A day before taking on the advanced human-like algorithm in the ancient board game Go, also known as “baduk” here, Lee said he now feels some pressure and the score he predicted last month — 5-0 or 4-1 in his favor after five matches — may not happen.
“I now think AlphaGo can imitate human intuition a little bit,” Lee said at Tuesday’s press conference in Seoul. “I should be a little nervous about the match.”
The historic man-versus-computer tournament in the ancient board game begins at 1 p.m. Wednesday at the Four Seasons Hotel in Seoul, with the winner’s prize of US$1 million at stake. The matches will be also held at the same venue on Thursday, Saturday and Sunday before closing out next Tuesday. If AlphaGo wins, the money will be donated to UNICEF and other charities.
Lee, who has won 18 international Go events, said he has lowered his expectations after listening to a presentation on AlphaGo’s algorithm by Demis Hassabis, the CEO of Google’s London-based AI company DeepMind. Though he didn’t fully understand how Google’s AI works, Lee now knows what the program is capable of.
“If the maximum number of human calculation on possible moves is 1,000, the computer is thinking millions and it is narrowing it down,” he said. “I thought it can be a threat.”
The 33-year-old South Korean, who went pro at the age of 12, said that the key to beating the machine will be reducing “human errors.” Despite feeling tension, Lee said that doesn’t mean that he will allow a loss.
“I’m still confident I can win,” he said. “First of all, I have never thought about getting beaten in the first match.”
Lee, who acquired Go’s highest level of ninth dan in 2003, said he has been preparing for the landmark match with image training about one or two hours a day. He has also stayed sharp, coming off from tough matches at the 17th Nongshim Cup in China last week where he had three consecutive wins before getting beaten by Chinese young star Ke Jie to hand over the trophy to the Chinese team.
“It will feel like I’m playing the game alone,” he said. “As the opponent is not human, I’m preparing for it differently.”
Although he openly discussed the possibility of losing to AlphaGo, Lee said that beauty of Go will continue even if he ultimately falls to machine.
“Since the computer plays Go without understanding the beauty of the game, the greatness of Go will continue,” he said. “In this match, I will try to protect that value of human.”
Go originated in China more than 2,500 years ago. It involves two players alternately putting black and white stones on a checkerboard-like grid of 19 lines by 19 lines. The object is to have larger territories than the opponent by surrounding vacant areas of the board using their stones. Go has been viewed as one of the hardest games for computers to master because of its infinite number of possibilities.
In this special match organized by Google and the Korea Baduk Association, the matches will be played under Chinese rules in which a player who places white stones receives compensation points of 7.5 at the end of the match. Scoring calculation will be also based on Chinese rules, also known as “area scoring” which a player’s score is determined by the number of stones that the player has on the board and territories earned. In “territory scoring,” which Koreans generally apply, a player’s score is calculated by the player’s territories plus the number of stones that the player has captured.
In this landmark showdown, each player will receive two hours per match with three lots of 60 seconds overtime counting after they have finished their allotted time. Google said the algorithm’s programmer will place the markers on the board after watching AlphaGo’s moves from a monitor.
AlphaGo surprised the Go community after shutting out European Go champion Fan Hui 5-0 last October. Hassabis said that its AI has been improved since then.
“We focused on improving algorithms, rather than using more hardware,” he said. “We made improvements that can challenge somewhat creativity against Lee.”
Hassabis emphasized that AlphaGo, which has won 494 of 495 matches against other computer opponents like Crazy Stone and Zen, is much more advanced than Deep Blue, a supercomputer from IBM that is best known for beating chess champion Garry Kasparov in 1997.
“Deep Blue was handcrafted by designers and it didn’t learn from itself,” he said. “I think the message we want to send is that how flexible and powerful that learning algorithms can be.”
The 39-year-old Briton added that AlphaGo’s strengths against Lee is that it doesn’t get tired and the AI will not feel intimidated by Lee though he admitted that the program will take more time if the South Korean makes a difficult movement. He said that the reason why they are having the match is to find any weaknesses on AlphaGo that they don’t know about.
“We have not experienced the ceiling level of Alpha Go and how far it can improve.”
Hassabis emphasized that Google wants to apply the technology to more than just playing a game, using the AI to help real-world problems in healthcare systems, robotics, household robots and smarter home management systems. He called it Artificial Global Intelligence (AGI), adding that it will not be a threat to humans in the future.
“We think the AI as a powerful tool,” he said. “Just like any other new powerful technology, it’s neutral but it depends how humans use it.”
Eric Schmidt, former Google CEO and current Alphabet executive chairman, said that he is also expecting much from this special Go match and the AI’s learning and experience will further help humans.
“Reinforcement learning that allows us to do something which I thought it would never happen: challenging the world’s best Go player,” he said. “It will be a great day for humanity. Advanced AI and machine learning will make humans smarter and a better world.”
The tournament will be livestreamed on DeepMind’s YouTube channel and will be broadcast throughout Asia on Korea’s Baduk TV.’