DeepMind teamed up with Liverpool FC to create TacticAI, an AI system set to transform how football teams approach tactics, starting with corner kicks.
Remember the unforgettable moment in the 2019 UEFA Champions League semi-finals, with the quick corner that led to Origi’s legendary goal?
You will if you’re a Liverpool or Barcelona fan – and that’s the kind of play TacticAI aims to create more of, blending the art of human intuition with AI optimization.
TacticAI, detailed in a study published in Nature Communications, uses geometric deep learning to produce models capable of predicting corner outcomes and analyzing tactics.
TacticAI actually started with an earlier project named Game Plan back in 2020, which sought to harness AI for football analysis and strategy, evolving into sophisticated predictive models capable of anticipating off-camera player movements.
For this new study, researchers collected data on 7,176 corner kicks from the Premier League’s 2020 to 2021 seasons. Each corner kick was transformed into a graph representation, where players were nodes featuring data on their positions, movements, and physical attributes at the time the corner was executed.
Researchers then applied geometric deep learning through graph neural networks (GNNs) to analyze spatial interactions and strategies during corner kicks.
Compared to traditional approaches, TacticAI’s predictions were the preferred choice of expert evaluators 90% of the time, and it could also predict both receivers and shot-takers.
At its core, TacticAI serves three primary functions:
- Predicting the likely outcomes of a given corner kick setup, such as identifying potential receivers and estimating shot attempt chances.
- Analyzing the effectiveness of past tactics in similar setups, providing insights into what worked or didn’t.
- Offering tactical adjustments to influence future outcomes, for example, suggesting defender repositions to lower the opposition’s shot attempt chances.
TacticAI is an AI system that can advise football coaches on tactics & plays https://t.co/ghSizpW5iQ This was a really fun project to work on in collaboration with my much loved Liverpool FC @LFC – fingers crossed we win the league this year to give Klopp a fitting send off!
— Demis Hassabis (@demishassabis) March 20, 2024
From there, TacticAI was measured against three benchmark tasks:
- Shot prediction: TacticAI reached an F1 score of 0.71 in predicting shot attempts from corner kicks. An F1 score closer to 1 indicates higher accuracy – so 0.71 is excellent.
- Receiver prediction: This involved determining the player likely first to receive the ball post-corner. TacticAI demonstrated a top-3 test accuracy of 0.782.
- Tactic recommendation: Proposing adjustments in player positions and movements to optimize tactical setups. This is where qualitative expert evaluators generally preferred TacticAI.
Are corners one of the most unpredictable sports set pieces? Perhaps not anymore.
TacticAI might result in more injury-time winners – and right now, Liverpool FC is set to take advantage.
Liverpool fans – you lost Klopp, but at least you have TacticAI.
