Matthias Sammer is to be envied. Borussia Dortmund leads the table in the Bundesliga. As a TV expert, the BVB consultant is popular with viewers. Regularly he acts tactics and touch screen for Eurosport. He shifts magnets, lets play scenes run back and forth, marks actors. What is there for patterns? Who is where? Who should have been anywhere else in the crucial scenes?

For years, neurons have networked in his brain, enabling him to answer such questions. Researchers have been attempting to replicate such neural networks artificially for several decades. As early as 2017, a research group from Disney Research presented an approach that deals explicitly with the above-mentioned questions and has been constantly being developed ever since.

Using so-called spirits who have learned a certain behavior, it should be shown how the defending team should have stood or walked in a particular enemy attack. For the spirits to learn this behavior, the researchers developed a process they call "Deep Imitation Learning."

Three million different times analyzed

As a basis, Disney Research took the position data of 100 Premier League games, from which they released a total of more than 17,400 defense sequences. These scenes again gave the researchers about three million different times for the spatial distribution of professionals including ball on the field. From this data, the computer should independently learn to predict the following behavior based on the coordinates of the players and the ball.

Since not only a time should be learned, but rather a sequence, recurrent neural networks were used. They are able to understand relationships over several times. For example, only the long-short-term-memory technique used here led to the breakthrough in the field of speech recognition. Watson, Siri and Alexa say hello.

In addition, the researchers from Disney Research let the neural network learn from its own mistakes. For each time point, the deviation of the player's predicted coordinates to the actual position was calculated. These data were then used directly to fit the model before predicting the next time. This allows the model to recover from errors and better predict the future events of a sequence. This process was now carried out first for the ten outfield players and then for all field players of the defending team together.

photo gallery


8 pictures

Ghosting in football: Eleven ghosts you should be

But is it really possible to show how the defending team should have acted positionally? The visualizations presented by the researchers are quite impressive, as our photo series shows. For a Fulham attack against Swansea, the ghosts show how either a defending Premier League team or Manchester City would have behaved in particular. Whether this simulated behavior is better than what Fulham actually showed in this scene is determined by the expected goal probability.

Ghosts in the NBA already successfully in use

In addition, it is already possible to take a look into the future. Using a valid model for mapping the goal probability from positional data, the ghosts could learn to behave in such a way that the probability of the goal is minimized. From that point on, you can really speak of showing whether a team behaved properly in a defensive action and how it could have behaved better.

In the NBA, Ghosting has been used to describe defensive behavior since 2013, when the Toronto Raptors opened the doors of its analytics department for the now discontinued blog project "Grantland." In basketball also strictly predetermined moves are basically possible, which is why there are already the first models that try to predict the expected defensive behavior for certain moves.

This predictive use is unlikely in football, as during the games no moves are implemented in the strict sense, so with meticulously given running and pass paths. For standard situations, it would be conceivable to cut back if the planned ball paths are complied with. The most effective would be the use of the method in penalties, which raises the question of how much added value this would bring over the Lehmann's slip.

It is quite possible, however, that the method of trainers and TV experts will be used later for analysis. The role of Matthias Sammer would change to increasingly describe how and why the spirits behave in a certain way. Graphics could make the complex tactical processes more tangible to the simple TV viewer, revolutionizing football coverage.