Subscribe
E-Paper
By Yoon Sung-won
Lee Se-dol will be able to exploit artificial intelligence (AI) AlphaGo's weaknesses with creative and unexpected moves in the final of the five-game match, today, according to analysts.
After scoring his first win following three straight losses in the Google DeepMind Challenge Match, Lee will be seeking revenge against AlphaGo in the historic final human versus machine game.
Professional go players and AI experts expect Lee will play more with his original aggressive, creative style.
"Lee has already learned that AlphaGo tends to make strange moves when it faces unfavorable conditions," said professional eighth-dan go player Kwon Gap-yong, who also has been a mentor to Lee. "Lee will have more of a chance of winning the final match if he makes key moves that are difficult to read."
Other go experts warned that the machine may not make a mistake in the final round.
"The chance of Lee winning is still below 50 percent considering AlphaGo's superb calculations though Lee found a strategy to beat the AI program in the fourth game," Korean ninth-dan professional go player Hong Min-pyo said.
Korea Baduk Association Vice President Park Chi-mun said, "Lee's task here will be to engage in a tussle without taking too many risks."
On Sunday, Lee clinched a long-awaited first win against the AI in the fourth game.
Lee forced a resignation from the program in 180 moves.
He made a key move in the 78th round of the game and led AlphaGo to commit to a "strange" position on the right edge of the board. Despite the program recovering its normal play mode and trying to make up the points it had lost, Lee succeeded in defending his advantage with safer moves.
Once the AI concluded that the probability of its loss was too high to overcome, it popped up a message that read, "AlphaGo resigns. The result ‘W+resign' was added to the game information." AlphaGo programmer Aja Huang, who placed the stones on the board on behalf of the machine throughout the competition, made the actual move of "throwing the stone" to resign.
After the game, Lee said he has discovered that the machine has two weaknesses.
"AlphaGo seems to have more difficulty playing with black than white," he said. "It also revealed some kind of bug when it faced unexpected positions."
DeepMind CEO Demis Hassabis said, "AlphaGo seems to have realized its mistake in the 87th move," admitting that its winning rate, which was about 70 percent in the 79th move, dropped below 50 percent in the 87th move.
He said it is difficult to tell AlphaGo's strategic move from a mistake because it is designed to make decisions that will lead to the highest win rate. Such an algorithm structure proved successful in the first three games when the machine maintained advantages throughout most of the match, but not when facing an unfavorable condition.
"We have designed AlphaGo to learn by itself by repeating self-plays," said David Silver, Google DeepMind's AlphaGo team leader. "But such an algorithm must have hulls in it that we cannot know about. Lee was able to push AlphaGo to expose such a weakness."
A Korea Baduk Association spokesman said, "AlphaGo, which had played like a top-class professional go player up to that point, abruptly changed into an entry-level player. Even when it could take a step back and head to a close match, it repeated preposterous moves."
Google and the association, which jointly organized the showdown, had previously agreed that the match will go through the fifth game even after the victory was settled with the first three wins by AlphaGo.
In the final match, Lee will play first with black stones, without flipping a coin or drawing straws, as the DeepMind CEO accepted his request to do so. Consequently, Lee will yield 7.5 compensation points to AlphaGo.
The game is scheduled to start at 1 p.m. today at the Four Seasons Hotel in central Seoul.
