Artificial intelligence is showing up in more and more parts of everyday life, from self-driving cars to chatbots. But now, it’s also making its way into sports. A team of researchers at the University of Waterloo is using AI simulations to rethink how sports analytics could work, especially in areas where data is hard to come by.
David Radke, who recently finished his PhD in computer science at Waterloo and now works as a senior research scientist for the NHL’s Chicago Blackhawks, says his background as both a student-athlete and sports fan helped spark the project.
“I played collegiate sports through my undergraduate degree and part of graduate school and we are both fans of sports in general,” Radke said. “When I began working in hockey and gauging the landscape of publicly available data, we realized there was a real opportunity for simulated AI data to help the community.”
To run their experiments, the researchers used Google Research Football, an open source simulation that works well with reinforcement learning. “Google Research Football was a good environment for us to use because it is open source and runs in a similar framework to other reinforcement learning environments that we had used for other research,” Radke explained.
The team simulated 3,000 matches and collected detailed data, such as player locations and movements on the field, as well as in-game events like passes, shots, and turnovers. “We set up the environment to record player locations at each timestep of the game… We then created scripts to extract events such as passes, shots, and turnovers to augment the ‘tracking’ dataset with ‘event’ data,” Radke said.
Radke explained that the goal is not to perfectly mirror real soccer. “We do not make any claims about the realism of the games other than the fact there are events since the agents were trained purely using reinforcement learning,” Radke said. “The point of the paper is generating similar data to what you would work with in real soccer.”
This kind of simulated data matters because professional leagues often keep their tracking data private. By building something similar, Waterloo researchers are giving other students and academics a chance to experiment with the same kind of detailed stats used in professional sports.
The approach could even be applied to other sports like hockey or basketball. “Yes, the approach would be the exact same… The difficult part about scaling to those sports is accessibility to realistic simulated environments,” Radke said.
While the project isn’t about perfectly recreating professional matches, for Radke, it’s a full-circle moment, bringing together his experiences in sports, computer science, and now professional hockey.
