AlphaGo Zero and AlphaZero
Era 4 · Deep learning and foundation models (2012–2021) › Deep reinforcement learning
- Lineage
- Data and learning
- Kind
- Implementation
Who
DeepMind
What happened
DeepMind publishes AlphaGo Zero in Nature, a program that learns Go with no human games and beats the version that defeated Lee Sedol 100–0 (October 19, 2017). Seven weeks later it presents AlphaZero, which uses the same method to master chess, shogi, and Go within 24 hours.
Why it mattered
Learning purely through self-play, with no human knowledge beyond the rules. A single algorithm reaches superhuman level in three different games.
The full story
The 2016 AlphaGo had started by imitating expert games. AlphaGo Zero did without them: it started from random moves, knew only the rules, and learned by playing against itself. It used a single network that both proposed moves and evaluated positions, guiding a tree search whose results were in turn used to train it.
The paper, published in Nature on October 19, 2017, under the title "Mastering the game of Go without human knowledge," charted its progress. In three days it surpassed the version that had beaten Lee Sedol, defeating it 100 games to 0. In 21 days it caught up with AlphaGo Master, the version that had beaten Ke Jie, and in 40 days it surpassed every earlier version.
On December 5 the team posted AlphaZero on arXiv, a generalization of the method. With no game-specific adaptations, and also starting from random moves, within 24 hours it reached superhuman level at chess, shogi, and Go, and in each one it beat a world-champion program. The best chess programs, by contrast, relied on evaluation functions tuned by hand by experts over decades. The conditions of the match against Stockfish, with one minute per move and no opening book, drew criticism, and the version published in Science in December 2018 repeated the contest with more time and the latest version of Stockfish: AlphaZero won 155 games, lost 6, and drew 839.
The result showed that human knowledge could be a starting point, but neither a requirement nor a ceiling.
Videos
- AlphaGo Zero: Starting from scratch (opens in a new tab) (Google DeepMind)
Builds on
AlphaGo beats Lee Sedol (March 9, 2016)
Documented direct influence
AlphaGo Zero is matched against the version that beat Lee Sedol and wins 100 games to 0 after three days of training
TPU (May 18, 2016)
Documented direct influence
AlphaGo Zero plays on four TPUs, the chip Google had unveiled in May 2016
Threads that pass through here
- From Atari to Go (step 4 of 4)
Sources
Original source
- Mastering the game of Go without human knowledge. Nature, 550(7676), pp. 354-359 (opens in a new tab). David Silver, Julian Schrittwieser et al., 2017
- Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm (arXiv:1712.01815) (opens in a new tab). David Silver, Thomas Hubert, Julian Schrittwieser et al., 2017
- A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play. Science, 362(6419), pp. 1140-1144 (opens in a new tab). David Silver, Thomas Hubert, Julian Schrittwieser, Ioannis Antonoglou et al., 2018
How to cite
“AlphaGo Zero and AlphaZero.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/alphago-zero-and-alphazero
This entry was last reviewed on November 13, 2026.