Artificial intelligence is everywhere these days even if we don’t recognize it. You can thank, or blame, AI for the suggestions Amazon gives you on what to buy next, or Google’s habit of finishing your sentences when searching. However, it’s not the sudden shift in search engine optimization that has artificial intelligence researchers celebrating, but the unexpected victory of an AI system at a game of Go against a professional human player.
While the victory itself is shocking, what is more exciting is how the artificial intelligence system managed to win. According to a report released on Jan 27 in the journal Nature, the system, called AlphaGo, won by teaching itself how to play the game through a process called “deep learning.”
Board games have long been the testing grounds for artificial intelligence and AI earned their place in the world of games when IBM’s Deep Blue defeated Gary Kasparov at chess in 1997. Since then, artificial intelligence has proved itself in games ranging from Jeopardy! to Texas Hold’em. But Go has long stood as the most difficult challenge for artificial intelligence programmers due to its simple rules masking the extraordinary complexity of the game.
In the game of Go, players place white or black stones on a large gridded board and try to capture the largest amount of territory. Pieces are captured when they are fully surrounded by an opponent’s stones and are ‘alive’ when they are still touching open areas on the board. The extreme simplicity of the rules does not make the game simple, and many of the best players have spent most of their lives improving their skill at the game.
“It’s probably the most complex game devised by humans,” said study co-author Demis Hassabis, a computer scientist at Google DeepMind in London, at a news conference. “It has 10 to the power 170 possible board positions, which is greater than the number of atoms in the universe.”
Live Science reported that Hassabis and his co-workers differed in their approach to designing an artificial intelligence system by using deep learning to teach the system Go rather than instructing the AI to use a specific tactical pattern or series of moves. With deep learning, the AI uses two sets of neural networks called the value network and the policy network. The value network allows the AI to ‘read’ the board and decide which player is winning, a difficult task in Go, and the policy network chooses the moves to make.
“Our search looks ahead by playing the game many times over in its imagination,” said study co-author David Silver, a computer scientist at Google DeepMind who helped build AlphaGo, at the news conference. “This makes AlphaGo search much more humanlike than previous approaches.”
By all accounts the “more humanlike” approach appears to be working. AlphaGo has defeated rival AI systems about 99.8 percent of the time, and its recent victory against Fan Hui, the reigning European Go champion, has secured the AI systems dominance at the game. It didn’t just beat Fan Hui either, it defeated him in all five games they played.
Now the researchers behind AlphaGo are setting up the next challenge for the system, this time it will play Lee Sedol, one of the world’s best Go players. “You can think of him as the Roger Federer of the Go world,” said Hassabis.