The second game of a million-dollar, man-vs.-machine Go showdown was a real nail-biter, but the outcome was a repeat of the first game: Google DeepMind’s AlphaGo artificial intelligence program vanquished Go champion Lee Sedol.
Today’s game in Seoul, South Korea, lasted almost four and a half hours. The battle went on so long that Lee ran out of regulation time and eventually was forced to make each of his moves in a minute or less. AlphaGo racked up an unassailable lead in points, and Lee resigned.
“Yesterday, I was surprised, but today, it’s more than that,” Lee said afterward at a news conference. “I’m quite speechless.”
Lee said that during the first game, AlphaGo may have made some questionable moves. In contrast, the program played a “near-perfect game” the second time around, he said.
DeepMind founder Demis Hassabis said AlphaGo’s playing style was more confident than it was the day before. “AlphaGo seemed to know what was happening,” he said.
Match commentator Michael Redmond, who is himself a top-rated Go player, said AlphaGo’s style of play had clearly improved since it beat European Go champion Fan Hui last October. “It’s playing much more aggressively,” he said during Google’s webcast. Redmond said the time pressure may have led Lee to miss the full significance of AlphaGo’s gutsy push through the center of the board.
#AlphaGo wins match 2, to take a 2-0 lead!! Hard for us to believe. AlphaGo played some beautiful creative moves in this game. Mega-tense…
— Demis Hassabis (@demishassabis) March 10, 2016
"AlphaGo used to be much more conservative in October. It's free now" -DM engineer
— American Go Assoc. (@theaga) March 10, 2016
This month’s five-game match is seen as a turning point for the field of artificial intelligence, and for the millennia-old game of Go.
The game looks simple, in that it merely involves placing white and black stones on a 19-by-19 board. But it’s considered much more complex than chess, due to the multiplicity of possible moves. Some experts had thought it would be a decade before software programs matched human champions in their ability to master the game.
AlphaGo’s developers surprised the experts by taking advantage of two strategies that gave the program humanlike learning capabilities: It was able to digest a huge database of Go game sequences, and then play itself millions of times to figure out how best to evaluate board positions.
Hassabis says the same strategies could be applied to produce smarter AI assistants for smartphones as well as researchers. “You can think of AI scientists, or AI-assisted science, working hand in hand with human expert scientists to help them in a complementary way,” he told GeekWire in January.
The next game will be webcast at 8 p.m. PT Friday. In order to win the $1 million prize, Lee would have to win all three of the remaining games in the series. If AlphaGo wins, the prize money will be donated to UNICEF, science education programs and Go societies.