In the latest battle between human and machine, Martin Mueller can perhaps be forgiven for his divided loyalties.
A professor of computer of science at the University of Alberta and an avid player of the ancient game of Go, Dr. Mueller has been following every move of the epic matchup that ended on Tuesday and saw, for the first time, a computer program beat the world’s best player at what has often been described as the world’s most complex game.
Dr. Mueller was not one of those musing about the future of the human race as AlphaGo, the program developed by Britain-based Google DeepMind, notched up its final win, taking four of five games against South Korea’s Lee Sedol and claiming a $1-million (U.S.) cash prize in the process.
For him, the more pertinent question concerns the future of an entire field of research now that one of the holy grails of artificial intelligence has been claimed at last – notably, by a group that was backed not by a university but by one of the world’s best-resourced technology companies.
“Are the most interesting problems in this area now only attackable if you have a big infrastructure behind you?” Dr. Mueller wondered after staying up most of Monday night, along with a small contingent of other department members and interested spectators, to take in a live webcast of Game 5 from Seoul.
The group has been gathering at the university’s computer science department for the past week to tune in to YouTube for the games, along with more than one million other viewers around the world.
But in Edmonton, interest was more than academic, Dr. Mueller said. One of AlphaGo’s lead authors is David Silver, who earned his PhD under Dr. Mueller and Rich Sutton, another faculty member. A second author, Aja Huang, did post-doctoral research in the same department. For the Alberta group, that made the experience a bit like watching a pair of hometown stars hoisting the Stanley Cup after a dramatic playoff series.
And in this case, it’s clear that the new Gretzkys of Go have pulled off something unprecedented.
“It’s obviously a milestone,” said Dr. Sutton, a pioneer in a form of computer programming known as reinforcement learning, which uses a trial-and-error method to improve performance. Reinforcement learning is one of the key ingredients behind AlphaGo’s success. It was used to hone the program, which is built like a layered network with millions of connections. Together, they can work somewhat like neurons in the human brain to evaluate the state of a game and choose the most promising next moves.
Those moves cannot be arrived at by brute-force calculation, because with more possible board positions than atoms in the universe, the game of Go is simply too massive a problem for a computer to tackle that way.
Playing on a 19-by-19 grid with black and white markers, the two opponents in a game of Go are bound by a simple set of rules as they try to surround the most territory by the end of the game. But the many ways in which a game can unfold mean that superior players are those who best understand which moves are most likely to prove beneficial in the long run, a subtle skill that has thwarted computer programs for years – much longer than it took to develop a machine that could beat the world’s top-ranked chess player, which was accomplished in 1997.
AlphaGo has managed to be the first to cross the threshold thanks to its double-barrelled approach that involves both a learning neural network coupled with a sophisticated search technique. The program managed to sweep Mr. Sedol in the first three games, firmly demonstrating its power, despite his efforts to find creative ways around the onslaught.
“He tried so many different things. Against any human something would have worked, but the program just countered everything perfectly,” Dr. Mueller said.
At their fourth meeting, having already lost the series, Mr. Sedol was free to play AlphaGo in a more daring fashion. He was able to spot an unlikely move that the program did not see, and it ultimately won him the game. But in the fifth game, which some experts have judged the most hard-fought, the computer ultimately prevailed again.
Dr. Sutton said the stunning win signalled that it might be time for artificial intelligence (AI) to move beyond the realm of games into areas where conditions are far less idealized, such as controlling robots that have to move and interact with the physical world.
Dr. Mueller agreed that “the air is getting thinner” for researchers interested in building programs that can mimic human game-playing. However, he added, AlphaGo still relied on studying human expert players to learn how to defeat them. The next step will be to design a system that can teach itself how to play – and be a champion – without any guidance.
And while it may seem like an esoteric achievement to the uninitiated, the success of such systems ultimately bodes well for future applications in areas that involve complex decision making of a more practical nature.
Chris Maddison, a graduate student at the University of Toronto and a member of the AlphaGo team, reflected that optimism in an e-mail to The Globe and Mail, sent as he boarded his flight back home from Seoul. “With the help of AI, we have the opportunity to tackle some of humanity's biggest problems,” he wrote.Report Typo/Error