How machines learn
Reinforcement learning is machine learning where a program learns by trial and error, receiving rewards for good actions and penalties for bad ones.
Think of training a dog. You do not explain the trick; you reward it when it gets closer to the right behavior. Reinforcement learning works the same way. A program called an agent takes actions in some environment, gets a score, and gradually learns which actions lead to higher scores over time.
The hard part is that rewards can come late. A chess move may only pay off twenty moves later. The agent has to figure out which earlier choices deserve credit. It also must balance trying new things against repeating what already works. Both problems are central to the field.
Reinforcement learning taught programs to beat world champions at Go, to control robots, and to make chatbots more helpful through RLHF, where the reward comes from human feedback. It is powerful but data-hungry, often needing millions of practice runs in a simulator.
A program learns a platform game by playing it thousands of times, discovering that jumping over pits earns more points than falling into them.