Working with player tracking data from 7,176 corners taken in the Premier League in 2020 and 2021, the researchers began by graphing the layout of players, with the position, movement, height and weight of players. players coded as nodes on the graph. , and the relationships between players like the lines between them. They then used an approach called geometric deep learning, which takes advantage of the symmetry of a football field to reduce the amount of processing needed by the neural network. (This is not a new strategy: a similar approach was used in DeepMind’s influential AlphaGo research.)

The resulting model led to the creation of a number of tools that could be useful to football coaches. Based on the players’ arrangement at the time the kick is taken, TacticAI can predict which player is most likely to make first contact with the ball and whether a shot will be taken as a result. It can then generate recommendations on the best ways to adjust the player’s position and movements to maximize the chances of a shot being made (for the attacking team) or minimize them (for the defending team) by moving a defender to cover the first post. , for example, or place a man on the edge of the zone.

Liverpool’s football experts particularly liked how TacticAI’s recommendations could identify which attackers were critical to the success of a particular tactic, or which defenders were “asleep at the wheel”, says Veličković. Analysts spend hours combing through video footage looking for weak spots in their opponents’ defensive setups that they can target, or trying to find flaws in their own team’s performance to double down on in practice . “But it’s really difficult to follow 22 people, in many different situations,” says Veličković. “If you have a tool like this, it immediately helps you see which players are not moving the right way, which players should be doing something different.”

TacticAI can also be used to find other coins with a similar pattern of players and moves, again saving analysts hours. According to DeepMind, the suggestions made by the model were found useful by Liverpool coaches twice as often as current techniques, which are based solely on players’ physical coordinates and do not take into account their movements or physical attributes. (Two corners can look the same, but if the big striker is on the edge of the box in one and running towards the near post in the other, it’s probably important.)