A football data model using artificial intelligence called a graph neural network has helped the American Football Federation (USSF) improve the success rate of counter-attacks.
For Joris Bekkers, football data is something more exciting than Christmas.
A few winters ago, when Bekkers colleagues at the US Soccer Federation (USSF) went home for the holidays, the Dutch data scientist was still taking advantage of the time to build football data models. stone.
It was this period of time that also allowed him to pursue his secret passion of building a football data model of his own.
Teaching computers to learn football tactics
For Bekkers and many other data experts, teaching computers to learn soccer tactics is more appealing than drinking and shouting in front of the TV on Boxing Day.
Speaking to The Athletic, Bekkers admitted he was first intrigued by a YouTube talk about using a type of artificial intelligence called graph neural networks to understand data. whether football is complicated.
As the name implies, a neural network is a computer system with interconnected nodes that act like neurons in the human brain to be more efficient.
The biggest advantage of this model is that they can study a large number of resources and gradually connect the dots to learn how football works.
That’s the idea, but unfortunately for anyone thinking about firing a real coach and replacing it with a machine, data science isn’t as easy as asking Siri to watch a thousand games and then have it. Can lead a team.
First, the Bekkers must combine event data (logs of human actions with the ball) with logs of tracking data (computer vision records of how the player and the ball move) to get computer- and human-language logs of what’s going on in the match.
The Dutch data scientist then had to translate all that information into a graph form that the neural network could understand.
In fact, this chart is somewhat similar to the passing diagram recorded by the statistics pages, with the players on the field being the dots, and the direction of the arrow will be the connection between them and their teammates with the ball. .
With neural networks, the dots and arrows on the graph are more than simply recording the player’s position. They can include the entire context of the player and surrounding links, such as which direction they are moving, and at what speed.
“You can even add color to the player’s boots if you think that’s important,” joked Bekkers. The computer will need a complete view of the player’s position and movement without the ball to start processing tactical tasks.
However, there is still a big problem as to how Bekkers can build his own model and this is also when this expert has hit a giant wall.
Specifically, when I tried to run models for my neural networks, they didn’t work. The silicon brain of the computer coach that Bekkers worked so hard on was broken.
At the same time, the Christmas break is over and the Bekkers must return to normal football work at the USSF. Bekkers’ passion would have been unfinished without a surprise message from someone named Amod Sahasrabudhe.
Unlike Bekkers who is a football data scientist and has deep understanding of artificial intelligence, Sahasrabudhe just happened to join this project as a fresh graduate and also a longtime Manchester fan. United.
On the advice of a friend, Sahasrabudhe watched the entire YouTube series that inspired the Bekkers. The graduate student suddenly realized that people with such expertise were contributing to cutting-edge technological research for some of the world’s largest football organizations.
An avid soccer fan, Sahasrabudhe immediately contacted Bekkers and convinced the USSF to accept him as an intern and use his knowledge of artificial intelligence on a mission to “rescue” the neural network project- Neurons seem to have reached a dead end.
After some trial and error and different ways of building the model, Sahasrabudhe and Bekkers finally discovered an interesting relationship between field tactics and molecular bonding.
“Changing the shape of a molecule means it will have a different property. We believe the same thing happens in football,” Sahasrabudhe explained.
With just a slight change of position on the field, the whole project knot was removed. Before long, the two had successfully built a football-specific computer brain that could function smoothly.
The problem now is that the computer needs something to learn. After consulting with league performance analysts, Sahasrabudhe and the Bekkers decided to focus on one of the messiest phases of a game, the counter-attacks, which account for around 10% of goals. won at the American Professional League (MLS).
What makes a counter-attack a success or a failure? What exactly is a counterattack?
According to the USSF’s own definition, a counter-attack is a sequence of open situations that begins immediately after the defending team has successfully captured the ball and moves upwards at least 10 m at a speed of 4 m/s.
Bekkers and Sahasrabudhe have narrowed down this concept by marking a successful counter-attack if the attacking team gets into the opponent’s box by any means but a long dribble.
The job of the neural network is now to figure out what helped a team counterattack successfully. Given a large enough number of attempts, the answer the computer gives may sound a bit ridiculous.
Specifically, after studying thousands of counter-attacks, the artificial intelligence concluded that the secret to success was simply moving as quickly as possible.