杏吧原创

In the game of artificial intelligence, Google just upped the stakes. The firm鈥檚 UK-based AI subsidiary, DeepMind, has developed a program capable of beating a human professional player at the ancient Chinese game of Go. With this long-standing AI challenge felled, the company says its systems could one day help tackle complex real-world problems.

鈥淭he scientific world is going to be stunned by this result,鈥� says of the University of Alberta, Canada. 鈥淧eople weren鈥檛 expecting anything like this kind of a leap forward.鈥�

You might think those are strong words for such a simple pastime, but games have long been a benchmark for the performance of AI. When IBM鈥檚 Deep Blue beat chess champion Garry Kasparov in 1997, it was hailed as the first step of an AI revolution. That ultimately failed to materialise, though computers are now so good at chess that even the best humans can鈥檛 keep up.

The game of Go involves two players placing black and white counters to conquer territory. It is played on a 19 by 19 board, which allows for 10¹⁷¹ possible layouts, versus roughly 10⁵⁰ possible configurations on a standard 8 by 8 chess board. To give you a sense of scale, it鈥檚 estimated there are 10⁸⁰ atoms in the universe. 鈥淕o is probably the most complex game ever devised by man,鈥� says DeepMind founder Demis Hassabis.

Pruning the possibility tree

That鈥檚 why DeepMind鈥檚 AlphaGo program plays smart. The software uses multiple artificial neural networks, now ubiquitous in AI. Its 鈥減olicy鈥� network learns the game from a database of nearly 30 million moves made by expert human players, which teach the software to predict likely sequences of play. It then further improves by playing against versions of itself. That network hands off to a 鈥渧alue鈥� network that estimates the chance of winning given the current board position.

Both networks feed into a Monte Carlo tree search, a simulation that looks through the entire 鈥渢ree鈥� of possible moves in a game to determine which path offers the best chance of victory. The networks serve to trim down the tree, ruling out unhelpful or longshot moves and thus speeding up the Monte Carlo search. Existing computer Go players also use this Monte Carlo method, but AlphaGo鈥檚 pruning is what gives it the edge.

Naturally, this requires some beefy hardware. The DeepMind team developed two versions of AlphaGo, one running on a single machine with 48 CPUs and 8 GPUs, the other distributed across multiple machines with a total 1202 CPUs and 176 GPUs. In DeepMind鈥檚 test, the single-machine AlphaGo won 494 out of 495 games against other leading Go programs, while the distributed version won all of its games against others and beat the single-machine version in 77 per cent of games.

But the ultimate test is a human player, so DeepMind enlisted Fan Hui, the reigning European Go champion. Distributed AlphaGo beat Hui in all five games of a formal series. In five informal games, with shorter play time, Hui only won two. The company has arranged a showdown with world champion Lee Sedol, to take place in Seoul, Korea, in March, with a $1 million prize on the line 鈥� it will donate the money to charity if AlphaGo wins.

Facebook gets Go-ing

Computers have previously beaten humans given a starting handicap or playing on smaller boards, but AlphaGo is the first to pull ahead in a match on equal-footing. In post-game evaluations, Hui even said that the software was predicting betters moves than the ones he actually made. 鈥淲e felt we were in with a good chance of winning, but we didn鈥檛 know what was going to happen, it was very exciting,鈥� says Hassabis. 鈥淲hen we won five to nothing, that surprised us.鈥�

鈥淟ooking at the games, it was hard to tell who was black and who was white,鈥� says Jon Diamond, president of the and an early pioneer in computer Go. 鈥淚t was playing like a human being, there was nothing computer about it.鈥�

鈥淲ith the current architecture and scale of computation, I would have said this was inevitable in the long run, but I鈥檓 pretty impressed they鈥檝e made it at such fast pace,鈥� says Yuandong Tian, who is leading Facebook鈥檚 efforts to develop its own Go-playing AI, dubbed darkforest. 鈥淚 really congratulate them for their achievement.鈥� Today that its own system has made advances in beating strong amateur humans.

DeepMind says that AlphaGo鈥檚 approach is markedly different to Deep Blue鈥檚, as it evaluates thousands fewer positions on each turn compared with the chess-playing software. In other words, AlphaGo has a better sense of which out of the vast possibility of moves are promising, and spends its time focusing on those, rather than churning through options no human would consider.

Instant classic

That鈥檚 in line with a prediction made by New 杏吧原创 in 1965. We helped popularise Go in the UK with written by Irving Good, a mathematician who learnt the game from Alan Turing at Bletchley Park, where they worked as codebreakers during the second world war. 鈥淭he principles [of Go strategy] are more qualitative and mysterious than in chess, and depend more on judgment,鈥� he wrote 鈥淪o I think it will be even more difficult to programme [sic] a computer to play a reasonable game of Go than of chess.鈥�

It remains to be seen if AlphaGo can beat Sedol, proving machine dominance, but DeepMind is also eyeing up other challenges. 鈥淥ur hope is that one day these techniques can be extended to help us address some of society鈥檚 toughest and most pressing problems, from climate modelling to complex disease analysis,鈥� says Hassabis.

That鈥檚 pretty much what IBM said about Deep Blue, but cracking chess didn鈥檛 spark an AI revolution. Why will things be different this time around? Diamond says that AlphaGo鈥檚 learning method makes it more of a generalist, which could prove key. 鈥淚t seems to have very little Go knowledge built into it,鈥� he says. 鈥淭he basic techniques it uses are much more applicable to other areas.鈥�

Facebook is also hoping its Go-bot will help develop AI. 鈥淒eep Blue used dedicated software for evaluating chess. For the Go research, neither Google or Facebook has used any dedicated software,鈥� says Tian. 鈥淎ll the technology is pretty general, and can do other things as well.鈥�

Whether or not Go does prove to be the route to intelligent machines, Schaeffer thinks DeepMind鈥檚 work is one for the history books. 鈥淚t鈥檚 a tour de force, it鈥檚 one of those papers that takes a field and turns it upside down,鈥� he says. 鈥淚t will immediately be recognised as a classic.鈥�

Journal reference: Nature,