Score More from Set Plays: Tapping into the Power of AI for Effective Corner Kicks

Google’s DeepMind has introduced TacticAI:

An AI system that might be capable of helping experts get tactical information, especially on corner kicks using predictive and generative AI.

In the 2019 Champions League semifinal, Liverpool defeated Barcelona 4-0. That particular goal scored by Divock Origi, which won the title of the greatest in the club’s 130-year history, was later voted as the best. The right-back Trent Alexander-Arnold of Liverpool won a corner for his side and placed the ball on the corner arc. He then walked away pretending to do so and suddenly turned back and hit a fast, low kick directly to Origi, who lobbed the ball into the net, top left corner.

Although football does not often happen in predictable situations, corner kicks are the exception because they are very repeatable, and teams practice for them a lot. The goal by Alexander-Arnold-Origi was clearly the outcome of a well-rehearsed movement, and even slight advantages in the approach to corner kicks can be game-changers for a team in the tournament.

This is where AI (artificial intelligence) – which is usually thought of as having minimal impact in sports other than maybe board games such as chess or individual track and field events – could be a factor.

DeepMind’s TacticAI

If the ball is passed over the goal line after touching a player of the defending team, a corner is given. In the usual Champions League game, there are about 10 corner kicks and the probability of predicting their outcomes is very low because of the individual player's randomness and the dynamics among them.

Although exact gold-standard data about corner kicks is not available, Tactic AI says it has promising results by using “a geometric deep learning approach” that helps build more generalization models.

Predictive and Generative

DeepMind had been working on the TacticAI with the experts from Liverpool Football Club in the multi-year research collaboration.

It is written that the human expert judges liked TacticAI’s recommendations 90% of the time instead of the real-life tactical arrangements seen on the field. Petar Velickovic, the team leader who has presented the TacticAI tool to the team managers, has suggested it be used as an assistant for tactics.

The main idea behind modern football — and TacticAI, according to Velickovic, is to predict and generate the most effective counter-tactics, and to do so, the AI has a predictive and generative component to allow coaches to sample and explore alternative player setups for each corner kick, and to select those that have the highest probability of success.

Sports, such as football, are a complex area for AI development, which involves real-world, multi-agent interactions with multi-modal data. While TacticAI has shown the capabilities of AI assistive techniques to transform sports for players, coaches, and fans, AI in sports could also be used to create actionable models that could be applied off the field — from computer games and robotics to traffic coordination.

AI in corner kicks, however, is not only used in the game itself but also the pre-game preparation.

Velickovic’s team at DeepMind published a paper titled Game Plan, which focused on why AI should be employed to help football tactics, illustrating instances such as examining penalty kicks. Next, in 2022, it came up with ‘Graph Imputer’ which showed how AI can combine with a prototype of a predictive system in football analytics for downstream tasks.

(A) Illustration of how corner kick situations are converted to a graph representation. Every player is considered as a node in a graph and their links are as edges. A neural network runs over the graph and the messages are passed from one node to another for each node to be updated.

(B) The way TacticAI handles a corner kick is described in the following paragraphs. All four reflections are used in the corner, and then the output is fed to the core TacticAI model. They work together to compute the final representation of players, which can be used to predict the outcome.

(Source: DeepMind)

The most recent version, TacticAI, is advertised as an “ AI system in full”. According to the DeepMind team, it is built to address three core questions: According to the DeepMind team, it is built to address three core questions:

What will the corner kick tactical setup appear like? For e.g., who is the most probable player to receive a pass, and will there be a shot attempt?

How can we know that a given sequence is the correct one after the setup has been played? For example, what are the success stories of using similar tactics?

What should we do to adapt ourselves to make a particular result possible? For example, what is the best way to reposition the defensive players to make it less likely for the opponent to take a shot?

The tool consists of graphs, where players are shown as nodes and their relationships are represented as edges, to represent corner kick situations. The next step is to feed the graph into a neural network which will then recompute each node’s representation. This allows for the interaction between players to be modeled, which might be more important than the actual distances between them.

Why specifically corners

The corner kicks are highlighted in a way that is a bit strategic too. The models began with trying to predict aspects of open play in a football match, but there are pitfalls: even if a model would be able to give a hint about the current state of open play, a coach is not always able to take appropriate action to it at that very moment. Instructions given out in the open may be misunderstood by the players or inform the other side about the team's strategies.

AI tools are useful for corner kicks as they are the moments of the game when it is frozen. In addition, they are the moments when the game starts from the same position at the corner of the pitch, with the opportunity being immediate to score. The corner strategies, which are often selected well in advance, are also usually decided before the players go onto the pitch, which avoids any confusion on the match day.