Reinforcement Learning (RL)

A learning approach in which an agent improves through rewards and penalties from its environment.

Which choice earns the reward three moves later?

In a toy exhibit puzzle, the visitor gets three moves. Picking up a key, opening a cabinet and placing a tile earns one point when the tile reaches the display. All earlier moves earn zero.

Policy A explores the key

Its first move earns zero. The cabinet becomes accessible on the second move and the third completes the display. The undiscounted return is one.

Policy B repeats a familiar move

It presses an illuminated button three times. The puzzle visibly responds each time, but the display stays empty. Its return is zero.

Why can an unrewarded action matter?

A policy selects actions from the current state. To learn the value of taking the key, the agent needs experience connecting that early move to a later reward. Comparing complete episodes reveals a difference the first reward alone cannot show.

Common Confusions

Reinforcement learning is often confused with any system that improves through feedback. In strict use, RL involves actions, state, reward, and policy learning. Human feedback methods such as RLHF borrow related ideas, but they are usually part of a broader post-training workflow rather than a direct robot-like environment.

Follow the idea

AI Agent · Consider how an agent chooses its next action.

RLHF · See where preference judgments enter a reward signal.

Related concepts (7)

Examples in the AI guides

Related Yenra articles (11)

Source · OpenAI’s introduction to reinforcement learning

Return to Reinforcement Learning (RL) in the A–Z glossary