Reinforcement learning

Learning better actions by trial and error, guided by a reward given according to the outcome, rather than by being told the right answer.

The learner (the agent) chooses an action in some state and receives a reward as a result. Repeating this, it learns how to choose actions that maximise reward over the long run. Q-learning, devised by Christopher Watkins in 1989, is a representative method; in 1992 it was proved to learn correctly^[1].

Besides Go-playing programs and learning robot movements, it has begun to be used to keep adjusting control settings during error correction in quantum computers.