In March 1950, Charles Reep, RAF wing commander and trained accountant, turned to football for the numbers. Reep, who had taken an interest in the sport in the 1930s and was fascinated by Herbert Chapman’s pioneering team, Arsenal, had returned from World War II to find that the tactical revolution he had witnessed before was in place. dead.
Finally, at halftime in a lackluster Division 3 game between Swindon Town and Bristol City, in which he saw countless attacks cost nothing, Reep’s patience ran out. He grabbed a notebook and pencil and began to furiously note what had happened on the pitch: he started counting the number of passes and shots in one of the first systematic attempts to use data to analyze football.
Seven decades later, the data revolution has reached the grassroots – fans speak fluent xG and net spending, and the best teams pull the statistics of doctoral students straight from the university in search of an edge. Now defending Premier League champions Liverpool have teamed up with DeepMind to explore the use of artificial intelligence in the world of football. An article by researchers from both organizations, published today by the Artificial Intelligence Research Journal, describes some of the potential applications.
“The timing is right,” says Karl Tuyls, AI researcher at DeepMind and one of the lead authors of the article. DeepMind’s collaboration in Liverpool grew out of his previous role at the city’s university. (Founder of DeepMind Demis Hassabis is also a longtime Liverpool fan and has served as a research advisor.) The two groups came together to discuss how the AI could help football players and coaches. Liverpool also provided DeepMind with data on every Premier League game the club played from 2017 to 2019.
In recent years, the amount of data available in football has grown through the use of sensors, GPS trackers and computer vision algorithms to track the movement of players and the ball. For soccer teams, AI offers a way to spot patterns that coaches can’t; for DeepMind researchers, football offers them a constrained but stimulating environment to test their algorithms on the road. “A game like [soccer] it’s super interesting, because there are a lot of agents present, there are aspects of competition and collaboration, ”says Tuyls. Unlike chess, or Go, football has an inherent uncertainty because it is played in the real world.
That doesn’t mean you can’t make predictions, though – and this is one area where AI could prove particularly useful. The paper shows how you can train a model on data about a specific team and roster to predict how their players will react in a particular situation: If you hit a long ball in the right channel against Manchester City, for example, Kyle Walker will work. in a particular direction, while John Stones can do something else.
This is known as “ghost” – because the alternate trajectories are superimposed on what actually happened, like in a video game – and has a range of different applications. It could be used, for example, to predict the implications of a tactical change or how an opponent might play if a key player gets injured. These are things that coaches would probably notice themselves, and Tuyls stresses that the point isn’t to design tools to replace them. “There’s a lot of data, a lot of stuff to digest, and it’s not necessarily that easy to manage this mass of data,” he says. “We are trying to develop assistive technologies.”
For this article, the researchers also analyzed more than 12,000 shots on goal taken across Europe in recent seasons – categorizing players into groups based on their style of play, then using that information to make predictions about where they happened to be the most. likely to hit a penalty and whether they were likely to score. Strikers were, for example, more likely to aim for the lower left corner than midfielders – who took a more balanced approach, and the data showed that the optimal strategy for penalty shooters was, perhaps unsurprisingly. , to kick their stronger side.