What happened when AlphaGo played Lee Sedol, and why did the match matter?

In March 2016, a quiet board game became front-page news around the world. For five games in a Seoul hotel, the South Korean Go master Lee Sedol sat across from a person placing stones on behalf of a computer program, DeepMind's AlphaGo. When it was over, the program had won four games to one.

Two months earlier, DeepMind had revealed that AlphaGo had beaten Europe's champion, Fan Hui, five games to none. But Fan Hui was a 2-dan professional; Lee was at the very top of the game. Most observers expected Lee to win. The match in Seoul was the real test.

Who was playing, when and where

Lee Sedol is a 9-dan professional, the highest rank, who turned professional at age 12 in 1996 and had won 18 international titles. In South Korea he is a national hero, known for creative, unconventional play. Before the match, he predicted he would win.

Officially named the Google DeepMind Challenge Match, the event took place at Seoul's Four Seasons Hotel. The first game was played on 9 March 2016 and the fifth on 15 March. The rules followed Chinese scoring with a 7.5-point bonus (called komi) for the player with white, who moves second. Each player had two hours, followed by three 60-second overtime periods. The winner would receive $1 million.

AlphaGo itself was not in the room. It ran on Google's cloud computers in the United States. Aja Huang, a DeepMind researcher and strong amateur player, sat at the board and placed the stones AlphaGo chose. Demis Hassabis, who co-founded DeepMind in 2010, spent the match in Seoul with the team.

The five games

Every game ended by resignation. Two moves became famous. In game 2, AlphaGo's move 37 struck commentators as creative and original; DeepMind's own data suggested a human would have played it only about once in 10,000 times. In game 4, Lee answered with a move of his own that DeepMind put at the same 1-in-10,000 odds. AlphaGo misjudged it, its estimate of its chances collapsed several moves later, and it played a run of poor moves before resigning.

How AlphaGo chose its moves

AlphaGo used the design DeepMind had described in the journal Nature that January. A policy network, a neural network trained first on expert human games and then by playing millions of games against itself, suggested promising moves. A value network estimated who was winning from any position. A search method called Monte Carlo tree search used both to look ahead and pick a move.

One detail explains much of its style. AlphaGo tries to maximize its chance of winning, not its margin of victory. Faced with a choice between a likely big win and an almost certain narrow one, it takes the narrow one. To human eyes, that sometimes looked like slack or timid play, especially late in games when it gave up points it did not need.

Its weakness showed in game 4. Tree search prunes away lines of play that look unlikely, and Lee's move 78 was exactly the kind of rare move that can fall off the search's radar. Once AlphaGo realized its position was worse than it had thought, it struggled to recover.

Why it mattered

The match drew an enormous audience. Live streams in China were followed by tens of millions of people, and Chinese coverage of the first game alone reached about 60 million viewers. Sales of Go boards rose in Korea. After the match, the Korea Baduk Association gave AlphaGo an honorary 9-dan rank, and Science named the match a runner-up for its 2016 Breakthrough of the Year. DeepMind gave the $1 million prize to charities including UNICEF and to groups supporting science education and Go.

For AI research, the result marked the fall of the last great classic board game. Many experts had believed a program this strong was at least five to ten years away. Just as important, AlphaGo had learned much of its skill through general learning methods rather than handcrafted rules, which is why Hassabis described it as a step toward systems that could help with problems far beyond games, while stressing that machines able to learn any task a human can were still a long way off. Lee himself took the loss gracefully, saying it was "not a defeat of mankind."

What happened to AlphaGo and Lee Sedol after the match?

AlphaGo kept playing. In 2017 it faced China's Ke Jie, then the world's top-ranked player, and won all three games. DeepMind has since built successors, starting with AlphaGo Zero in 2017, and has turned its methods toward science with AlphaFold, a system for predicting the shapes of proteins.

The match also changed how Go players think about their own game. Go officials, including the referee of the Fan Hui match, predicted that players would come to use computers to find their mistakes and improve, and top professionals described AlphaGo's moves as original ideas worth studying rather than errors.

Lee Sedol continued playing professionally for three more years. On 19 November 2019 he announced his retirement, saying that with AI's growing strength he could never be the top player, and describing the programs as an entity that cannot be defeated. An award-winning documentary about the match, AlphaGo, was released in 2017, and in March 2020 DeepMind made it free to watch online.