Computer says Go

IN 1996 IBM challenged Garry Kasparov to a game of chess against one of its computers, Deep Blue. Mr Kasparov, regarded as one of the best-ever players, won—but Deep Blue won the rematch. Two decades on, computers are much better than humans at chess but remain amateurs when it comes to the much tougher, ancient game of Go. Or at least, they did. Now a computer has managed to thrash a top-drawer human player.

The computer used a program, called AlphaGo, developed by DeepMind, a London-based artificial intelligence (AI) company bought by Google in 2014 for $400m. It took on Fan Hui, European Go champion, beating him 5-0, according to a report in Nature. Beating a champion at Go has long been considered a “grand challenge” in AI research, for the game is far harder for computers than chess. Go players alternately place black or white stones on a grid of 19x19 squares with the aim of occupying the most territory. The size of the board, and the number and complexity of potential moves, make the game impossible to play via brute-force calculation. Demis Hassabis, DeepMind’s founder and one of the paper’s authors, reckons that a typical Go turn offers around 200 legal moves, compared with just 20 or so in chess.

Whereas a chess-playing computer like Deep Blue was programmed directly by humans, AlphaGo used AI to teach itself about how to play Go and then make its own decisions. This was done with a technique called machine learning, which allows computers to figure out for themselves how to do things, such as to recognise faces, respond to speech and even translate between languages.

AlphaGo works in two parts. When it is the computer’s turn, the program first suggests moves based on the sorts of general tactics that human players have used in the past—much as Deep Blue would. Then the second part of the system sifts those moves for those that look like they might lead to a win, again based on patterns it has picked up through memorising zillions of previous games.

The ultimate test of AlphaGo’s capabilities, though, will come in March. DeepMind has persuaded Lee Sedol, a Korean player widely regarded—like Mr Kasparov in his day—as one of the best-ever players, to take on their machine in a series of games in Seoul. If AlphaGo wins—and given its performance against Mr Hui, that seems like a distinct possibility—then human brains, and their possessors, will have to cede another defeat to the machines.