By Sunny Um WIRED Korea
In November, Lee Sedol declared that he was quitting Chinese board game Go, because he found it impossible to defeat artificially intelligent (AI) Go computing programs. But a month later, he came out for another round of matches with an AI player.
This week, 18-time Go world champion Lee had first two matches against an AI Go program HanDol, developed by Korean company NHN Entertainment, in Seoul. He won the first match and lost the second.
What is Go?
Go, known as “Baduk” in Korea, is played by two people placing black or white stones on 19-lines-by-19-lines grid board. Each point of the grid can be emptied or occupied under certain rules. One who conquers more territory wins.
Coding a winning Go program is not easy. As Demis Hassabis, the founder and CEO of DeepMind, puts it, “There are more possible Go positions than there are atoms in the universe.”
HanDol the AI Go Player
HanDol is an AI Go player introduced two years ago. Its system predicts the opponent’s next move based on the big data from NHN’s online Go game, collected since 1999. It calculates the overall winning chance from each move, designs a pattern, and finds the best move possible with its algorithm. In January, HanDol defeated the top five Korean Go players.
Lee challenged HanDol after he went into retirement. He no longer competes against human rivals. It is Lee’s first game with an AI Go Player since his match in 2016. The competition with HanDol is a round of three matches, with the first two games held on Wednesday and Thursday in Seoul. The final match is scheduled for Saturday in Lee’s hometown of Sinan-gun, South Jeolla Province.
In Game One, Lee started with the two of his stones on the board as a handicap given to an underdog. Even with the advantage, Lee was not sure about his victory. A month ago, Lee was quoted as saying: “I feel like I will lose the first game to HanDol. ... I want to play comfortably against HanDol as I have already retired, but I will do my best.”
A move by Lee took HanDol by surprise. NHN said HanDol could not see this move coming and its developers were confused because it was not a tricky one. “This was a move that you would definitely make if you’re a professional Go player. It is surprising that HanDol did not see it,” said Lee.
In Game Two, Lee was not given the kind of advantage that he had previously received. At his 40th stone, Lee made a mistake with the move on the upper left side of the board. He tried to make up for the mistake when he put his stones on the lower side. HanDol remained on the upper side and took four points.
Lee’s challenge against HanDol resonated with the historic game he had with AlphaGo, an AI player developed by Google’s DeepMind Technologies. He won just once in a five-match showdown. Yet, he was the first to beat AlphaGo.
Lee felt deeply frustrated when he lost the first three games to AlphaGo, according to Yonhap News Agency. In Game Two, Lee was taken aback by AlphaGo’s Move 37, as it was not a move that a human player would consider making. It was a moment to see what artificial intelligence could do.
"Frankly, I had sensed kind of a defeat even before the start of the matches against AlphaGo,” Lee said. “People from Google's DeepMind Technologies looked very confident from the beginning.”
In Game Four, however, Lee made a clever move, one that the machine apparently could not predict. His “White 78” is said to have been a “divine move.” Lee’s subsequent moves left AlphaGo confused.