Lee Se-dol shows AlphaGo beatable
.jpg?w=728)
World go champion Lee Se-dol smiles at a news conference after he defeated AlphaGo for the first time in the fourth of a five-game match against the Google’s artificial intelligence program at the Four Seasons Hotel, Seoul, Sunday. / Yonhap
Go champion strikes back to score first victory against AI
By Yoon Sung-won
Korean professional go player Lee Se-dol redeemed himself by beating AlphaGo by a resignation in 180 moves in the fourth of a five-game match, Sunday, proving that the Google’s artificial intelligence (AI) is beatable.
Lee’s win, which ended his three-game losing streak, is considered as a distant hope as the AI has shown superior calculations that far exceed human capability.
His long-awaited win was hailed as a “victory for humans.”
After AlphaGo repeated several “strange” moves in the middle of the game, the world go champion managed to focus and retain his advantage to the end of the game without making any critical mistakes.
Lee expressed delight after the hard-fought game.
“This win cannot be more joyful because it came after three consecutive defeats,” he said. “It is the single priceless win that I will not exchange for anything.”
Lee said he has seized on the AI’s weaknesses in the game.
“AlphaGo seemed to feel more difficulties playing with black than white,” he said. “It also revealed some kind of bug when it faced unexpected positions.”
When asked about the asymmetry issue that Lee lacked information about AlphaGo, he said, “Basically, my lack of ability is the problem. I do not think it is a big problem.”
During the post-game interview, the go grandmaster asked the Google DeepMind team if he could play with black stones in the final round. According to Chinese go rules, which this competition follows, the player playing with black stones yields 7.5 compensation points to his opponent at the end of the game for playing first, and thus this considered as a disadvantage.
“As I won the game with white stones, I would like to play with black stones because winning with them will be more valuable,” Lee said.
DeepMind CEO Demis Hassabis accepted Lee’s request.
Self-destruction
Pointing out that the AI’s defeat was due to its mistakes, Hassabis said, “Lee was too strong for AlphaGo today. It estimated itself that it was doing well before the middle of the game but Lee’s incredible plays pressured it into make some mistakes.”
After Lee’s key move with the 79th stone, AlphaGo placed a stone in a strange position on the right edge of the board in the 87th move, which commentators said was “difficult to understand.”
“AlphaGo yielded its own territory more while allowing its opponent to expand his own,” said commentator Song Tae-gon, a Korean nine-dan professional go player. “This can be the starting point of AlphaGo’s self-destruction.”
The AI again placed a stone in the bottom left corner without reinforcing its territory in the center. Afterwards, it recovered its normal calculations, which Song said would be difficult for human players under such pressure, but Lee maintained his advantage by continuing to adopt safer moves.
Song said, “If this was a match between human players, any professional go player reviewing this game would call it an irreversible win for Lee.”
After losing the five-game Google DeepMind Challenge Match with his third defeat, Lee said, “I, Lee Se-dol, lost, but mankind did not,” adding that he would be able to play under less pressure in the remaining game.
He continued to be pushed for time in the fourth game, spending much more of his allocated time after repeating deep thinking over his moves.
Lee entered overtime counting after two hours and 53 minutes into the game. But he did not lose concentration which led to the win.
In the third game, Saturday, AlphaGo beat Lee again by a resignation in 176 moves as it did in the previous two rounds. The three consecutive wins gave the AI victory in the five-game competition as well as $1 million that Google offered as prize money.
Aiming at early domination, Lee started with an aggressive strategy. But AlphaGo successfully defended it and drove him to resign the game.
Previously Lee had anticipated a 5-0 or a 4-1 win. After the third defeat, however, he admitted he had misjudged the AI’s capabilities.
Lee added AlphaGo “clearly has weaknesses” even though it surprised him with the victories.
Demis Hassabis said AlphaGo wasn't specifically trained to face Lee and the way that its AI gets stronger is by playing more games against an “incredible player” like Lee.
“We are also very happy to test AlphaGo, push it to its limit and find out its weaknesses. We will go back to the U.K. to figure out what has happened,” he said.
Google co-founder and Alphabet president Sergey Brin also said he was very “impressed and excited” to have this match, adding that he will be looking forward to watching the last game on Tuesday.