DeepMind Publishes AlphaGo, a Deep Reinforcement Learning System That Defeated Professional Go Players
In January 2016, researchers at Google DeepMind published a paper in Nature describing AlphaGo, a system combining deep convolutional neural networks with Monte Carlo tree search and reinforcement learning that defeated the European Go champion Fan Hui 5–0, marking the first time a computer program had beaten a professional Go player at full-board Go.
Background
Chess fell to computers in 1997, when IBM’s Deep Blue defeated Garry Kasparov. Checkers had been solved earlier. The pattern looked clear enough: given enough processing power and clever search, machines would eventually master any board game.
Go was different. A chess position has an average of about 35 legal moves at each turn. A Go position has about 250. That branching factor makes exhaustive search completely unworkable, even for modern hardware. A program that tries to look ahead by brute force will run out of time and memory long before it finds anything useful.
The approaches that had worked in chess relied heavily on hand-crafted rules and position evaluations written by experts. Go players tried the same thing, and it didn’t work well. The game’s sheer complexity resisted that kind of encoding, and even the best Go programs of the early 2010s played at the level of a strong amateur. Professional players were nowhere near threatened.
What happened
In January 2016, David Silver, Aja Huang, and colleagues at Google DeepMind published a paper in Nature describing a system they called AlphaGo. It combined several techniques that, taken individually, were already known. The team’s contribution was putting them together in a way that actually worked.
AlphaGo used two deep convolutional neural networks, each scanning the board the way an image recognition system scans a photograph, in overlapping patches rather than pixel by pixel. One, called the policy network, was trained first on a large set of recorded games between human expert players, learning to predict which moves strong players tended to make. A second network, the value network, was trained to estimate how likely a given board position was to lead to a win. Once that foundation was in place, the system played games against itself using reinforcement learning, updating both networks based on the outcomes. It found moves that human training data had never shown it.
Monte Carlo tree search, a method that builds a search tree by running many random sample games forward rather than examining every branch, tied it together. AlphaGo used the policy network to decide which branches were worth exploring, and the value network to judge positions without playing them all the way to the end. The result was a system that could look ahead sensibly without having to consider the full breadth of possible play.
The team tested AlphaGo against Fan Hui, who held the European Go champion title at the time. AlphaGo won all five games in that match, becoming the first computer program to defeat a professional Go player in full-board Go without a handicap. The match was held before the paper’s publication and refereed formally to make the result credible to the Go community. Shortly after publication, AlphaGo went on to play Lee Sedol, one of the highest-ranked players in the world, in a match in March 2016, winning 4–1.
Why it mattered
Go had long resisted the search-based approaches that had proven effective in chess, because its branching factor, the number of possible moves per position, is far too large for exhaustive look-ahead to be tractable. AlphaGo's victory demonstrated that combining supervised learning from human expert games with self-play reinforcement learning could produce superhuman performance in a domain previously considered beyond the reach of machine methods. The result accelerated interest in reinforcement learning and deep learning as general problem-solving frameworks, well beyond game-playing contexts.
People
David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George van den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, Sander Dieleman, Dominik Grewe, John Nham, Nal Kalchbrenner, Ilya Sutskever, Timothy Lillicrap, Madeleine Leach, Koray Kavukcuoglu, Thore Graepel, Demis Hassabis
Organisations
Sources
- Mastering the game of Go with deep neural networks and tree search.Nature.Primary source
- Mastering the game of Go without human knowledge.Nature.Secondary
- AlphaGo: The story so far.Google DeepMind.Official
Cite this page
AI Achievements. (2016). DeepMind Publishes AlphaGo, a Deep Reinforcement Learning System That Defeated Professional Go Players. Retrieved 2026-08-22, from https://achievements.ai/milestone/an-algorithm-alphago-by-google
@misc{achievements_an_algorithm_alphago_by_google,
title = {DeepMind Publishes AlphaGo, a Deep Reinforcement Learning System That Defeated Professional Go Players},
author = {{AI Achievements}},
year = {2016},
url = {https://achievements.ai/milestone/an-algorithm-alphago-by-google}
}