AlphaGo is a Go-playing program built by DeepMind that combined deep neural networks with Monte Carlo tree search. In March 2016 it beat Lee Sedol, an 18-time world champion, by four games to one in Seoul, after defeating the European champion Fan Hui five games to zero in October 2015. Later versions, AlphaGo Zero and AlphaZero, learned through self-play alone, and the same research group went on to build AlphaFold for protein structure prediction.
Key takeaways
- The Lee Sedol match ran from March 9 to March 15, 2016. AlphaGo won games one, two, three, and five, and Lee won game four.
- AlphaGo used a policy network to pick candidate moves and a value network to judge positions, both guided by tree search.
- Training combined supervised learning from human expert games with reinforcement learning from self-play.
- AlphaGo Zero dropped human data entirely and won 100 games to 0 against the earlier champion-defeating version.
- Games are useful test beds because rules are fixed, outcomes are unambiguous, and unlimited practice games can be generated by simulation.
The two matches that made headlines
DeepMind's account lists October 2015 as the first time an AI program beat a professional Go player, when AlphaGo won five games to zero against Fan Hui. The result was published in the journal Nature in January 2016 in the paper "Mastering the game of Go with deep neural networks and tree search," which also reported a 99.8% winning rate against other Go programs.
The better-known match took place at the Four Seasons Hotel in Seoul. The British Go Association's record gives the dates and outcomes below.
| Game | Date in 2016 | Winner |
|---|---|---|
| 1 | March 9 | AlphaGo |
| 2 | March 10 | AlphaGo |
| 3 | March 12 | AlphaGo |
| 4 | March 13 | Lee Sedol |
| 5 | March 15 | AlphaGo |
Two moves are remembered. In game two, AlphaGo's move 37 was a choice DeepMind says had about a one in 10,000 chance of being played by a human. In game four, Lee's move 78, nicknamed "God's Touch," carried a similar estimate and helped him win the only game he took.
How AlphaGo chose a move
Go is hard for brute force. DeepMind puts the number of possible board configurations at about 10 to the power of 170. Searching every line is impossible, so the program needs two kinds of judgment: which moves deserve attention, and who is probably winning in a given position.
The Nature paper supplies both with neural networks. A policy network selects promising moves, which narrows the search. A value network evaluates board positions, which shortens it. Monte Carlo tree search ties them together. It grows a tree of possible continuations by running many simulated games from the current position, and it sends more simulations down branches that keep scoring well. After the budget is spent, the most-visited move is played.
An illustrative example: suppose a position has three candidate moves, and after 100 simulations the moves have been visited 70, 20, and 10 times. The program plays the first. These counts are invented to show the selection rule, not taken from a real game.
How it was trained
DeepMind describes a two-stage process. The networks first learned from thousands of expert human games through supervised learning, which taught them to imitate strong play. They then improved by playing against themselves, using reinforcement learning, where the program receives a win or loss at the end of each game and adjusts its networks so that moves leading to wins become more likely. The self-play stage let the system exceed the level of the human games it began with, because it could keep generating new positions on its own.
AlphaGo Zero and AlphaZero
Later papers removed assumptions one at a time.
| System | Publication | What changed | Reported result |
|---|---|---|---|
| AlphaGo | Nature, January 2016 | Human games plus self-play, with tree search | 5-0 against Fan Hui, 4-1 against Lee Sedol |
| AlphaGo Zero | Nature, October 2017 | Reinforcement learning only, no human data or domain knowledge beyond the rules | 100-0 against the earlier champion-defeating AlphaGo |
| AlphaZero | Science, December 2018 | One algorithm for chess, shogi, and Go | Defeated a world champion program in each game |
The AlphaGo Zero abstract says the network is trained to predict both AlphaGo's own move choices and the winner of its games, and that this improves the tree search, which in turn produces stronger self-play data in the next round.
Why games matter for AI research
Board games make a clean laboratory. The rules are complete, the state is fully visible, and a game ends with an unambiguous result, so a learning signal is always available. Simulated games cost little, which makes millions of practice rounds feasible. Progress is also easy to measure against humans and earlier programs. These properties explain why methods first proven on games are tested on other problems later. They also mark the limit: real-world tasks usually have hidden information, unclear goals, and costly mistakes.
From Go to protein folding
DeepMind's AlphaFold page says that AlphaGo's 2016 win suggested its techniques might be advanced enough for scientific problems, and a small team then began work on protein structure prediction. AlphaFold2 won the CASP14 assessment on November 30, 2020, and the Nature paper appeared on July 15, 2021. Demis Hassabis and John Jumper shared half of the 2024 Nobel Prize in Chemistry for protein structure prediction, announced on October 9, 2024. AlphaFold is a different system from AlphaGo, built around a neural network that encodes biological and physical knowledge about proteins, so the link is shared research lineage, not shared code.
FAQ
Did AlphaGo learn only from human games?
The 2016 version did not. It started from human expert games and then improved through self-play. AlphaGo Zero, published in 2017, used no human data.
What is Monte Carlo tree search?
It is a method that estimates which move is best by running many simulated continuations and favoring the branches that perform well. In AlphaGo, neural networks guide which branches to explore and how to judge positions.
Why did AlphaGo lose game four?
Records show only that Lee Sedol won on March 13, 2016, and that his move 78 was widely praised. Accounts of the cause differ, so no single explanation is established here.
Is AlphaGo the same as AlphaFold?
No. They are separate systems for different problems. DeepMind has said that the 2016 match helped prompt its move into scientific problems.
Related reading
- DeepMind courses for learners
- DeepMind research presented at ICLR 2023
- A research paper view of AI progress
- Google's LaMDA language model
More in DeepMind.
Last verified: October 2026
Sources
- Silver, D. et al., "Mastering the game of Go with deep neural networks and tree search," Nature, 2016: source page
- DeepMind, AlphaGo research page, accessed October 2026: deepmind.google
- Silver, D. et al., "Mastering the game of Go without human knowledge," Nature, 2017
- Silver, D. et al., "A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play," Science, 2018
- British Go Association, "Google DeepMind Challenge Match: Lee Sedol v AlphaGo, match report," 2016
- DeepMind, AlphaFold research page, accessed October 2026
- Jumper, J. et al., "Highly accurate protein structure prediction with AlphaFold," Nature, 2021
- Science News, "Work on protein structure and design wins the 2024 chemistry Nobel," October 2024