Reinforcement Learning

What is reinforcement learning?

Think about training a puppy: it does the right thing, you give it a treat; it does the wrong thing, no treat. After a few rounds, it gets the idea. Reinforcement learning (RL) runs on that same logic — an AI tries things in an environment, gets rewarded for good moves and penalized for bad ones, and gradually learns an optimal strategy.

How is it different from supervised learning?

Supervised learning: learn from the answer key
Someone labels the correct answer for every example, and the model imitates it.
Reinforcement learning: learn by trial and error
There's no answer key. The agent explores on its own and figures out what's right from reward signals.

A few key ideas

Agent and environment
The agent takes actions in an environment, and the environment feeds back.
Reward
Points for good moves, penalties for bad ones — the only learning signal it gets.
Policy
The rule it eventually learns: "given this situation, do that."

Why it matters

RL is what let AI beat humans at Go and video games, and today it's a key ingredient in RLHF and reasoning models. Any problem with no answer key — only trial and error — needs it.

Bottom line: reinforcement learning is "treat for doing it right, nothing for doing it wrong" — the AI teaches itself the best move through trial and error.