Tests
As part of our long-standing partnership with Liverpool FC, we are developing a full AI system that can advise coaches on corner kicks
“Corner quickly taken… Origi!”
Liverpool FC made a historic comeback in the semi-finals of the 2019 UEFA Champions League. One of the most iconic moments was Trent Alexander-Arnold’s corner kick, after which Divock Origi scored a goal that went down in history as Liverpool FC’s greatest goal.
Corners have great scoring potential, but developing a routine relies on a combination of human intuition and game design to identify patterns in competing teams and react on the fly.
Today in Nature communication, we present TacticAI: an artificial intelligence (AI) system that can provide experts with tactical insights, especially on corner kicks, thanks to predictive and generative AI. Despite the restricted availability of gold standard corner kick data, TacticAI achieves state-of-the-art results using a geometric deep learning approach that helps create more generalizable models.
We developed and evaluated TacticAI together with experts from Liverpool Football Club as part of a multi-year research collaboration. TacticAI suggestions were preferred by human experts 90% of the time over tactical configurations used in practice.
TacticAI demonstrates the potential of AI-enabled techniques to revolutionize sports for players, coaches and fans. Sports like soccer are also a energetic field for AI development because they involve real-world interactions with multiple agents and multimodal data. The development of artificial intelligence in sports may translate into many areas on and off the pitch – from computer games and robotics to movement coordination.
TacticAI is a full AI system with combined predictive and generative models, allowing you to analyze what happened in previous games and how to make changes to augment the likelihood of a specific outcome.
Developing a game plan with Liverpool FC
Five years ago, we began a multi-year collaboration with Liverpool FC to improve artificial intelligence for sports analytics.
In our first article, Game Plan, we analyzed why AI should be used to aid football tactics, highlighting examples such as penalty kick analysis. In 2022, we have grown Chart assignment, which showed how artificial intelligence could be used in a prototype predictive system to perform further tasks in football analytics. The system could predict players’ movements off-camera when no tracking data was available – otherwise the club would have had to send a scout to watch the match in person.
We have now developed TacticAI as a full AI system with combined predictive and generative models. Our system allows coaches to try alternative player alignments for any formation of interest and then directly evaluate the possible outcomes of such alternatives.
TacticAI was built to answer three basic questions:
- What happens with a given tactical corner kick formation? e.g., who is most likely to receive the ball and will a shot be attempted?
- After playing the setup, can we understand what happened? e.g. has a similar tactic worked in the past?
- How to adapt tactics to achieve a specific result? e.g. how should the position of defending players be changed to reduce the likelihood of shooting?
Corner kick prediction using geometric deep learning
A corner kick is awarded when the ball crosses the goal line after touching a player of the defending team. Predicting the results of corner kicks is intricate due to the randomness of individual players’ play and the dynamics between them. This problem is also tough for AI to model due to the restricted amount of gold-standard corner kick data available – only around 10 corners are taken in each Premier League match each season.
(A) How corner kick situations are converted into a graphical representation. Each player is treated as a node in the graph. A graph neural network operates on this graph, updating the representation of each node through message passing.
(B) How TacticAI processes a given corner kick. All four possible reflection combinations are applied to the corners and passed to the base TacticAI model. They interact with each other to calculate the final representation of the player, which can be used to predict outcomes.
TacticAI successfully predicts corners using a geometric deep learning approach. First, we directly model the implicit relationships between players by representing corner setups as graphs where nodes represent players (with features such as position, speed, height, etc.) and edges represent the relationships between them. We then employ the approximate symmetry of a football field. Our geometric architecture is a variation Group-equivalent convolutional network this generates all four possible reflections of a given situation (original, reverse H, reverse V, reverse HV) and forces our predictions about receivers and shot attempts to be identical in all four of them. This approach limits the search space of possible features that our neural network can represent to those that take reflection symmetry into account, and produces more generalizable models with less training data.
Providing constructive suggestions to human experts
Using its predictive and generative models, TacticAI can lend a hand coaches find similar corners and test different tactics.
Traditionally, to develop tactics and counter tactics, analysts would re-watch many match videos, looking for similar examples and studying competing teams. TacticAI automatically calculates a numerical representation of players, allowing experts to easily and efficiently find relevant past routines. We then confirmed this intuitive observation through extensive qualitative research with soccer experts, who found that TacticAI’s top results searches were right 63% of the time, which is almost double the benchmark of 33% seen in matchmaking approaches. based on direct analysis of the similarity of the players’ positions.
TacticAI’s generative model also enables coaches to redesign corner kick tactics to optimize the likelihood of specific outcomes, for example reducing the likelihood of a shot being taken in a defensive formation. TacticAI provides tactical recommendations that adjust the positions of all players on a given team. Based on the proposed adjustments, coaches can more quickly identify vital patterns, as well as key factors influencing the success or failure of tactics.
(A) Example of a corner kick where a shot was actually attempted.
(B) TacticAI can generate a counterfactual in which the shot probability was reduced by adjusting the position and speed of defenders.
(C) Suggested defender positions reduce receiver probability for attacking players 2-4.
(D) The model is able to generate many such scenarios, and trainers can explore different options.
In our quantitative analysis, we showed that TacticAI successfully predicted corner receivers and shooting situations, and that player positioning was similar to real action. We also assessed these recommendations qualitatively in a blind case study in which the evaluators did not know what tactics were used. from a real game and which were generated by TacticAI. Human football experts at Liverpool FC found that our suggestions were indistinguishable from real corner kicks and were preferred over the original situations 90% of the time. This shows that TacticAI’s predictions are not only right, but also useful and implementable.
Examples of strategic improvements that evaluators preferred over the original plays TacticAI suggested:
(A) The four players’ recommendations are more favorable by most evaluators.
(B) Defenders furthest from the corner are better at covering coverage
(C) Corrected marking runs by the central group of defenders in the penalty box
(D) Much better runs for the two center backs, along with better positioning for the two other defenders in the goal area.
The development of artificial intelligence in sports
TacticAI is a full artificial intelligence system that can provide coaches with instant, comprehensive and right tactical insights that are also actionable on the pitch. With TacticAI, we have developed a capable AI assistant for football tactics and achieved a milestone in developing useful assistants in sports AI. We hope that future research will lend a hand develop assistants that employ more multimodal inputs beyond player data, as well as assist 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. This is a very energetic and analytical game in which many human factors are involved, from body structure to psychology. Detecting all the patterns is a challenge even for experts such as experienced trainers. With TacticAI, we hope to learn many lessons from developing broader enabling technologies that combine human knowledge and AI analysis to lend a hand people in the real world.