Google DeepMind’s AlphaGo AI program may have won the $1 million five-game Go match with three straight wins, but Go champion Lee Sedol struck back with a consolation win today.
“Because I lost three matches, and I was able to get one single win, I think this one win is so valuable I would not trade it for anything in the world,” Lee said during a post-game news conference that was webcast from Seoul, South Korea.
Lee said he was driven on by the “cheers and encouragement” of his fans.
The Korean Go master is part of one of the most closely watched experiments in artificial intelligence since IBM’s Watson computer software took on two human champions in the “Jeopardy” TV quiz show in 2011. The past week’s match has also been compared to the duels between IBM’s Deep Blue computer and chess champion Garry Kasparov in the 1990s.
Go, a game that was created thousands of years ago in China, is considered more intellectually challenging than chess because of the multiplicity of moves that can be played on the 19-by-19 game board.
Google DeepMind took on the challenge of designing a Go-playing AI because it’s a game that can’t be solved by brute force. Instead, AlphaGo combines two methods of analysis, drawing upon a database of well-played Go games as well as Monte Carlo simulations to evaluate board positions. The two methods are known as the policy network and the value network.
AlphaGo’s methods served it well in the first three games, clinching the match. But in Game 4, Lee executed a brilliant series of moves in the middle of the game. It took the program several moves to register that it was in trouble, as explained by DeepMind founder Demis Hassabis in a series of tweets:
Lee Sedol is playing brilliantly! #AlphaGo thought it was doing well, but got confused on move 87. We are in trouble now…
— Demis Hassabis (@demishassabis) March 13, 2016
Mistake was on move 79, but #AlphaGo only came to that realisation on around move 87
— Demis Hassabis (@demishassabis) March 13, 2016
When I say 'thought' and 'realisation' I just mean the output of #AlphaGo value net. It was around 70% at move 79 and then dived on move 87
— Demis Hassabis (@demishassabis) March 13, 2016
Lee Sedol’s brilliant moves deliver a comeback against #AlphaGo in game 4 → https://t.co/MbtYm64lhL pic.twitter.com/NchPMDZTU6
— Google (@Google) March 13, 2016
AlphaGo eventually determined that it didn’t have a substantial chance of winning – and resigned after 180 moves.
Hassabis congratulated Lee on the win. “I think he has proved again today what an incredible player he is, and why he is such a legend,” he said at the news conference.
Google DeepMind plans to use the technology behind AlphaGo in a variety of applications, ranging from smartphone assistants to self-driving cars to medical diagnosis. But Hassabis said interactions with human experts were an essential part of finding the flaws in AI programs.
“This is why we came here, to test AlphaGo to its limits and find out what its weaknesses were, so we can try and improve the program,” he said. “We need a creative genius like Lee Sedol to be able to find out these issues and actually expose them.”
If Lee had won the best-of-five series, he would have received the $1 million prize put up by Google DeepMind. Instead, Hassabis has said the money will be donated to UNICEF, science education programs and Go organizations.
The final game of the series is due to be webcast at 8 p.m. PT Monday. The Seattle Go Center plans to stay open late for a watch party.