Reinforcement learning in AI, explained simply

In plain words

Nobody gives the computer the right answers. It is dropped into a situation, such as a game, a robot's body or a conversation, and has to act.

After its actions it gets a score: a reward when things go well, a penalty when they go badly. At first its moves are mostly random. Over many thousands of tries it learns which actions lead to the most reward over time, not just right away.

This is how AlphaGo got good enough at Go to beat a professional player, and how robots learn to walk. It is also a key step in making chatbots helpful: people rate answers, and the model is rewarded for the kind people prefer. That step is called RLHF.

Everyday example

Learning a new phone game

You download a game and skip the tutorial. You tap around. Some moves make the score jump, some end the level. Nobody explains the rules, but after a dozen rounds you play far better, because you repeat what raised the score and avoid what lost it. The score is the reward, your taps are the actions, and the game is the environment.

Where the example breaks: You can guess the rules after a few rounds. A reinforcement learning system usually needs far more tries, often millions in a simulation, and it chases the exact score it is given even when that score misses what we really wanted.

Go deeper

How it actually works

Reinforcement learning (RL) is learning from consequences instead of from answers.

Agent, environment, reward
The learner is called the agent. Everything it acts on is the environment. At each step the agent sees the current state, picks an action, and the environment returns a new state and a reward: a single number.
Learn a policy
The agent's strategy, its rule for choosing an action in each state, is called a policy. The goal is the policy that earns the most total reward over the long run, called the return.
Explore versus exploit
The agent must balance trying new actions (explore) with repeating ones that already pay (exploit). A common trick is to act randomly some of the time at first, and less often as it learns.
Credit for delayed rewards
A win may come many moves after the moves that earned it. RL methods estimate how much each earlier action contributed, so good early moves get credit too.
Inside language models
In RLHF, a reward model trained on human rankings scores a chatbot's answers and RL nudges the language model toward higher scores. Newer reasoning models, such as DeepSeek-R1 (January 2025), were trained with RL on tasks whose answers can be checked, like maths and code.

Turing Award: Andrew Barto and Richard Sutton received the 2024 ACM A.M. Turing Award, announced March 2025, for the foundations of RL
Standard textbook: Reinforcement Learning: An Introduction by Sutton and Barto (second edition, MIT Press, 2018)
Landmark: AlphaGo mixed supervised learning from expert games with RL from self-play, and beat the European Go champion 5 games to 0 (Nature, January 2016)
In chatbots: OpenAI says ChatGPT (November 2022) was trained with RLHF, using the same methods as InstructGPT

Questions people ask

What is reinforcement learning in simple words?

Learning by trial and error with a score. The AI tries something, gets a reward or a penalty, and adjusts so it earns more reward next time. Repeated many thousands of times, this can make it very good at a task nobody explained to it.

How is reinforcement learning different from supervised learning?

Supervised learning is told the correct answer for every example. Reinforcement learning is only told how good its result was, often long after the action that mattered, and has to work out for itself which actions earned the reward.

What is reinforcement learning used for?

Game-playing systems like AlphaGo, robots learning to walk or grip, and training chatbots and reasoning models to give answers that people prefer or that check out as correct.

What is RLHF?

Reinforcement learning from human feedback. People compare a chatbot's answers and pick the better ones. A reward model learns those preferences, and reinforcement learning then pushes the chatbot toward answers the reward model scores highly.

Who invented reinforcement learning?

Its ideas draw on decades of psychology and control theory. Andrew Barto and Richard Sutton laid its modern foundations in a series of papers beginning in the 1980s, wrote the standard textbook, and received the 2024 Turing Award for it.

Reinforcement learning

In plain words

Learning a new phone game

How it actually works

Agent, environment, reward

Learn a policy

Explore versus exploit

Credit for delayed rewards

Inside language models