Quantcast
The opening ceremony of the Future of Go Summit. Image: AP Photo/Peng Peng

An Alberta AlphaGo Pioneer Is in China to Watch the AI Wallop Human Opponents

Lisa Cumming

Lisa Cumming

Meet the man behind the machine.

The opening ceremony of the Future of Go Summit. Image: AP Photo/Peng Peng

Earlier this week, Google DeepMind's AlphaGo beat Ke Jie, the world's best Go player, for the second time—wrapping up a best-two-out-of-three match.

Martin Müller, a computer scientist at the University of Alberta and an expert Go player himself, was in the town of Wuzhen in China to witness the historic event. In the early 2000s, Müller helped to lay the groundwork for the machine that has now beaten the most talented humans at their own game.

Canada has contributed a lot of research to deep learning technology thanks to scientists like Geoff Hinton and Yoshua Bengio, who developed many of today's fundamentals. Canadians also played a more direct role in AlphaGo's success. The two main authors of the program, David Silver and Aja Huang, both spent time at the University of Alberta, respectively doing a PhD and postdoctoral work there. Both were supervised by Müller.

"Before AlphaGo, much of the fundamental games and machine learning research was done here," Müller wrote in an email. "If you look through the references list of the AlphaGo paper in the journal Nature, over 40% of these references have a University of Alberta (co-)author. Then, DeepMind greatly surpassed all of these previous efforts with their new ideas."

Müller has been around long enough to see technological advances shape AI, especially when it comes to his specialties, which just so happen to be the secret sauce behind AlphaGo: the Monte Carlo method and heuristic algorithms.

Read More: A New AI Hub In Canada Is Welcoming the Scientists Trump Is Pushing Away

Heuristics for machines work similarly as they do with humans; we often use an incomplete idea or theory about the future to guide our next action. Heuristic algorithms help AIs to act more quickly by ranking the actions most likely to lead to an approximation of the desired solution (say, winning a game of Go) at every decision point.

AlphaGo, for its part, uses a Monte Carlo Tree Search heuristic algorithm, which quickly "searches" branches of a decision tree to select the AI's next move.

"The main breakthrough [in AlphaGo] compared to previous programs is that it uses much better knowledge during the search," wrote Müller. "This knowledge is learned before the games and is encoded in deep neural networks. It is used for move selection, and for position evaluation, ie. how promising is a Go position for the player."

Müller sees the back-to-back wins against Ke Jie as a milestone for AI and machine learning. But board games aren't the only applications for the powerful machine learning techniques behind AlphaGo.

"The techniques underpinning AlphaGo and much of our other work are general-purpose and could potentially be applied to a wide range of other domains," a spokesperson for DeepMind wrote me in an email. "We believe that in the next few years scientists and researchers using similar approaches will generate insights in a multitude of areas, from superconductor material design to drug discovery."

Despite AlphaGo's power, Müller doesn't think that winning in Go against AlphaGo is impossible.

"I don't think absolute zero, but very very close to zero in practice," he wrote. However, "it is important to remember that AlphaGo is a heuristic program. While it plays extremely strongly, it does not come with any guarantees."

Kind of like humans.

Get six of our favorite Motherboard stories every day by signing up for our newsletter.