Google DeepMind and Liverpool join forces again to improve sports analytics
As part of Google's long-term collaboration with the Liverpool football club, Google DeepMind introduced the TacticAI artificial intelligence system, which is designed to advise football coaches on corner kick tactics.
The system, created in partnership with one of the UK's most prolific teams, provides tactical information using predictive and generative artificial intelligence models. Blind testing, in which the experts did not know which tactics were chosen by the real trainer and which were suggested by TacticAI, showed that in 90% of cases TacticAI's suggestions were preferable to tactical settings observed in practice.
TacticAI — it is a full-fledged artificial intelligence system that combines predictive and generative models. It allows coaches to select alternative player formations and evaluate the possible results of such alternatives.
TacticAI addresses three main issues. The first — what will happen when using certain tactics on a corner kick? For example, who is most likely to take possession of the ball? Second — what were the results after using such tactics previously? For example, were such tactics successful in the past? And the third — how you can adjust tactics to achieve a specific result?
TacticAI was trained using a dataset of 7,176 corner kicks taken in the Premier League during the 2020–2021 seasons. This dataset was randomly shuffled and divided into training and test sets. Details were published in the journal Nature Communications.
Predicting the outcome of corner kicks is difficult due to the randomness of gameplay and the interaction of individual players. The process is also complicated by artificial intelligence modeling due to the limited amount of data available on corner kick tactics — There are only around 10 corner kicks in every Premier League match this season. TacticAI uses deep learning techniques to create more generalized models.
The system models the implicit connections between players by representing the tactical corner kick pattern as graphs, where nodes represent players (with characteristics such as position, speed, etc.) and edges represent connections between them. TacticAI also uses the approximate symmetry of a football field. The system architecture is a version of a convolutional neural network that generates all four possible reflections of a given situation (original, horizontally flipped, vertically flipped, and horizontally and vertically flipped) and brings the predictions to identity in all cases. This approach reduces the search space of possible functions represented by a neural network and produces more generalized models using less training data.
A qualitative analysis of the system, conducted with the participation of football experts, showed that the results obtained by TacticAI were relevant in 63% of cases, which is almost double the benchmark of 33%. Quantitative analysis conducted by the developers confirms that TacticAI's predictions regarding corner kick recipients, kick situations, and player position changes are consistent with those observed in real games.
In the first collaboration between Google DeepMind and FC «Liverpool» of 2021, «Advancing Sports Analytics through AI Research», football was highlighted as a promising area for the application of artificial intelligence. A year later, the Graph Imputer was developed, a prototype forecasting system that allows the use of AI for football analytics. This system is capable of predicting players' movements outside the frame when tracking data is not available.
TacticAI demonstrates the potential of assisted artificial intelligence techniques to dramatically transform sports for players, coaches and fans. Football and similar sports are a dynamic area for AI development because they are highly dynamic and complex multi-agent systems with multi-modal data. The use of artificial intelligence in the field of sports can find application in various fields — from computer games and robotics to traffic coordination.