Reinforcement learning

Learning from reward

Supervised learning needs the right answer for every example. But for many problems nobody knows the right move — only whether things worked out in the end. A game is won or lost; a robot reaches the door or falls over. Reinforcement learning learns from that sparse signal.

An agent sees a state, takes an action, and the environment returns a new state and a reward. The agent’s goal: maximise the total reward over time. It has to discover, by trial and error, which actions eventually pay off.