pays off where "good" and "bad" outcomes can be clearly defined.