There’s a hall of champions at the University of Alberta that only computer science students know where to find — more of a hallway, really, one office after the next, the achievements archived on hard drives and written in code.
It’s there you’ll find the professors who solved the game of checkers, beat a top human player in the game of Go and used cutting-edge artificial intelligence to outsmart a handful of professional poker players for the very first time.
But lately it’s Richard Sutton who is catching people’s attention on the Edmonton campus.
He’s a pioneer in a branch of artificial intelligence research known as reinforcement learning — the computer science equivalent of treat-training a dog, except in this case the dog is an algorithm that’s been incentivized to behave in a certain way.
It’s a problem that’s preoccupied Sutton for decades, one on which he literally wrote the book, and it’s this wealth of experience that’s brought a growing number of the tech industry’s AI labs right to his doorstep.
Last week, Google’s AI subsidiary DeepMind announced it was opening its first international office in Edmonton, where Sutton — alongside professors Michael Bowling and Patrick Pilarski — will work part-time. And earlier in the year, the research arm of the Royal Bank of Canada announced it was also opening an office in the city, where Sutton also will advise.
Dr. Jonathan Schaeffer, dean of the school’s computer science department, says there are more announcements to come.
Edmonton — which Schaeffer describes as “just off the beaten path” — has not experienced the same frenzied pace of investment as cities like Toronto and Montreal, nor are tech companies opening offices or acquiring startups there with the same fervour. But the city — and the university in particular — has been a hotbed for world-class artificial intelligence research longer than outsiders might realize.
Those efforts date all the way back to the 1980s, when some of the school’s researchers first entertained the notion of building a computer program that could play chess.
The faculty came together “organically” over the years, Shaeffer says. “It wasn’t like there was a deliberate, brilliant strategy to build a strong group here.”
Decades in the making
While artificial intelligence is linked nowadays with advances in virtual assistants, robotics and self-driving vehicles, students and faculty at the university have spent decades working on one of the field’s oldest challenges: games.
In 2007, Schaeffer and his team solved the game of checkers with a program they developed named Chinook, finishing a project that began nearly 20 years earlier.
In 2010, researcher Martin Muller and his colleagues detailed their work on Fuego — then one of the world’s most advanced computer programs capable of playing Go. The ancient Chinese game is notoriously difficult, owing to the incredible number of possible moves a computer has to evaluate, but Fuego managed to beat a top professional on a smaller version of the game’s board.
And earlier this year, a team led by Bowling presented DeepStack, a poker-playing program they taught to bluff and learn from its previously played games. DeepStack beat 11 professional poker players, one of two academic teams to recently take on the task — and a feat the school’s Computer Poker Research Group has been working on since its founding in 1996.
David Churchill — an assistant professor at Memorial University in Newfoundland and formerly a PhD student at the U of A — says that games are particularly well suited to artificial intelligence research, in part because they have well-defined rules, a clear goal and no shortage of human players to evaluate a program’s progress and skill.
“We’re not necessarily playing games for the sake of games,” says Churchill — who spent his PhD teaching computers to play the popular real-time strategy video game StarCraft — but rather “using games as a test bed” to make artificial intelligence better.
‘We don’t seek the spotlight’
The school’s researchers haven’t solely been focused on games, Schaeffer says — even if those are the projects that get the most press. He points to a professor named Russ Greiner, who has been using AI to more accurately identify brain tumours in MRI scans, and Pilarski, who has been working on algorithms that make it easier for amputees to control their prosthetic limbs.
But it is Sutton’s work on reinforcement learning that has the greatest potential to turn the city into Canada’s next budding AI research hub.
Montreal and Toronto have received the bulk of attention in recent years, thanks to the rise of a particular branch of artificial intelligence research known as deep learning. Pioneered by the University of Toronto’s Geoffrey Hinton, and the Montreal Institute for Learning Algorithms’ Yoshua Bengio, among others, the technique has transformed everything from speech recognition to the development of self-driving cars.
But reinforcement learning — which some say is complementary to deep learning — is now getting its fair share of attention too.
Carnegie Mellon used the technique this year in its poker-playing program Libratus, which beat one of the best players in the world. Apple’s director of artificial intelligence, Ruslan Salakhutdinov, has called it an “exciting area of research” that he believes could help solve challenging problems in robotics and self-driving cars.
And most famously, DeepMind relied on reinforcement learning — and the handful of U of A graduates it hired — to develop AlphaGo, the AI that beat Go grandmaster Lee Sedol.
“We don’t seek the spotlight,” says Schaeffer. “We’re very proud of what we’ve done. We don’t necessarily toot our own horn as much as other people do.”
The Lord is my Shepard [to feed, guide, and shield me], I shall not lack. – Psalm 23:1