Working with player data from 7,176 corners taken in the Premier League in 2020 and 2021, the researchers began by displaying player lineups on a graph, with players’ position, movement, height and weight as nodes in the graph were coded. and relationships between players as the lines between them. They then used an approach called geometric deep learning, which uses the symmetry of a football field to reduce the amount of processing the neural network has to do. (This isn’t a new strategy; a similar approach was used in DeepMind’s influential AlphaGo study.)
The resulting model led to the creation of a number of tools that could be useful to football coaches. Based on the position of the players 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 for the best ways to adjust the player’s position and movement to maximize (for the attacking team) or minimize (for the defending team) the chance of a shot – moving a defender across to cover the nearest post. , for example, or placing a man at the edge of the area.
The football experts at Liverpool particularly liked how TacticAI’s recommendations could identify attackers who were crucial to the success of a particular tactic, or defenders who were ‘asleep at the wheel’, says Veličković. Analysts spend hours sifting through video footage looking for weaknesses in their opponents’ defensive lineups to target, or trying to find holes in their own team’s performance that they can double down on during training. “But it is very difficult to trace 22 people in many different situations,” says Veličković. “If you have a tool like that, you can immediately see which players are not moving in the right way, which players should be doing something different.”
TacticAI can also be used to find other angles that exhibit similar player and movement patterns, again saving analysts hours of time. According to DeepMind, the model’s suggestions were rated as useful by Liverpool coaches twice as often as current techniques, which are based only on the players’ physical coordinates and do not take their movements or physical attributes into account. (Two corners may look the same, but if the tall striker is in one corner at the edge of the penalty area and running towards the near post on the other side, that’s probably important.)
One thing it also does, according to DeepMind’s Zhe Wang, another lead contributor to the paper, is compensate for the lack of appropriate language to describe the huge range of different things that can happen on a corner. Unlike American football, which has a deep and storied nomenclature for different plays and running routes, choreographing football sets in such detail is a relatively new phenomenon. “Different coaches may have their own expressions for the corner kick patterns they observe,” says Wang. “So with TacticAI we hope to use the power of deep learning to develop a common language to describe corner kick patterns.”
According to the article, in the future, the researchers hope to build TacticAI into a natural language interface so that coaches can query it in text and get answers to the problems they are trying to solve in the field. Veličković says the model can be used during a match to help coaches fine-tune their corner routines on the fly, but it will likely be useful in the days leading up to a match, where it will free up coaches’ time. “We don’t want to build AI systems that replace experts,” says Veličković. “We want to build AI systems that strengthen the capabilities of experts, so that they can do their work a lot more efficiently and have more time for the creative part of coaching.”