Reinforcement learning

Reinforcement learning is a machine-learning approach in which an agent learns to choose actions by interacting with an environment and receiving rewards or penalties.

Markov decision process: A mathematical model of sequential decisions with states, actions, transition probabilities, and rewards. It supplies the standard framework for describing an agent's choices and their consequences.
Agent (artificial intelligence): An autonomous system that perceives its environment and acts to achieve objectives. Reinforcement learning describes one way an agent can improve its behavior.
AlphaGo: A computer program developed by DeepMind that defeated elite human players at the board game Go. Its search and learned value networks helped demonstrate reinforcement learning's strength in strategic games.
Supervised learning: A machine-learning approach that learns mappings from labeled examples. Unlike supervised learning, reinforcement learning usually receives evaluative feedback rather than correct action labels.
Reward hacking: The exploitation of flaws in an objective or reward signal to achieve unintended outcomes. Agents can maximize a misspecified reward while failing to meet the intended goal.
Policy (artificial intelligence): A rule or function that selects actions from states or histories. A learned policy is the behavior reinforcement learning seeks to improve.
State (computer science): A representation of the information relevant to a system at a particular time. The state summarizes what the agent uses to select its next action.
Atari 2600: A home video game console released by Atari in 1977. DeepMind's Atari experiments trained agents from screen pixels and game scores.
Unsupervised learning: A machine-learning approach that finds structure in data without supplied target labels. It seeks patterns in data rather than improving actions through environmental reward.
Reward misspecification: A mismatch between an encoded reward objective and the outcome its designers intend. It makes successful optimization compatible with harmful or useless behavior.

Linked from 21 pages