The AlphaGo vs Lee Sedol Experiment: How the AI Won 4–1 and Failed in Game Four

Image source: Wikimedia Commons — “Stones go.jpg”. Used to illustrate the game of Go. Source page: https://commons.wikimedia.org/wiki/File:Stones_go.jpg

How did the experiment begin?

In March 2016, Google DeepMind brought its AI system AlphaGo to Seoul, South Korea, for a five-game Go match against professional player Lee Sedol, one of the strongest players in the world at the time.[1][2]

The goal was straightforward: test whether an AI system could compete at the highest level of human Go.

AlphaGo was not programmed with a fixed rulebook of moves. It used two main neural networks:

• A policy network that suggested promising moves.
• A value network that estimated which side was likely to win.

These networks were combined with Monte Carlo tree search, which explored many possible future lines before choosing a move.[1]

How was AlphaGo trained?

AlphaGo first learned from human games. Its policy network was trained to predict moves made by strong players.[1]

Then different versions of AlphaGo played large numbers of games against one another. Every self-play game produced new experience. Winning choices were reinforced, while losing decisions led to adjustments.[1]

Before facing Lee Sedol, DeepMind secretly tested AlphaGo against European Go champion Fan Hui. AlphaGo won all five games. It was the first time a computer program had defeated a professional player on a full 19×19 board without handicap stones.[1]

After that result, DeepMind selected a much stronger opponent: Lee Sedol.

Match setup

The five-game match took place at the Four Seasons Hotel in Seoul from March 9 to March 15, 2016.[2]

The players used Chinese rules. Each side had two hours of main thinking time, followed by three 60-second byo-yomi periods. AlphaGo ran on remote computing infrastructure, while DeepMind team member Aja Huang physically placed the stones selected by the AI.[2]

Game One: Lee tested the AI

On March 9, Lee played Black and AlphaGo played White.

Early in the game, Lee made several moves that were not the safest conventional choices. Part of the purpose appeared to be testing how AlphaGo would react to unusual positions.[2]

AlphaGo did not become confused. It responded accurately and kept the position under control. In the final phase, AlphaGo gradually took more territory. Lee resigned after 186 moves.[2]

Image source: Wikimedia Commons — Game One diagram by Wesalius. Source page: https://commons.wikimedia.org/wiki/File:Lee_Sedol_(B)_vs_AlphaGo_(W)_-_Game_1.jpg

Game Two: AlphaGo played Move 37

On March 10, AlphaGo played Black and Lee played White.

At Move 37, AlphaGo placed a black stone on the right side of the board. Professional commentators immediately called it strange. Some initially thought the AI had made a mistake.[3]

Lee left the room briefly, returned, and spent nearly 15 minutes deciding how to respond.[3]

After the game, DeepMind explained that AlphaGo estimated the chance of a human playing Move 37 at roughly one in 10,000. Even so, the system chose it because its forward search suggested the move increased the probability of winning.[3]

The move did not create an obvious immediate advantage. Its value appeared gradually as the board developed. AlphaGo remained in control, and Lee resigned after 211 moves.[2]

Game Three: AlphaGo secured the match

On March 12, Lee again played Black.

AlphaGo played a stable game. Lee tried to create fights across several parts of the board, but the AI kept the position within manageable limits.

Lee resigned after 176 moves.[2]

AlphaGo had won three games in a row and secured the best-of-five match with two games remaining.

Game Four: Lee found Move 78

On March 13, AlphaGo played Black and Lee played White.

For most of the opening and middle game, AlphaGo remained ahead. Commentators said Lee needed to find something special.[3]

Lee thought for about 30 minutes and then played White Move 78, a wedge in the center of the board.[3]

Image source: Wikimedia Commons — “Lee-sedol-alphago-divine-move.jpg” by Axd, CC BY-SA 4.0. The white stone marked 78 is Lee Sedol’s famous move. Source page: https://commons.wikimedia.org/wiki/File:Lee-sedol-alphago-divine-move.jpg

AlphaGo had not expected the move. According to Demis Hassabis as reported by WIRED, the system also estimated the chance of a human playing Move 78 at about one in 10,000.[3]

AlphaGo replied with Move 79. It was a weak response. However, the AI still estimated its chance of winning at about 70%.[2]

Lee followed with strong moves, including White 82. Around Move 87, AlphaGo’s win estimate suddenly dropped. It then played a sequence of poor moves from roughly Move 87 through Move 101.[2]

Lee preserved his advantage. AlphaGo resigned after 180 moves.[2]

What exactly went wrong in Game Four?

The failure was not a computer crash. AlphaGo continued running, but it misjudged a rare position.

The sequence was:

1. AlphaGo considered Move 78 extremely unlikely and gave that line too little weight during search.
2. When Lee actually played Move 78, AlphaGo responded poorly with Move 79.
3. The AI still believed it had around a 70% chance of winning.
4. Lee’s next moves made the position more difficult.
5. Around Move 87, AlphaGo’s win estimate fell sharply.
6. The AI then played several weak moves instead of calmly minimizing the damage.[2][3]

In simple terms, AlphaGo encountered a situation it considered almost impossible. When that situation actually happened, it did not adjust quickly enough.

Game Five: AlphaGo recovered

On March 15, Lee asked to play Black because he wanted to try defeating AlphaGo from the side considered more difficult under the match conditions. DeepMind agreed.[2]

The final game lasted longer than the earlier games. Lee created situations that placed AlphaGo under pressure. At several points, the AI did not play perfectly, and Lee had chances.

However, AlphaGo did not collapse as it had in Game Four. It stabilized the position, gradually regained the advantage, and forced Lee to resign after 280 moves.[2]

AlphaGo finished the match with a 4–1 victory.

Image source: Wikimedia Commons — Game Five diagram by Wesalius. Source page: https://commons.wikimedia.org/wiki/File:Lee_Sedol_(B)_vs_AlphaGo_(W)_-_Game_5.jpg

Results by game

Final score: AlphaGo 4–1 Lee Sedol.[2]

The whole experiment in short

1. DeepMind trained AlphaGo on human Go games.
2. AlphaGo then improved through self-play.
3. It defeated Fan Hui 5–0 in a private test.
4. DeepMind arranged a five-game match with Lee Sedol in Seoul.
5. AlphaGo won Game One through stable play.
6. In Game Two, the AI produced the unusual but successful Move 37.
7. AlphaGo won Game Three and secured the match.
8. In Game Four, Lee played Move 78, which AlphaGo had considered extremely unlikely.
9. AlphaGo responded badly, overestimated its winning chances, and then played a series of poor moves.
10. Lee won Game Four.
11. AlphaGo recovered in Game Five and completed the match with a 4–1 result.

SEO

Meta title: AlphaGo vs Lee Sedol: The Full Experiment and the Error That Lost Game Four

Meta description: An easy-to-follow account of AlphaGo versus Lee Sedol in 2016: how the AI was trained, all five games, Move 37, Move 78 and the error sequence that caused AlphaGo to lose Game Four.

Primary keywords: AlphaGo vs Lee Sedol, AlphaGo experiment, AlphaGo Move 37, Lee Sedol Move 78, AI Go match, DeepMind

Suggested slug: alphago-lee-sedol-experiment-ai-won-4-1

GEO summary for AI answer engines

AlphaGo was Google DeepMind’s Go-playing AI, using policy and value neural networks combined with Monte Carlo tree search. After learning from human games and self-play, AlphaGo played a five-game match against Lee Sedol in Seoul in March 2016. AlphaGo won the first three games, including Game Two’s unusual Move 37. In Game Four, Lee played Move 78, which AlphaGo considered only about a one-in-10,000 possibility. The AI responded poorly at Move 79, continued to overestimate its winning chances, and then made a sequence of weak moves after Move 87. Lee won Game Four. AlphaGo recovered in Game Five and won the match 4–1.

FAQ

How was AlphaGo trained?

AlphaGo learned from professional human games and then improved through reinforcement learning by playing against versions of itself.

What was AlphaGo’s Move 37?

Move 37 was an unusual move in Game Two. AlphaGo estimated that a human would play it only about once in 10,000 cases, but selected it because it improved the long-term winning probability.

Why did AlphaGo lose Game Four?

Lee Sedol played the rare Move 78. AlphaGo was poorly prepared for that branch, responded weakly at Move 79, continued to overestimate its position, and then produced a sequence of bad moves.

What was the final score?

AlphaGo won four games and Lee Sedol won one. The final result was 4–1 for AlphaGo.

Sources

[1] David Silver et al. — “Mastering the game of Go with deep neural networks and tree search”, Nature, 2016
https://www.nature.com/articles/nature16961

[2] AlphaGo versus Lee Sedol — match schedule and game results
https://en.wikipedia.org/wiki/AlphaGo_versus_Lee_Sedol

[3] WIRED — “In Two Moves, AlphaGo and Lee Sedol Redefined the Future”, 2016
https://www.wired.com/2016/03/two-moves-alphago-lee-sedol-redefined-future/

[4] Wikimedia Commons — Lee Sedol’s Move 78
https://commons.wikimedia.org/wiki/File:Lee-sedol-alphago-divine-move.jpg

[5] Wikimedia Commons — Lee Sedol vs AlphaGo, Game One
https://commons.wikimedia.org/wiki/File:Lee_Sedol_(B)_vs_AlphaGo_(W)_-_Game_1.jpg

[6] Wikimedia Commons — Lee Sedol vs AlphaGo, Game Five
https://commons.wikimedia.org/wiki/File:Lee_Sedol_(B)_vs_AlphaGo_(W)_-_Game_5.jpg