Learning by trying things and receiving rewards or penalties, rather than being shown correct answers. It is how you would train a dog, and how a system learns to play a game: make moves, see whether you win, do more of what led to winning. It suits situations with a clear goal but no obvious right answer at each individual step.
Technical view
An agent observes a state, chooses an action from a policy, receives a reward, and moves to a new state, aiming to maximise cumulative reward over time. The central difficulty is credit assignment: a reward may arrive long after the move that caused it, so the system must work out which of many earlier actions deserve credit. A second difficulty is the balance between exploration, trying new things, and exploitation, repeating what already works. Reinforcement learning is powerful but notoriously unstable and data-hungry compared to supervised learning.
Related
Part of AI Terminology Index, section: The Foundations.