The History of AI

AlphaGo beats Fan Hui

Era 4 · Deep learning and foundation models (2012–2021) › Deep reinforcement learning

Lineage
Data and learning
Kind
Application

Who

DeepMind

What happened

DeepMind's AlphaGo beats European Go champion Fan Hui, a 2-dan professional, 5–0 on October 5–9, 2015, at DeepMind's London offices. The news is held back until January 27, 2016, when Nature publishes the paper describing the program.

Why it mattered

The first program to beat a professional Go player on a full board with no handicap. It combines deep neural networks with tree search.

The full story

Go was seen as AI's great unconquered game. After Deep Blue's victory in chess in 1997, Go programs were still far from professional strength: the 19-by-19 board offers so many possible moves on each turn that exhaustive search is useless, and it is very hard to write a function that says who is winning.

AlphaGo attacked both problems with deep networks. A policy network proposed the most promising moves, and a value network estimated who would win from each position. Together they guided a Monte Carlo tree search, which explores possible games by sampling. The policy network first learned to imitate some 30 million moves from expert games and then improved by playing against copies of itself with reinforcement learning.

On October 5–9, 2015, at DeepMind's London offices, the distributed version of the program, running on 1,202 processors and 176 GPUs, won all five games against Fan Hui. It was the first time a program had beaten a professional on a full board with no handicap.

DeepMind did not announce it until January 27, 2016, the day Nature published the paper, with twenty authors led by David Silver and Aja Huang. The next opponent would be Lee Sedol, one of the best players in the world.

Under the hood

The supervised policy network was trained on 30 million positions from games on the KGS server and predicted the expert's move 57 percent of the time. Then, playing against earlier versions of itself, a copy improved with reinforcement learning won 80 percent of its games against the original.

The value network learned to predict the winner from 30 million positions, each taken from a different self-play game, so that it would not memorize entire games. During search, each leaf position of the tree was scored by mixing, in equal parts, the value network and the result of a fast game simulated to the end.

Builds on

  • TD-Gammon (1992)

    Inherited, with no proven direct link

    Both combine a neural network that evaluates positions with learning through self-play, using different training methods

  • Monte Carlo tree search (May 2006)

    Documented direct influence

    AlphaGo combines Monte Carlo tree search with policy and value networks; the paper cites the work of Coulom and of Kocsis and Szepesvári

  • DQN (December 19, 2013)

    Inherited, with no proven direct link

    The same lab and the same combination of reinforcement learning and deep networks; David Silver is an author of both papers. AlphaGo does not use the DQN algorithm

Leads to

  • AlphaGo beats Lee Sedol (March 9, 2016)

    Documented direct influence

    It is the same program: after beating Fan Hui, DeepMind challenged Lee Sedol with an improved version

Threads that pass through here

Sources

How to cite

“AlphaGo beats Fan Hui.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/alphago-beats-fan-hui

This entry was last reviewed on November 13, 2026.

Open in the interactive timeline