AlphaGo and Lee Sedol: Seoul, 2016
2016 · ai
In March 2016 the South Korean player Lee Sedol, one of the strongest go players of his generation, sat down in a Seoul hotel for a five game match against a program called AlphaGo. Go is played with black and white stones on a nineteen by nineteen grid and has more legal positions than there are atoms in the observable universe, which had made it the classic example of a game where human intuition still beat calculation. Most experts expected Lee to win comfortably. He lost four games to one.
AlphaGo had been built by the London company DeepMind and combined two ideas, neural networks that learned to judge positions and moves from millions of examples, and a search method that explores promising lines by playing them out. It first studied recorded human games and then improved by playing against versions of itself.
The moment that professionals still discuss came in the second game. AlphaGo played a stone on the fifth line in a way that no strong player would have chosen, and commentators assumed it was an error until the shape of the board turned in its favour many moves later. Lee answered in the fourth game with a move of his own that the program had rated as very unlikely, and won that game, the only defeat this version of AlphaGo ever suffered.
Lee retired from professional play in 2019, saying that an entity now existed that could not be defeated. The programs that followed learned the game from nothing but its rules and reached a higher level within days. Go clubs report that human play has changed since 2016, with openings and shapes borrowed from the machines, and the match is now used as the standard example of what learning systems can do in a domain thought to require judgement rather than force.