AI football coach joins Liverpool and improves shooting chances by 13%! From DeepMind, netizen: This is unfair

Mengchen comes from Aofei Temple
Qubits | Public account QbitAI

AI football coach appeared in the Nature sub-magazine, Google DeepMind and Liverpool team collaborated for three years to create:

Just as AlphaGo subverted the game of Go, it changed the way teams formulate tactics.

For example, to whom the offensive team passes the ball is more likely to create shooting opportunities, and how the defensive team adjusts its formation... The efficient tactics easily designed by AI are indistinguishable from real tactics , and human experts favor AI's suggestions in 90% of cases !

Petar Veličković, co-author of the paper , said that football is a more challenging problem than Go.

Football is a dynamic sport, and there are many unobserved factors that can influence outcomes.

Some netizens believe, "If sports can use AI, then all human activities will be able to use AI."

Some people also encourage DeepMind not to be distracted by ChatGPT, and to go in the direction it is good at and one day develop better products.

There are even fans of other teams who are angry that Liverpool was chosen as the partner.

It’s really unfair to speculate whether the founder of DeepMind has selfish motives and uses AI black technology to enhance his favorite team.

To be fair, Arsenal is the one whose home stadium is closest to the DeepMind headquarters (both are in London).

AI takes the corner kick

What is the strength of TacticAI? I thoroughly enjoyed the corner kick mechanism .

The DeepMind team stated that corner kicks are a great opportunity for offense in football matches. According to statistics, 30% of goals come from corner kicks.

For example, in the 2019 Champions League semi-finals, Alexander-Arnold of the Liverpool team suddenly turned around and kicked off the ball quickly, catching Barcelona off guard and was named one of the best corner kicks. Messi was dumbfounded at the time.

(It seems there are many real fans in DeepMind)

Not every player can achieve such wonderful coordination, and whether they can do it depends on whether they are in good condition at the time.

Therefore, TacticAI’s research and development goals aim to solve three core problems :

• What happens for a given corner kick tactic? For example, who is most likely to receive the ball and what is the chance of shooting?

• How to analyze the tactics after they are executed? For example, have similar strategies worked in the past?

• How do you adjust your strategy to achieve specific results? How can the offense increase shooting opportunities, and how should the defense set up their formation?

As for how to solve it, let’s look at some data first.

First of all, TacticAI can predict who of the 22 players in the field is most likely to receive the ball after a corner kick is passed , with an accuracy rate of 78.2% , which is far better than human experts.

This will help the server choose who to pass the ball to.

For the offensive side, passing the ball is not enough. The key is to create shooting opportunities. TacticAI also takes this into consideration.

By analyzing the relationship between catch probability and shot probability, it can predict whether a corner kick will result in a shot with 71% accuracy .

What's even more powerful is that it can also dig out the intrinsic connections between different corner kick tactics and propose targeted improvement measures.

In the end, for the offense, the tactics proposed by the AI ​​increased the probability of making a shot from 18% to 31% .

For the defender, after the AI ​​adjusts the formation, the opponent's probability of shooting is reduced from 75% to 69% .

Just ask which team’s coach can’t be tempted?

Graph neural network + geometric deep learning

So how did DeepMind develop this killer weapon?

Data, collected from more than 7,000 corner kicks in Premier League matches between 2020-2023.

Three core technologies: graph neural network + geometric deep learning + conditional variational autoencoder .

First, express the status of each corner kick as a graph (Graph) .

Each player serves as a node (Node) , and the connections (Edges) between nodes represent possible interactions between players. This graphical representation naturally captures the spatial relationships between players and underlying tactical patterns.

Next, a graph neural network (GNN) is used to learn features in the graph representation.

Through the information transfer mechanism of nodes and edges, GNN can learn high-dimensional latent features of nodes such as players' roles, positions, motion states and other information.

The classic GAT (Graph Attention Networks) model is used here, which uses an attention mechanism common on large models to enhance graph representation learning.

GAT was proposed by the Turing Award winner Bengio team, and the co-author Petar Veličković is also the co-author of TacticAI this time.

To improve data efficiency, TacticAI also uses geometric deep learning to exploit symmetries in football matches (such as the horizontal and vertical symmetry of a square football field) .

By explicitly introducing symmetry constraints into the model, the model can maintain prediction consistency in the face of graph symmetry transformations.

Finally, the generation component uses a conditional variational autoencoder (CVAE) to generate the possible positions and speeds of players under specific tactics.

CVAE is able to learn the underlying distribution of input data and sample from it to generate new data and propose tactical adjustments.

Do all players have to wear AR for training?

The potential of TacticAI goes far beyond this. Once this method is extended to other set pieces and more tactical links, a universal AI football coach may really emerge in the future.

However, the paper does not explicitly mention the running speed of the current system.

Whether it is possible to analyze and give suggestions in real time during the game is a question that many people are concerned about (such as CV master Xie Saining) .

What the majority of fans are more concerned about is that if AI really becomes popular, will the enjoyment of football matches be increased or weakened?

Liverpool FC, the partner of this study, did not respond to requests for comment on whether the AI ​​suggestions have been used in real matches.

However, the intelligence director of the Italian team Atalanta is very optimistic about this technology and believes that compared with the big data analysis that has been widely used before, the suggestions made by AI can also be understood by humans.

AI can help us analyze football in a chunked or categorical way - rather than thinking of everything as just one continuous stream of data that humans can't understand what's going on.

In short, what is more likely to happen in the future is that all athletes will wear AR glasses during training.

Paper address:
https://www.nature.com/articles/s41467-024-45965-x

Reference links:
[1] https://deepmind.google/discover/blog/tacticai-ai-assistant-for-football-tactics
[2] https://www.ft.com/content/e5a64dd3-7fe0-4db4- 9f65-6f7517c2c573
[3] https://x.com/GoogleDeepMind/status/1770121564085707082 ‍ ‍ ‍ ‍ ‍