Google DeepMind and Premier League club «Liverpool» have created a prototype of the TacticAI artificial intelligence model for football training. The DeepMind lab, which has previously used AI for complex board games such as Go, notes that AI can be trained on patterns that can be seen on sports fields.
According to an article published in NatureThe artificial intelligence suggestions for improving players’ positions during corner kicks, a major potential source of goals, have been largely approved by human experts. Currently, the technology can provide advantages in planning situations with predictable initial conditions, such as corner kicks. The global challenge is to apply AI to more variable open play conditions.
The DeepMind project is the result of three years of work with «Liverpool». The DeepMind model uses deep geometric learning on a dataset of 7176 English Premier League corner kicks between 2020 and 2023. The model learned to predict which player will be the first to touch the ball after a corner kick. In the tests, the one who receives the ball was among TacticAI’s top three candidates 78% of the time. Corner kicks, together with free kicks, statistically account for up to 30% of all goals.
TacticAI analyzed the results of corner kicks with different player configurations and suggested positional improvements. Their usefulness was evaluated in a blind study conducted by five experts: three data scientists, one video analyst, and one assistant coach. The experts could not distinguish between the AI-generated scenarios and real match situations, and they preferred TacticAI’s advice to their existing strategies in 90% of cases.
Petar Velickovic from Google DeepMind says that the research has broader applications beyond sports. «If we can model the game of soccer, we can also model several aspects of human psychology. As AI becomes more capable, it will need to understand the world better, especially under conditions of uncertainty. Our system is capable of making decisions and suggestions under conditions of uncertainty. These skills, we believe, can be transferred to future AI systems, so this is a good testing ground,» he says.
Sources: Financial Times, New Scientist