Human champ loses to AlphaGo

World go champion Lee Se-dol, right, presents a go board with his autograph to DeepMind CEO Demis Hassabis after the final game of the Google DeepMind Challenge Match at the Four Seasons Hotel in central Seoul, Tuesday. / Korea Times photo by Hong In-ki

Lee says mental element key factor behind 4-1 defeat

By Yoon Sung-won

World go champion Lee Se-dol lost the final game in a five-match series to the artificial intelligence (AI) AlphaGo program, Tuesday, by resigning after 280 moves. The historic series ended 4 to 1 in favor of the AI.

In the five-hour game, the longest yet, Lee struggled to break a tie after he failed to overcome the 7.5 compensation points that he yielded to AlphaGo by playing first with black stones.

“I regret that the match is over and sorry for failing to bring it to a successful conclusion,” Lee said after the game. “I started thinking I had the advantage in the early stages. But I was defeated and I think it revealed my shortcomings.”

Lee said the mental element was one of the key factors behind his defeat, even though the program was not that much superior in terms of skill.

“Basically it is not human, never shaken psychologically and can concentrate endlessly. It is true that I needed time to adapt to such an unaccustomed playing style,” he said. “Even if I face it again, I doubt if I can win.”

The world go champion added, “AlphaGo’s skills made me doubt if what we have known about go is truly right. I will look into it in the future.”

He said he will make greater efforts to improve his skill to meet the expectations of those who have supported him.

Korean ninth-dan professional go player Kim Seong-ryong, who was a commentator for the match, said, “Lee faced AlphaGo with calculations, which we know computers are best at. I doubted if he could win, but at the same time, I loved it.”

It was a neck-and-neck game with only about 1.5 points difference when the scores were counted, according to the commentator.

Lee played first with black stones, something he requested after the fourth game.

During the post-game interview Sunday, he asked the DeepMind AlphaGo team if he could play with black stones in the fifth game without drawing lots.

He said winning with black stones, considered as a disadvantage because of the 7.5 compensation points that the player yields to the opponent for playing first would be more valuable. DeepMind CEO Demis Hassabis accepted the request.

AlphaGo’s next move undecided

While several other go players including the Chinese professional ninth dan Ke Jie have been referred as the next opponents of AlphaGo, the DeepMind CEO said he has not decided on any plans.

“We have concentrated on the match with Lee so far and have not established firm and detailed plans for the future,” Hassabis said. “We have learned how to improve AlphaGo. Once back in the U.K., we will analyze what we have observed for the next few weeks and decide whether to do more matches.”

He also said the AI technology has more room for improvement.

“AI technology has both opportunities and problems,” he said. “I would like to stress that we still remain in the early stage of AI development. At DeepMind, we consider AI as one of the powerful tools that can help humans do things better. There will be various interesting areas such as healthcare and science research that the technology can be applied to.”

He also pledged to research and use AI technologies in a proper manner.

“DeepMind will consider ethical ways of developing AI; discuss with others at conferences; and share what we have researched through science journals,” the CEO said. “We think openness and cooperation is important here and AlphaGo should benefit the majority not only a limited number of people.”

Before the final game, many professional go players expected Lee to adopt unorthodox, unpredictable plays called the “cosmic style,” aiming at exploiting AlphaGo’s weaknesses as he did in the fourth game.

Lee started on the top right corner of the board and gained points from a knight’s enclosure, a practical move to secure as many points as possible in the early stages. The AI responded by putting stones on the flower points of the left edge.

Hassabis wrote on his twitter account, “AlphaGo made a bad mistake early in the game. It didn’t know a known tesuji. But now it is trying hard to claw back.” Tesuji, or “knack,” refers to a tactical and clever move that leads to sequences that no other move can accomplish.

Lee secured extensive territory in the bottom right corner of the board about 60 minutes into the game.

Twenty minutes later, Lee turned to the top right corner and continued to secure territory in corners while AlphaGo concentrated on the central part of the board.

Around 132 minutes into the game, AlphaGo made aggressive moves to catch up with the go champion; but Lee chose to remain passive. This allowed the AI to establish extensive territory from the right edge to the center.

Lee used up his two hours of allocated time about 220 minutes into the game, and entered overtime counting. Compared to the previous round, Lee seemed more comfortable managing his allocated time while AlphaGo used more time than before.

After the match, Korea Baduk Association presented AlphaGo a certificate of honorary ninth dan and Lee gave Hassabis the go board with his autograph as a gift.

International Go Federation Chairman Hong Seok-hyun said, “The event served as an opportunity to promote the game of go as an essence of eastern culture and its beauty.”