Google’s AI Is About to Battle a Go Champion—But This Is No Game
Today, inside the towering glass and steel Four Seasons Hotel in downtown Seoul, South Korea, Google will put the future of artificial intelligence to the test. At one o’clock in the afternoon local time, a digital Google creation will challenge one of the world’s top players at the game of Go, the ancient Eastern pastime that’s often compared to chess—though it’s exponentially more complex. This Google machine is called AlphaGo, and to win, it must mimic not just the analytical skills of a human, but at least a bit of human intuition.
Over the years, machines have topped the best humans as checkers, chess, Othello, Scrabble, Jeopardy!, and so many other contests of human intellect. But they haven’t beat the very best at Go. As Google likes to point out, there are more possible positions on a Go board than atoms in the universe—more than even the most powerful computers can contemplate. The scope of the game is so enormous that top human players must rely on more than careful analysis to succeed. They play based on what the board looks like, how it feels. To beat these humans, a machine must, in some way, reproduce this magic.
Over the past eighteen months, a team of researchers at a Google AI lab in London have worked to build an artificially intelligent system that can make this kind of leap, and AlphaGo has already shown its worth. In October, during a closed-door match, it beat the three-time European Go champion, Fan Hui. But now comes the bigger test. Today at the Four Seasons, AlphaGo begins a five-game, seven-day, one-million-dollar match against the Korea-born Lee Sedol, who has won more international Go titles than all but one other player. Google bills it as a battle with “the Roger Federer of the Go world.”
For both hardcore techies and the obsessive community that surrounds Go in Korea and across Asia, the match—which Google will stream live on YouTube—is high entertainment. “The interest in this match is huge,” says Hajin Lee, a professional Korean Go player who helped organize the match. “It’s unprecedented.” But given the technologies that underpin AlphaGo—and the extreme complexity of the ancient Eastern game—this contest is also an opportunity to test the progress of modern AI, to measure its potential not just to win a game but to rapidly reinvent everything from Internet search engines and digital assistants to robotics and scientific research.
WIRED is on the ground at the Four Seasons through the final game next Tuesday, delivering regular dispatches on the match and all the activity—both human and digital—surrounding it. In South Korea, the contest is no niche pastime. Of the country’s 50 million citizens, an estimated 8 million play Go. “Lee is kind of a representative of Korea,” says May Jang, a Korean journalist. “So people even who don’t know Go have heard his name.” But the match is an even bigger deal in the world of technology—whether the public realizes it or not. Go players are rooting for Lee Sedol, Hajin Lee says. But the close-knit world of artificial intelligence research is clearly on the side of Google.
Some call this a replay of the 1997 chess match between IBM’s Deep Blue supercomputer and world champion Gary Kasparov—or the 2011 Jeopardy! match between IBM Watson and trivia-happy humans Brad Rutter and Ken Jennings. Those matches also tested the power of AI. But AlphaGo versus Lee Sedol is a little different.
Part of it is that the game of Go is different. As Hajin Lee will tell you, Go grandmasters do play by intuition. They don’t necessarily play the game like a chess grandmaster, methodically examining the possible outcomes of each possible move. But the larger point is that the technologies at the heart of Google’s artificially intelligent machines are by no means limited to Go. Unlike the earlier AI contests, which were more proof-of-concept exhibitions, the systems behind AlphaGo have already leapfrogged so many other technologies in the marketplace, proving that they can, say, recognize images or identify spoken words at a level that was previously impossible. With these technologies, machines can autonomously learn tasks at speed and a level that just wasn’t possible in years past. And that means they could significantly accelerate the movement towards so many other forms of artificial intelligence.
Improving on Its Own
As recently as two years ago, even Rémi Coulom—the guy who built the best computer Go player at the time—assumed that at least another decade would pass before a machine beat the top humans at the ancient game. But then Demis Hassabis, David Silver, and other researchers inside DeepMind—a London startup that Google acquired in early 2014—tackled the problem with help from AI techniques known as deep learning and reinforcement learning. The result was AlphaGo. And after less than two years of development, it topped three-time European Go champion Fan Hui in a five-game match, winning all five games.
The win surprised just about everyone in the AI world. A week before Google revealed its victory, Yann LeCun, the head of AI at Facebook, expressed doubt that Google had beaten a grandmaster. And some still doubt that AlphaGo will beat Lee Sedol. Fan Hui is ranked 633rd in the world, while Lee Sedol is ranked 5th. But in a speech last month, Hassabis made a point of saying that AlphaGo continues to learn. “They give us a less than 5 percent chance of winning,” he said of the world’s Go players. “But what they don’t realize is how much our system has improved. … It’s improving while I’m talking with you.” This ability for the machine to so quickly learn on its own is what makes this week’s match so intriguing.
Hassabis and crew bootstrapped the system using deep neural networks—networks of hardware and software that approximate the web of neurons in the human brain. Basically, deep neural nets learn to perform tasks by analyzing large amounts of digital data. If you feed enough photos of a flamingo in to a neural net, it can learn to identify a flamingo. If you feed it enough human dialogue, it can learn to converse (kinda) like a human. And if you feed it enough Go moves from the world’s grandmasters, it can learn to play Go—and play it well.
But that was merely a start. After using neural nets to build a system that could play Go, DeepMind matched this system against itself. In playing itself and tracking which moves are most successful, the system can improve its skills even more. This is called reinforcement learning. The result was a system that could beat the European Go champion. And as Hassabis points out, in the months since, this system has only improved. Humans like Hassabis are helping it improve, tweaking the code here and there. But AlphaGo is also improving on its own.
Sure, AlphaGo relies on more than just machine learning. It still leans on a technology called Monte Carlo tree search, the same technique of trying to calculate all possible future outcomes that Rémi Coulom used in building the previous digital Go champion. But it’s those ascendant forms of machine learning—methods that go beyond the brute force of calculating all possible outcomes—that make AlphaGo so powerful. Go is so complex that the tree search can’t come anywhere close to analyzing all the possible moves. But with its machine learning techniques, AlphaGo can limit the possible outcomes, making tree search much more effective. “It’s narrowing the beam of probable actions so it doesn’t have to calculate the rest,” says Chris Nicholson, the founder of a deep learning startup called Skymind. “The rest was too much.”
Pushing Past Go
Yes, Go is just a game. But Hassabis says these same techniques can reinvent robotics—that machines can learn real world tasks in much the same way they learn to make moves in a game. He also sees them as a path to a new kind of scientific research, where machines learn to identify promising areas of research and push human scientists in the right direction.
These claims carry some serious weight because deep learning has already proven quite powerful in at least some real world situations, including image recognition and speech recognition. And its progress is already promising when it comes to understanding the natural language that we humans use and, indeed, giving robots the power to learn on the job. At the University of California, Berkeley, robots use neural nets to learn how to screw bottle caps onto bottles.
Nonetheless, Oren Etzioni, a professor of computer science at the University of Washington and the executive director of the Allen Institute for Artificial Intelligence, says we should remain a little skeptical. After all, the real world is far more complex than Go. “[Go] is still a game—an artificial environment with discrete moves. And at the end, you know who won and who lost. On the other hand, take Libya,” he says , referring to the American intervention that brought down Muammar Gaddafi but left a country torn by civil war. “Did we win or did we lose?”
He also points out that in beating Gary Kasparov, Deep Blue relied on machine learning. And Deep Blue didn’t exactly deliver sentient robots. But then he acknowledges this was machine learning of a much less powerful variety.
The winner of this week’s match receives a million-dollar prize, but Etzioni believes that the real stakes lie elsewhere. “The million-dollar question is: ‘Who’s gonna win?’” he says. “But the billion-dollar question—or maybe even the trillion-dollar question—is: ‘How do we build AI systems for fuzzy situations that are less artificial than a board game?’”
Soon, we’ll at least have an answer to the million-dollar question. What will that answer be? There’s no consensus. But many AI experts believe AlphaGo will triumph. “I wouldn’t bet against Demis Hassabis,” Nicholson says. And, well, Etzioni says much the same thing. “I’m betting on the humans behind AlphaGo,” he says. “People see this as machines versus Lee Sedol. But I see this as humans using technology versus Lee’s brilliance.”
AlphaGo may not win this week, but if it doesn’t, it will win soon enough. As Hassabis says, it’s always learning.