Reinforcement Learning
Supervised learning copies labelled examples. Reinforcement learning has no examples at all — only an agent, an environment, and a REWARD signal. The agent acts, sees what reward follows, and slowly learns a policy: which action is best in each situation. This is the paradigm behind game-playing AIs, robotics — and the RL in RLHF.
Run one episode: the robot wanders randomly (exploration), stumbling into the pit or the trophy. Every step updates its estimate of ‘how good is this action here?’. Run 25 more and arrows emerge — the learned policy — routing around the pit toward the trophy. Note the tension it must balance: EXPLORE new paths or EXPLOIT the best known one. The tiny −0.1 per step also matters: it teaches the agent that shorter is better.
The robot knows nothing — no examples, no labels. Only: +10 at the trophy, −10 in the pit, −0.1 per step.
Check yourself
Nobody ever tells the agent the correct action. What, exactly, is the teaching signal — and why does it arrive late?
Go deeper (free): OpenAI Spinning Up — Intro to RL ↗