reinforcement learning
Reinforcement learning is how a machine learns the way you'd train a puppy: not by being shown the right answer, but by trying things and being rewarded or scolded for the results. The learner — called an agent — takes an action, sees what happens, and pockets a reward or a penalty. Over thousands of attempts, it drifts toward whatever earns the most reward. Nobody ever tells it the correct move; it has to discover that by playing.
This is what makes it so powerful for games and robots. A program learning to play chess or a video game gets a single signal — you won, you lost, you scored — and must figure out for itself which of its many moves actually helped. A robot learning to walk gets rewarded for staying upright and moving forward, and stumbles its way to a working gait. The agent is always caught between two urges: exploit the trick it already knows, or explore something new that might pay off even better.
The landmark moment came in 2013–2015, when DeepMind's DQN learned to play dozens of Atari games from nothing but the raw pixels on screen and the score — and beat human experts at many of them, using the same algorithm, network design, and settings for every game (with a fresh copy trained separately on each one). The common misconception is that the machine was 'programmed to win.' It wasn't: it started out flailing randomly, and learned purely from the consequences of its own fumbling.
AlphaGo learned to play Go partly by playing millions of games against itself, rewarded only by winning — and went on to beat the world's best human player.
The name comes from psychology: B. F. Skinner's mid-20th-century experiments showed that behaviour followed by a reward tends to be repeated — 'reinforced.' The modern machine version was forged in the 1980s–90s, when researchers like Richard Sutton and Andrew Barto gave the idea its mathematical backbone.