Reinforcement learning (RL)
강화 학습
A method of learning by trial and error. Instead of being told the answer, the AI tries things and gets a score when it does well.
Overview
When an AI takes an action, it gets a reward (a score) based on the result. It adjusts its behavior little by little toward getting more points, and by repeating this many times, it ends up finding a good way on its own. It suits jobs that are hard for a person to teach step by step.
AlphaGo, the Go AI that beat Lee Sedol (9 dan) in 2016, used this method to build its skill by playing itself countless times. It is also used for robots learning to walk and for polishing language models so they give better answers. If the reward is set wrongly, the AI can also learn odd tricks that only chase the score.
An everyday example
It is like giving a puppy a treat when it does well while teaching "sit." The puppy does the action that got the treat more and more often.
Why it matters
It is the method behind AI that has gone beyond people at Go and games, and behind robots that walk on their own. It comes up often, along with words like "reward" and "trial and error," in news about training robots and agents.
Also calledRL