A computer just beat a champion of the complex strategy game Go, a feat that may have enormous implications for artificial intelligence (AI) research. Go isn't a particularly popular or well-known game in the west, but it is popular worldwide, where it is played by about 40 million. There are Go tournaments that are held, at which regional and world champs are crowned. One such celebrated player, European Champ Fan Hui, just had his hat handed to him (five games to zip) by AlphaGo, a computer-based Go-playing AI built by Google's DeepMind. The results of DeepMind's work and the no-holds-barred tournament are detailed in a new study published Wednesday in the journal Nature. "It's a historical milestone in artificial intelligence," said Nature senior editor Tanguy Chouard, who was present at the tournament between Hui and the AlphaGo program. Go is often described as the "Chinese version of Chess," but that description barely does the deceivingly simplistic game justice. The object of the game is to have majority control of the board. You do so by placing your white (or black) pieces (stones) on the board and using them to surround your opponent's pieces so that they are forced to remove them. If it sounds less complicated than chess, it's not. To put things in perspective, for each move in chess you have about 40 options. Each move on the 19-by-19 Go grid affords you 200 choices. "There are more configurations on the board than there are atoms in the universe," explained Google DeepMind researcher David Silver in a Nature video describing his group's achievement. This is, by some measures, as significant a moment in AI as when IBM's Deep Blue beat chess champion Gary Kasparov in 1997. Perhaps more importantly, this breakthrough is arriving ahead of schedule. Two years ago, a Wired report[1] said it might take another "10 years or so" before anyone could build a computer program capable of beating a human Go champion. "Go is the pinnacle of AI research," Google DeepMind CEO Demis Hassabis told Mashable, adding that it's stood as an "irresistible challenge." In fact, the challenge is so irresistible that Facebook, one of Google's main Silicon Valley rivals, claimed on Wednesday morning that his company has neared its own Go achievement: an AI capable of making strong Go tournament-level moves in as little as 0.1 seconds. Facebook's announcement[2], made on CEO Mark Zuckerberg's Facebook page, may have been timed to blunt the impact of the Google-related news, considering they were published in one of the most prominent science journals worldwide. Just as there were chess-playing programs long before Deep Blue, there have been many Go programs, but none were able to beat leading human players without the application of some handicaps, which are measured in game pieces. "The game of Go is intractable to brute force search," explained DeepMind's Silver. This is in contrast to chess which, with its named pieces, piece values and grid of 64 squares, is almost perfect for a powerful computer running millions of move possibilities in seconds. Mastering Go required something more than the techniques Deep Blue used to beat Chess champions. A year ago, DeepMind's AI learned how to play and win at the classic video game Breakout as an audience watched. Another Nature study revealed that DeepMind had actually mastered a number of classic Atari Console games[3] including Fishing Derby, Freeway and Robot Tank. The not-so-secret sauce to DeepMind's gaming skills is machine learning. For Go, DeepMind would once again apply machine learning, but this time with not one, but two neural networks called "Policy" and "Value." Both look at Go's myriad game play possibilities, but in two quite specific ways. Policy narrows the field of possible moves to a handful of promising ones, while Value looks for positive outcomes without driving all the way to every possible game conclusion. Silver said the Policy network looks at some 30 million games by human Go experts to accurately predict moves up to 57% of the time. The previous record was 44%. AlphaGo essentially plays millions of games between its two neural networks and learns how to be a better Go player through trial and error and reinforcement learning, said Silver. Beating a human Go champ "was initially thought that this was too hard for AI programs," wrote Oren Etzioni, CEO of the Allen Institute for Artificial Intelligence, in an email to Mashable. Etzioni called DeepMind's solution, a combination of deep learning and reinforcement learning, "a strong technical contribution" to the field of AI research. He was not involved in the new study. Etzoni, who does play Go, agrees that AI researchers are making rapid progress "on narrow tasks (like Go) but [are going] very slow on broad tasks (like natural language understanding)." Initially, DeepMind pitted AlphaGo, running on a standard PC, against the leading Go programs, where it won all but one of 500 matches. Then it was time to play three-time European Go champ Fui. AlphaGo won every game. For that match, though, DeepMind upped the hardware ante considerably, putting AlphaGo on a distributed system with hundreds of CPUs. The five games Go European Champ Fan Hui lost to AlphaGo. During the match Hui apparently made several comments that sound like he was associating human characteristics with the AI. Image: Nature "It was one of the most exciting moments in my career," said Nature's Chouard, who recalled the cheering from the programmers upstairs and the people near the defeated champ Fui. "One could not help but root for [a] poor human being being beaten... It was chilling to watch." Hassabis said DeepMind is well aware of the ethical issues surrounding AI. "With any new powerful technology… you have to think very carefully about how to ethically use it and responsibly deploy it. You have to make sure those benefits accrue to the many as opposed to the few," he said. However, he also told Mashable, "We're still talking about a game here, a fabulously complex game." The AlphaGo AI is, ultimately more suitable, said Hassabis, to solving computer-based problems than those in the real world. That said, a triumph in AI is rarely just about winning a game or beating a human opponent. DeepMind, like everyone else in the field, is taking the long view and has big dreams for a system like AphaGo. This kind of AI is "applicable to any problem where you have a large amount of data where you need to find insights and structures, long term plans and decisions to figure out what to do next to reach some sort of goal," said DeepMind's Silver. During the near future, AlphaGo's smarts could end up in something like a smartphone assistant. Since DeepMind is owned by Google, perhaps we should eventually expect Google Now to become very adept at playing Go. Further down the road, DeepMind hopes to apply the AI to medical diagnostics and, more broadly, big science questions like climate change. Up next, though, is AlphaGo's latest human Go pro challenge. His name is Lee Sedol, and he is perhaps Go's greatest player ever (Silver called the South Korean "the Roger Federer of the Go world"). Sedol will play AlphaGo in March. Just as Deep Blue's besting of Kasparov did not mark the end of humans playing chess and even beating computers, Hassabis doesn't expect AlphaGo's success to mark the beginning of the end for Go. "It may show some new levels of Go that are attainable," Hassabis told Mashable. There are, he noted, few strong Western Go players, but with access to Google DeepMind's AlphaGo, they may be able to become world class players without moving to Asia. Have something to add to this story? Share it in the comments.
To get Go-ing
Winning and wondering
The next move
References
- ^ a Wired report (www.wired.com)
- ^ announcement (mashable.com)
- ^ classic Atari Console games (mashable.com)
