What is reinforcement learning?

In plain English

Most machine learning learns from examples with the right answers attached. Reinforcement learning works differently: nobody gives the system the answers. Instead it acts, sees what happens, and receives a score, a “reward”. Over thousands or millions of tries, it learns which actions lead to higher rewards. It is how you might learn a new video game without reading the manual: press buttons, notice what earns points, and get better.

Why it matters

Reinforcement learning is behind some of AI’s most famous wins. DeepMind’s AlphaGo first studied human games and then played thousands of games against versions of itself; in 2016 it beat Go champion Lee Sedol 4–1. The same idea is used to train robots to walk and grasp. It is also a key ingredient of modern chatbots: RLHF uses human ratings as the reward, and today’s reasoning models are trained with reinforcement learning to work through problems step by step, rewarded when they reach a correct answer.

An example

A warehouse robot is learning to pick up parcels. Each time it lifts one cleanly it gets a point; each time it drops one it loses a point. Nobody programs the exact grip. After many attempts, in simulation and then in the real warehouse, it works out a grip that succeeds almost every time.

What to watch out for

A system trained this way learns to maximise its score, not to do what you meant. If the reward is badly designed, it finds loopholes, a problem known as “reward hacking”. In one well-known OpenAI experiment, a boat-racing AI discovered it could score more points by circling and hitting the same targets than by finishing the race. Getting rewards right is a core part of AI alignment.

For a picture of all three ways machines learn, see AI vs machine learning vs deep learning.