AI football coach qualified? Drawing tactics is better than human coaches! It can also "master" other sports

Football is the most popular, most accepted and most widespread sport of all sports.

A corner kick is a way of scoring quickly and directly in football, but it is extremely difficult and requires real-time tactics. It requires not only careful planning by human coaches, but also tacit cooperation among human football players.

Today, artificial intelligence (AI) far outperforms human coaches when it comes to getting corner kicks into the net .

Recently, an "AI football coach" called TacticAI proposed by Google DeepMind, Liverpool Football Club and their partners was featured in Nature Communications, a subsidiary of Nature, with a 90% winning rate .

According to reports, TacticAI can predict corner kick results in football matches and provide practical and accurate tactical advice. The survey results show that Liverpool Football Club's experts choose TacticAI's advice 90% of the time instead of existing tactics from human coaches.

The research team said the research may lay the foundation for the next generation of football AI assistants, helping coaches determine the best player configuration and develop counterattack tactics that are most conducive to winning .

In addition, they believe that this technology may be expanded to other set pieces , such as throw-ins, and other team sports where timeouts can be called.

Who will catch the ball? Can it go in?

Corner kicks are very important in football matches because they can lead to direct goals and give coaches a direct opportunity to intervene and improve game performance.

Therefore, identifying key patterns in the opponent team's tactics and developing effective countermeasures are crucial elements to winning in modern football games.

Especially in real-world situations, where corner kicks are determined before each match, systems that help analyse and improve scoring rates are expected to support human experts well.

However, how to do this algorithmically remains an open research challenge.

In this study, the research team trained TacticAI using a dataset of 7,176 corner kicks from the history of the English Premier League provided by Liverpool Football Club, and identified key strategic patterns that can output predictable and generative outcomes through geometric deep learning technology.

Figure | Bird’s-eye view of TacticAI. (A) How a corner kick situation is converted into a graph representation. Each player is treated as a node in the graph. The graph neural network then operates on this graph by performing message passing; the representation of each node is updated based on the messages sent to it by its neighbors. B) How TacticAI handles a given corner kick. To ensure that TacticAI can respond robustly to horizontal or vertical reflections, all possible combinations of reflections are applied to the input corner kick. These four views are then fed into the TacticAI model to compute the final player representation by influencing each other - each inner blue arrow corresponds to a message passing layer in (A). Once the player representations are computed, they can be used to predict the receiver of the corner kick, whether a shot has been taken, and to make auxiliary adjustments to the player's position and speed to increase or decrease the probability of a shot.

According to the paper, TacticAI consists of two major components: prediction and generation, allowing coaches to effectively sample and explore alternative player settings for each corner kick procedure and select those players with the highest predicted probability of success .

Using this method, TacticAI can accurately predict the first receiver after a corner kick and the immediate outcome of the corner kick.

Figure | Example of using TacticAI to improve corner kick tactics. TacticAI makes it possible for human coaches to redesign corner kick tactics, helping to maximize the probability of a positive outcome for the attacking or defending team by identifying key players and providing time-coordinated tactical suggestions that take all players into account. As shown in this example (A), for a corner kick with a real-life shot attempt (B), TacticAI can generate a tactically adjusted setup to reduce the shot probability by adjusting the defender's position (D). The suggested defender position results in a lower catch probability for attacking players 2-5, while an increased catch probability for attacking player 1 who is farther from the goal post (C). The model is able to generate multiple such scenarios. Coaches can visually view the different options and can also consult TacticAI's quantitative analysis of the proposed tactics.

It is worth mentioning that the research team not only proved that TacticAI can accurately predict the first receiver after a corner kick, the probability of a corner kick directly leading to a shot, and that these tactical settings are feasible, but also asked five football experts (three data scientists, a video analyst, and a coaching assistant of Liverpool Football Club) to determine that it is no different from real-world scenarios.

AI has already been involved in football

In fact, AI’s involvement in football is not unprecedented.

Take Google DeepMind for example. It launched the "AI football player" as early as 2022 and published the relevant research papers in Science Robotics, a subsidiary of Science.

It is reported that this "AI football player" can not only complete a variety of actions such as dribbling and physical confrontation, but also complete accurate shots. Interestingly, two years ago it did not know how to take set pieces such as corner kicks , penalty kicks and free kicks.

In addition, the research team also stated that their method at the time was not suitable for learning directly on robot hardware, and the research results would not be quickly transferred from the simulated world to the real world. However, they believe that their research has promoted AI to move towards human-level sports intelligence.

Back to this study, the research team said that future research will integrate a natural language interface to enable dialogue with the "Football AI Assistant" with the aim of retrieving specific situations of interest, predicting and comparing given tactical variants, and providing guidance through an interactive process to derive tactical recommendations.

In the future, with the further development of AI technologies such as big models, what will human football look like?

It can be said that this is full of imagination.

Reference Links:

https://www.nature.com/articles/s41467-024-45965-xhttps://deepmind.google/discover/blog/tacticai-ai-assistant-for-football-tactics/