In March 1950, an RAF wing commander and trained accountant called Charles Reep turned his eye for numbers to football. Reep, who had become interested in the sport in the 1930s and was fascinated by Herbert Chapman’s pioneering Arsenal team, had returned from the Second World War to find that the tactical revolution he’d witnessed before had stalled.
Finally, at half-time during a drab Division Three game between Swindon Town and Bristol City during which he watched countless attacks amount to nothing, Reep’s patience ran out. He grabbed a notebook and a pencil and began furiously jotting down what happened on the pitch – he started counting the number of passes and shots, in one of the first systematic attempts to use data to analyse football.
Seven decades later, the data revolution has reached the grassroots – fans are fluent in xG and net spend, and the top teams pluck statistics PhD students straight from university in the search for an edge. Now, defending Premier League champions Liverpool have joined forces with DeepMind to explore the use of artificial intelligence in the football world. A paper by researchers at the two organisations, published today by the Journal of Artificial Intelligence Research, outlines some of the potential applications.
“The timing is just right,” says Karl Tuyls, an AI researcher at DeepMind and one of the lead authors on the paper. DeepMind’s collaboration at Liverpool arose from his previous role at the city’s university (DeepMind founder Demis Hassabis is also a lifelong Liverpool fan and was an advisor on the research). The two groups got together to discuss where AI might be able to help football players and coaches. Liverpool also provided DeepMind with data on every Premier League game the club played between 2017 and 2019.
In recent years, the amount of data available in football has swelled with the use of sensors, GPS trackers and computer vision algorithms to track the movement of both players ball. For football teams, AI offers a way to spot patterns that coaches can’t; for DeepMind researchers, football offers a constrained but challenging environment for them to roadtest their algorithms. “A game like football is super interesting because there are a lot of agents present, there’s competition and collaborative aspects,” says Tuyls. Unlike chess, or Go, football has inherent uncertainty built into it because it’s played in the real world.
That doesn’t mean you can’t make predictions, though – and that’s one area where AI could prove particularly useful. The paper demonstrates how you can train a model on data about a specific team and lineup to predict how its players will react in a particular situation: if you knock a long ball into the right hand channel against Manchester City, for example, Kyle Walker will run in a particular direction, while John Stones may do something else.
This is known as ‘ghosting’ – because the alternative trajectories are overlaid 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 goes off injured. These are things that coaches would likely notice themselves, and Tuyls stresses that the aim isn’t to design tools to replace them. “There’s lots of data, lots to digest and it’s not necessarily so easy to handle these masses of data,” he says. “We’re trying to build assistive technology.”
As part of the paper, the researchers also conducted analysis on more than 12,000 penalty kicks taken across Europe in the last few seasons – categorising players into clusters based on their style of play, and then using that information to make predictions about where they were most likely to hit a penalty, and whether they were likely to score. Strikers were, for instance, more likely to aim for the bottom left corner than midfielders – who took a more balanced approach, and the data demonstrated that the optimal strategy for penalty takers was, perhaps unsurprisingly, to kick to their strongest side.
ther models may be able to crunch the numbers on counterfactuals – to estimate how much a specific action like a pass or a missed tackle contributed to a goal, or xG. They could be used in the post-match analysis to demonstrate to players why they should have passed the ball in a particular situation instead of trying to shoot. A model trained on player performance data – strength and fitness – might be able to track fatigue better than human coaches and recommend resting players before they get injured.
There are echoes here of what Reep tried to do in the 1950s – he used his data to (erroneously) calculate that most goals were scored after moves of four passes or fewer, and his analysis helped usher in a style of long-ball football which became the hallmark of the English game for decades. There have been high-profile examples of AI in other domains spitting out answers that are nonsensical or just plain wrong – in the past, AIs trained on video games have won by breaking the rules of the game, or ignoring the laws of physics. An AI trained on football data might decide, like a robot Jose Mourinho, that actually the best way to get results is to let the opponent keep the ball and wait for them to make a mistake.
That’s why it’s important that the findings of the model are mediated by experts, Tuyls says, to ward against faulty reasoning by the AI systems. But an AI could spot a pass that – in the heat of the moment – even the best player might miss. “We’re not trying to build robots, we’re trying to improve human football play,” he says.
AI won’t replace football managers, Tuyls says, but its impacts could be felt within the next decade. “The purpose is to have a seamless system that integrates well with the human player on the pitch and facilitates their work,” he says. “I don’t think you will see big impacts in the next six months or a year, but in the next five years some of the tools will be more developed, and you could see something like an ‘Automated Video Assistant Coach’ that can help with pre and post-match analysis, or can look at the first half of a game and give you advice on what could be changed in the second half.”
DeepMind is hoping to combine computer vision, statistical learning and game theory to help teams spot patterns in the reams of data they’re collecting that they wouldn’t be able to see otherwise. Applying artificial intelligence to football could make players and coaches smarter –now if only it could do the same for the owners.
