Mark another milestone in the rise of the machines: An artificial intelligence program pioneered by Google DeepMind has learned how to play the game of Go well enough to beat a human champion decisively in a fair match.
That’s a quantum leap for artificial intelligence: Go is looked upon as the “holy grail of AI research,” said Demis Hassabis, the senior author of a research paper on the project published today by the journal Nature.
The game seems simple enough, involving the placement of alternating black and white stones on a 19-by-19 grid. The object is merely to avoid having your stones hemmed in on four sides by your opponent’s stones. But Go, which originated in China thousands of years ago, is considered the world’s most complex game. “It has 10170 possible board positions, which is greater than the number of atoms in the universe,” Hassabis noted.
That means a computer program can’t best humans with the same kind of approach used for checkers and and chess. The programs for those games combine brute-force searches through the possible moves with a weighted evaluation of patterns in moves. But researchers at Google DeepMind say their software, known as AlphaGo, takes a different approach.
AlphaGo was programmed to sift through a database of expert Go moves, and then play against itself millions of times to improve its performance. Researchers called that part of the program the “policy network.” Another part of the program runs through Monte Carlo simulations to evaluate board positions.
The combination of the policy network and the value network results in a “much more humanlike” decision process, said David Silver, one of the study’s lead authors.
AlphaGo could beat the other top game-playing programs 99.8 percent of the time. Then, last October, the computer took on Fan Hui, who has won the European Go championship for the past three years. The human lost all five games.
“One couldn’t help but root for the poor human being getting beaten,” said Tanguy Chouard, a senior editor at Nature who watched the match.
It marked the first time that a Go computer program defeated a human professional player in a full game, played without giving the computer an advantage. But the biggest match is yet to come: In March, AlphaGo is scheduled to play a match in Seoul against Lee Sedol, who is widely acknowledged as the world’s best Go player.
“You can think of him as the Roger Federer of the Go world,” said Hassabis, referring to the Swiss tennis star.
No matter how that match goes, the London-based researchers at Google DeepMind say their approach is a winner. Most experts expected that it would have been 10 years or more before an AI program beat a professional Go player.
In what may be the AI world’s equivalent of trash talking, Silver said there was “zero chance” that AlphaGo could be beaten by IBM’s Watson computer, which is optimized to handle a wide array of knowledge, but at a shallower depth. He also said the current version of Facebook’s Go-playing program “would be obviously no challenge” for AlphaGo.
On the other side of the board, Facebook founder Mark Zuckerberg said his A.I. team was “getting close” to having a Go program that can beat humans:
Oren Etzioni, CEO of the Allen Institute for Artificial Intelligence, said AlphaGo represents “an outstanding technical achievement, and it demonstrates that when the goal is crystal clear, and the rules of the game are simple … computers will dominate.”
“At the same time, when the problem is ‘ill-defined,’ as in understanding a sentence, writing an article, or even comforting a friend – this is still way beyond our abilities,” Etzioni wrote in an email to GeekWire. AI programs also tend to struggle in areas where there is only limited data to work with, or where communicating or collaborating with people is involved, he said.
“The truth is that superhuman performance on a narrow task, like the game of chess, or Go, is only a small step towards building ‘general intelligence’ into a machine,” Etzioni said. “It’s still the case that our 4-year-olds are ‘smarter’ than any computer program, including AlphaGo.”
Google DeepMind’s Hassabis told GeekWire that the insights gained from AlphaGo will go toward improving more practical applications, starting with smartphone virtual assistants. The same kinds of deep neural networks created for playing Go could soon point you to better restaurants when you ask your Android phone for recommendations.
In the medium term, deep neural networks may play a higher-profile role in medical diagnostics. IBM’s Watson is already getting into that field.
“Longer-term, what I’m most excited about … is to use these types of general learning systems to help with science,” Hassabis said. “So 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 make faster breakthroughs in scientific endeavors.”
But there are some things the researchers won’t do with AI. When Google acquired DeepMind in 2014, the companies agreed “that the technology developed by DeepMind would never be used for military purposes,” Hassabis said.
He said an internal ethics board was appointed to make sure that the agreement is followed, and that DeepMind’s AI doesn’t get the upper hand over us humans. He also pointed out that he’s one of the signers of an open letter calling for a ban on autonomous weapons.
Silver and Hassabis are among the 20 authors of the Nature paper, titled “Mastering the Game of Go With Deep Neural Networks and Tree Search.” Google DeepMind’s Aja Huang shares lead authorship with Silver.