After soundly beating the reigning European Go champion, Google's AI computer is looking to go head-to-head with one of the best players in the world in a match set to be held in South Korea in March.
In a study featured in the journal Nature, researchers from Google's London-based AI company DeepMind described how their AlphaGo program was able to win all five of its Go matches against the European champion Fan Hui.
Go, a board game that was invented in China some 2,500 years ago, involves having players alternately place white and black "stones" on a grid consisting of 19 vertical and 19 horizontal lines. The objective is to surround the stone pieces of the opponent without allowing a player's own pieces to be surrounded.
To become a master at Go, a player has to hone his or her skills at recognizing certain patterns that are formed when Go pieces are placed across the board. This can be done through continuous practice.
AlphaGo
Wanting to create an artificial intelligence capable of mastering Go, DeepMind researchers developed AlphaGo, a machine learning program that combines massive amounts of data sets with deep neural networks.
AlphaGo makes use of DeepMind's latest deep learning in combination with the Monte Carlo algorithm, which was designed to exhaustively identify large numbers of potential combinations of moves in Go. AlphaGo also received a considerable amount of input from master Go players.
When AlphaGo was matched against other Go programs, DeepMind's AI was able to achieve a 99.8 percent winning rate against its competition.
Match Against Fan Hui
In October, AlphaGo went up against Fan Hui, a 2-dan profession Go player, who is considered to be the best in Europe. The AI swept all five matches against Fan.
After the match, Fan said that playing Go with AlphaGo is like going up against a wall. Unlike human players who sometimes get tired or pressured to win a game, the AI was very strong and stable. This was a big difference, according to Fan.
"I know AlphaGo is a computer, but if no one told me, maybe I would think the player was a little strange, but a very strong player, a real person," Fan said.
Despite feeling bad about the loss, Fan said that he will study Go further and perhaps change the way he plays the game.
Toby Manning, an official from the British Go Association and the one who served as the referee during the AlphaGo-Fan match, said that it was difficult to tell which one was the human player and which one was the AI based on their moves.
He said that AlphaGo was better at managing its time compared to Fan. It also didn't appear to be as aggressive when making its moves as human players tend to become.
Manning said that some Go players would want to use the AI to work on their game.
DeepMind co-founder Demis Hassabis described Go as "the most complex and beautiful game ever devised by humans". By defeating Fan, he said that AlphaGo was able to overcome a long-standing challenge of artificial intelligence.
"AlphaGo is now going beyond - hopefully, eventually - what even the best humans in this area can do," Hassabis said.
"It's quite an amazing feeling to see what new things it's going to invent, within the constraints of the game of Go."
Upcoming Match Against Lee Sedol
DeepMind is now preparing for AlphaGo's Go match against the world champion Lee Sedol in March. Lee is a 9-dan professional player and has been the top Go master in the world for the past few decades.
"I haven't put any money on AlphaGo winning, but I do think we have a lot of reputation riding on this bet," DeepMind researcher David Silver said.
"So let's just say we'll be very disappointed if we lose the match in March."
