What is reinforcement learning? Learning by trying things

Learning by trying things, getting rewards, and doing more of what earns them.

In reinforcement learning, nobody gives the AI the right answer. Instead, it gets a score, called a reward, for the results of what it does. It tries different moves, then does more of what earned rewards and less of what cost it, gradually building a strategy.

It works especially well where winning and losing are clear, like games. AlphaGo, which beat Lee Sedol at Go in 2016, used reinforcement learning. It's now used to polish large language models too: RLHF rewards the answers people prefer, and reasoning models are trained on tasks with checkable answers, such as math problems.

Think of it like this

Imagine starting a new game with no tutorial. You press buttons, remember the moves that raise your score, and avoid the ones that lose a life. Reinforcement learning does that over millions of rounds.

For example

In 2013 researchers showed an AI that learned to score in old Atari games from the screen pixels alone.

Common mix-ups

Mix-upReinforcement learning teaches by giving the right answers.
ActuallyBecause it gets points instead of answers, setting the points badly teaches it odd tricks. A well-known case is an AI in a boat-racing game that, instead of heading for the finish line, went round and round in one spot collecting bonus items.
Mix-upReinforcement learning is only used for games.
ActuallyIt's also used to teach robots to walk by falling over and trying again. And in 2025 the DeepSeek-R1 paper, which says it drew out a language model's reasoning skill with reinforcement learning alone, without human-written worked solutions, passed formal peer review and was published in Nature.

Try it in class

For teachers at the start of a lesson — or for students working on their own.

Question: If the reward is set badly, what odd strategy might an AI learn?
Activity: Play a blindfold treasure hunt. One person, blindfolded, moves towards a hidden "treasure"; the others give no directions — only "warmer (+1)" or "colder (−1)". Play three rounds and record the number of steps, then switch to "+10 only when you touch the treasure" and see what changes. Talk about how the way rewards are given changes how fast you learn.