Go to Contents Go to Navigation

(News Focus) S. Korean Go champion proves Google AI can make mistakes

All Headlines 23:39 March 13, 2016

By Joo Kyung-don

SEOUL, March 13 (Yonhap) -- When South Korean Go player Lee Se-dol picked up his first victory against artificial intelligence (AI) program AlphaGo Sunday, there was applause even from the Google programmers who designed the self-learning program.

For them, they were thankful that Lee won because the 33-year-old South Korean actually pointed out what AlphaGo is capable of and what are its shortcomings.

"AlphaGo may have possible holes (in its program), but we were not strong players to find these holes," said David Silver, AlphaGo project team leader at Google's London-based firm DeepMind. "AlphaGo was pushed to its threshold and limit."

Lee's victory on Sunday was surely a victory for mankind, but at the same time it showed that the machine can make mistakes, too.

In previous matches, Lee had tried almost every possible strategy that he could come up with to beat AlphaGo.

For the opening match, Lee, who is known for his aggressive and unconventional style of playing the ancient board game, tried to tweak the game and lure AlphaGo into his traps, but was forced to cede in just 186 moves.

In the second match, Lee went opposite of his usual style, making his sequences calmly which ninth-dan player Yoo Chang-hyuk, who was commentating the match, described as "playing like a different person." Lee's strategy pushed AlphaGo into overtime, but he in the end was forced to yield again after 211 moves.

In the third round, Lee came up with strategy to shake AlphaGo's calculations by creating various "ko" or repetitive capture situations, but this didn't work either, resulting in him dropping the game in 176 moves.

In their fourth match, Lee wasn't testing AlphaGo from the start, though he was the first one to tweak the game with the 12th move. While alternating battles in the left and right center, AlphaGo was leading till midway through the game when at one point, South Korean commentators said the AI was 10 to 15 points ahead of Lee.

Making things worse, Lee, a ninth dan-player who went pro at the age of 12, made a critical mistake in the 70th move which Song Tae-gon, a ninth-dan player who commentated the match, described as "a move that could initiate Lee's defeat." In addition, he ran out his allotted time of two hours and entered 60-second overtime use.

At that point, Lee was still trying to break AlphaGo's presumed territories in the center with his white stones. While calculations were running for both sides, the 78th move from Lee turned out to be breakaway he needed in the center and AlphaGo seemed to lose control of the game as its next move wasn't standard.

Then from its 85th, 87th and 89th moves, AlphaGo placed stones in moves considered "suicidal" which let Lee to secure bigger territories, while ruining AlphaGo's territories.

"This is a really a (program) bug," Ha Ho-jung, a fourth-dan player who was commentating the match, said jokingly after watching the moves. "Did some South Koreans invade AlphaGo to force the errors?"

Though AlphaGo has surprised human Go players with mysterious moves that turned out to be correct in the end, this time, it was clear the program made a mistake. DeepMind CEO Demis Hassabis also noticed there was problem in AlphaGo that time.

He wrote on his Twitter that AlphaGo's chances of winning were around 70 percent after the 79th move, but then it "dived" after the 87th.

"He (Lee) was too good for us today and pressured AlphaGo into a mistake that it couldn't recover from," Hassbis tweeted.

The several strange moves at the end suddenly turned the game in Lee's favor. As the game went on, Lee was pressured by time and AlphaGo was slowly catching up in points, but it was too late to reverse the result.

The AI then displayed a message on the screen that read "AlphaGo resigns. The result 'W+ Resign' was added to the game of information," which made Aja Huang, a Google DeepMind programmer who was placing stones on the board for AlphaGo, to call game over.

"AlphaGo was breaking down itself during that fight," said match commentator Song. "It has been placing very odd moves consecutively."

Hassabis admitted that Lee's "fantastic moves" forced AlphaGo to make mistakes. When asked if AlphaGo's mistake could translate into larger system failures when the AI is used in other areas such as medical and home management systems, the 39-year-old Briton said that would not happen.

"AlphaGo is still prototype," he said. "It would be tested and strengthened before it is deployed in other applications such as healthcare."

Though AlphaGo did make mistakes, for Lee, it wasn't still easy to win. He said the 78th stone, which sparked Lee's revival in the center, was a move he had to place without a choice.

"I thought I could lose again because it was difficult to ruin it (AlphaGo's effort to secure territories) than I thought," he said. "At that point, I had no choice but to go for that move, but I'm confused that I'm getting praised by placing that move."

Lee has already lost the five-round Go tournament after suffering three straight defeats, but Sunday's victory will remain a memorable feat for the Go player, who has won 47 events, including 18 international titles, in his career.

"It's a priceless victory," he stressed.

This was also a precious victory for the South Korean Go community.

"Lee continuously tried to find AlphaGo's weaknesses," Korea Baduk Association (KBA) secretary-general Yang Jae-ho, who is also a ninth-dan Go player, told Yonhap News Agency Sunday. "He found out that it (AlphaGo) can make mistakes in large areas in complex situations."

Fellow South Korean Go players also cheered for the victory as Lee prepares to take on AlphaGo for the last round on Tuesday.

"As a professional player and a South Korean, I'm really happy for Lee," said Lee Da-hye, a fourth-dan player who will be the referee for the final match.

kdon@yna.co.kr
(END)

HOME TOP
Send Feedback
How can we improve?
Thanks for your feedback!