Milestone · 1998

Reinforcement Learning: An Introduction

Sutton and Barto gathered scattered methods into one discipline with shared notation, problems and theory.

Why it matters

A direction that had existed for twenty years as a set of tricks became a subject with a textbook, and so became teachable.

The book introduces a common language: Markov decision process, policy, value function, the trade-off between exploration and exploitation. It also separates honestly what is proved from what only works empirically. The authors put the full text online. The 2018 second edition added deep methods; this is the book the generation that built AlphaGo learned from.

Event record

Event date

1998

Timeline date

Event date

Verification

Sources gathered automatically · September 17, 2026

Lines

Learning theory

ID

evt-0180

Year of the MIT Press first edition.

Sources

primary Reinforcement Learning: An Introduction
MIT Press · Published 1998

Related events

Extends Temporal-difference learning
The book systematises the methods the temporal-difference paper began.
Reinforcement Learning: An Introduction
Extends Actor and critic
The actor-critic scheme is presented as part of a general theory.
Reinforcement Learning: An Introduction
Extends Q-learning
Q-learning is set out as one method within a shared frame.
Reinforcement Learning: An Introduction

Records that link to this one

Related Stanley wins the Grand Challenge
The vehicle's control was tuned by machine learning on recordings of human driving.
Builds on DQN plays Atari
The algorithm is Q-learning with the table replaced by a convolutional network.
Playing Atari with Deep Reinforcement Learning
Related InstructGPT and learning from human ratings
The fine-tuning is reinforcement learning against a reward derived from human ratings.