Research
As part of our multi-year partnership with Liverpool FC, we are developing a full artificial intelligence system that can advise coaches on corner kicks
“The corner was quickly turned… Origi!”
Liverpool FC made a historic return to the 2019 UEFA Champions League semi-finals. One of the most iconic moments was a Trent Alexander-Arnold corner that teed up Divock Origi to score what has gone down in history as Liverpool FC’s greatest goal.
Corners have high scoring potential, but devising a routine relies on a mix of human intuition and game planning to spot patterns in opposing teams and respond on the fly.
Today, at Nature communications, we present TacticAI: an artificial intelligence (AI) system that can provide experts with tactical information, particularly in corner kicks, through predictive and genetic artificial intelligence. Despite the limited availability of gold standard corner kick data, TacticAI achieves cutting-edge results by using a geometric deep learning approach to help build more generalizable models.
We developed and evaluated TacticAI together with experts from Liverpool Football Club as part of a multi-year research collaboration. TacticAI’s suggestions were preferred by the expert evaluators 90% of the time over tactical setups observed in practice.
TacticAI demonstrates the potential of AI assistive techniques to revolutionize sports for players, coaches and fans. Sports such as football are also a dynamic area for AI development, as they feature real multi-agent interactions with multi-modal data. The advancement of artificial intelligence for sports could translate into many areas on and off the field – from computer games and robotics, to traffic coordination.
TacticAI is a complete AI system with combined predictive and generational models to analyze what happened in previous games and how to make adjustments to make a certain outcome more likely.
Developing a game plan with Liverpool FC
Three years ago, we began a multi-year partnership with Liverpool FC to advance AI for sports analytics.
Our first paper, Game Plan, looked at why AI should be used to help football tactics, highlighting examples such as penalty kick analysis. In 2022 we developed Graphing calculator, which showed how artificial intelligence can be used with a prototype of a prediction system for downstream tasks in football analytics. The system could predict player movements off-camera when no tracking data was available – otherwise, a club would have to send a scout to watch the match in person.
Now, we have developed TacticAI as a complete AI system with combined prediction and generation models. Our system allows coaches to test alternative player setups for each routine of interest and then directly assess the potential outcomes of such alternatives.
TacticAI is built to answer three key questions:
- For a given corner tactical setup, what will happen? e.g. who is more likely to receive the ball and a shot will be attempted?
- Once a setup is played, can we tell what happened? eg, have similar tactics worked well in the past?
- How can we adjust tactics to achieve a specific result? e.g. how should defensive players be repositioned to reduce the chance of shot attempts?
Predicting corner results with geometric deep learning
A corner is awarded when the ball crosses the line after touching a player of the defending team. Predicting the outcome of corner kicks is complex due to the randomness in play by individual players and the dynamics between them. This is also difficult for AI to model due to the limited gold standard corner kick data available – only around 10 corners are played in each Premier League match each season.
(A) How corner kick situations are converted into a graph representation. Each player is treated as a node in a graph. A graph neural network operates on top of this graph by updating the representation of each node using message passing.
(B) How TacticAI processes a given corner. All four possible reflex combinations are applied to the corner and fed into the basic TacticAI model. They interact to calculate the player’s final representations, which can be used to predict outcomes.
TacticAI successfully predicts corner play by applying a geometric deep learning approach. First, we directly model the implicit relationships between players by representing the corner settings as graphs, in which nodes represent the players (with attributes such as position, speed, height, etc.) and edges represent the relationships between them . Next, we exploit an approximate symmetry of the football field. Our geometric architecture is a variation of it Group equivalent convolutional network which creates all four possible reflections of a given state (original, H-flipped, V-flipped, HV-flipped) and forces our receiver predictions and reception attempts to be identical in all four. This approach reduces the search space of possible functions that our neural network can represent to those that respect reflection symmetry — and yields more generalizable models, with less training data.
Providing constructive suggestions to human experts
Leveraging its predictive and production models, TacticAI can assist coaches by finding similar corner kicks and testing different tactics.
Traditionally, to develop tactics and counter-tactics, analysts would re-watch multiple game videos to look for similar examples and study opposing teams. TacticAI automatically calculates players’ numerical representations, which allows experts to easily and efficiently search for relevant past routines. We further validated this intuitive observation through extensive qualitative studies with soccer experts, who found TacticAI’s top-1 retrievals to be relevant 63% of the time, nearly double the 33% benchmark seen in approaches that propose pairs that based on direct analysis of player position similarity.
TacticAI’s production model also allows human coaches to redesign corner kick tactics to optimize the chances of certain outcomes, such as reducing the chance of a shot attempt for a defensive set-up. TacticAI provides tactical recommendations that adjust the positions of all players in a particular team. From these suggested adjustments, coaches can identify important patterns, as well as key players in the success or failure of a tactic, more quickly.
(A) Example of a corner where a shot was actually attempted.
(B) TacticAI can create an opposing setup in which the shot probability is reduced by adjusting the position and speeds of the defenders.
(C) Suggested linebacker positions result in reduced receiver probability for offensive linemen 2-4.
(D) The model is capable of generating multiple such scenarios and coaches can inspect the different options.
In our quantitative analysis, we showed that TacticAI was accurate in predicting corner takers and shooting situations, and that player repositioning was similar to the progression of real games. We also evaluated these recommendations qualitatively in a blind case study where the raters did not know which tactics were from the real game and which of them were generated by TacticAI. Human soccer experts from Liverpool FC found that our suggestions are indistinguishable from real corners and were preferred over their original state 90% of the time. This shows that TacticAI’s predictions are not only accurate, but useful and scalable.
Examples of strategic improvements that raters preferred from the original games, where TacticAI suggested:
(A) Four players’ recommendations are more favorable than most raters.
(B) Defenders farther from the corner make improved coverage runs
(C) Improved coverage routes for a central group of defenders in penalty kicks
(D) Substantially better running for two central defenders, along with better positioning for two other defenders in the goal area.
Advanced artificial intelligence for sports
TacticAI is a complete artificial intelligence system that could give coaches immediate, comprehensive and accurate tactical information – which is also practical on the pitch. With TacticAI, we have developed a capable artificial intelligence assistant for football tactics and achieved a milestone in the development of useful assistants in sports artificial intelligence. We hope that future research can help develop assistants that expand to more multimodal inputs beyond player data and help experts in more ways.
We show how artificial intelligence can be used in football, but football can also teach us a lot about artificial intelligence. It is an extremely dynamic and challenging game to analyze, with many human factors from physique to psychology. It is difficult even for experts such as experienced coaches to detect all the patterns. With TacticAI, we hope to learn many lessons for developing broader assistive technologies that combine human expertise and AI analysis to help people in the real world.