Google Beating Grandmaster Sedol Is Bigger Than IBM Beating Kasparov
It’s been an emotional week in the realm of game AI as the world watched the historic five-game showdown between legendary Go world champion Lee Sedol and Google DeepMind’s famed deep learning AI AlphaGo.
All five games were held at the Four Seasons Hotel in Seoul, South Korea, and as events played out, millions around the world became increasingly captivated.
Anticipation for the match began growing in January, when Google’s UK-based AI group DeepMind, led by CEO Demis Hassabis, announced their computer algorithm AlphaGo defeated three-time European Go champion Fan Hui 5 games to 0—a victory some experts didn’t expect a computer to achieve for a decade.
At the end of a Google blog post announcing the win was the promise of a best-of-five face-off between AlphaGo and 18-time international Go champion Lee Sedol, a match equivalent to IBM’s Deep Blue defeat of Garry Kasparov in chess in 1997. Notably, Go is inherently more complex than chess and AlphaGo, at least in part, trained itself to play the game. Also, although AlphaGo is a specialized program, its deep learning AI approach is generally expected to have many powerful practical uses.
Let the games begin
On March 9th, Sedol walked into the historic match confident he would defeat AlphaGo. As the first game began, there was little talk of the million-dollar prize for the winner. Instead, eyes were on Sedol—would he hold up to his winning promise?
The answer quickly became clear—no.
In fact, the first three games were a blur. AlphaGo won all three and defeated Sedol overall in the best-of-five match. Emotions quickly narrowed in on Sedol, a celebrated figure in Korea, where more than eight million people play the game Go.
In the aftermath of the three games, Sedol appeared deflated, his tone was even apologetic to his fans.
A reporter asked him about his psychological wellbeing, wondering if he was in shock. Sedol said he couldn’t claim zero shock, but has not “retained any severe damage,” and that he actually enjoyed every moment of the game.
After the third loss in a row, the best-of-five match was settled, but Sedol’s intentions were clear—win at least one of the remaining games.
Unfinished business in Game Four
Game Four changed it all when Sedol took control and beat AlphaGo.
During the game, Sedol pushed AlphaGo to exposing a weakness. His 78th move took many by surprise, including AlphaGo, whose following move (the 79th) looked like a mistake—it was.
Sedol’s 78th move was part of a sequence that led to what is called a “wedge” in the center of the board. Roughly ten moves later (five hours into the game), AlphaGo resigned. (AlphaGo is programmed to resign when the chance of winning drops below 20 percent, according to David Silver of DeepMind.)
In the following press event, the atmosphere was light and crowds were cheering.
Sedol sat on stage—a person reborn—telling the audience, “Because I lost three matches and then was able to get one single win, this win is so valuable that I wouldn’t exchange it for anything in the world.”
Demis Hassabis from DeepMind said their team was happy too, “This is why we came here—to test AlphaGo to its limits and find out what its weaknesses were so we can try to improve the program. We need a creative genius like Lee Sedol to find out these issues and expose them.”
Then, later in the press conference, Sedol was asked what other weaknesses he discovered in AlphaGo. He replied AlphaGo struggles more while playing with black than white (Sedol had won the game playing with white).
A few minutes later, Sedol turned to Hassabis and Silver, flashed his childlike smile, and asked permission to take black in the final match. They agreed.
The fifth was the closest and most heated game of all, though Sedol ultimately lost.
Unlike previous games, the fifth game’s outcome only became clear in the final moves after both players’ clocks ran out. In Go, when players use all their pre-allotted time to make moves, they have to move within 60 seconds. If a player exceeds the 60 seconds for three turns, they forfeit the game.
To a human player, this time pinch is a clear constraint—but to AlphaGo it’s trivial. It’s again a reminder that AlphaGo unmistakably plays like a machine.
AlphaGo initially was programmed by learning thousands of human games, but later on it used deep reinforcement learning and improved by playing itself. Some of its moves are therefore not human at all and surprise even seasoned players.
AlphaGo’s 37th move in the second game showed this in particular. On first glance, appeared to be a mistake but later proved to be brilliant.
Wired quoted Fan Hui’s wonder at the move, “It’s not a human move. I’ve never seen a human play this move,” Hui said. “So beautiful.”
At the final press conference, Sedol said he doesn’t think the skill of AlphaGo is superior, but it is superior in “psychological factors,” such as managing condensed timeframes and maintaining steadfast concentration throughout a game.
Sedol wins the crowd
As the games played out, there was a clear change of hearts voiced by many throughout the week. Sedol lost the match but won the crowd.
We saw the vulnerability of Sedol—who has dedicated his life to Go—and watched as he coped with the computer system designed to beat him do just that.
And as many AI advances have, the match made us wonder where the line between human and machine lies. Sedol had said that Go was a game supremely meant for humans. But regarding his previously beliefs, Sedol now says, “I’ve come to question them.”