TacticAI: Google built an artificial tactical assistant for Klopp, and it works quite well | Soccer | Sports


The experiment in which artificial intelligence (AI) has gone furthest in understanding the dynamics of football began with a neighbor conversation. In 2018, Julian Ward and Karl Tyls thought it would be cool to do something together. Ward was responsible for Liverpool’s loan program and Tyls worked at DeepMind, Google’s AI division. They organized a meeting in Melwood, the club’s old sports city, where they were received by Liverpool’s research director, Ian Graham: “The big idea was: What if AI could replace the coach?” he recalls via a videocall. “But we didn’t want to be too controversial, so we left it to see if the AI ​​could be an assistant coach that would advise the coach.”

Almost eight years later, that conversation between neighbors has led to the publication in the journal Nature Communication of the details of the construction of TacticAI, an artificial intelligence tool capable of helping Jürgen Klopp prepare corner plays. The system is capable of analyzing and distilling hours and hours of videos of corner kicks and providing suggestions indistinguishable from human ones. 90% of the time Liverpool analysts preferred the machine’s ideas over what had happened in reality.

Before building this tool, they explored broader ideas about what AI could achieve with football. They decided that the first challenge for AI would be some kind of game. They stopped the video of a match and asked him to guess how the footballers were going to continue moving, according to Graham, who has a doctorate in Theoretical Physics from the University of Cambridge: “With the players who are close to the ball, the predictions are really good. during the first ten seconds. If they are far away, he could deviate by about ten meters, but for those players it is not so important if he predicts it wrong.” This resulted in a publication in the journal Scientific Reports, part of the Nature family. They had managed to train the artificial intelligence to acquire a quite remarkable understanding of the complex dynamics that operate in a match.

Then, as Graham explains, they looked for ways to take advantage of what they had learned: “Corners are a controlled situation. It is very disorderly, there are many players in the area, grabbing each other’s shirts, elbowing each other… but we know that the ball is stopped and they are going to center it into the area.” There was also the possibility of a prize: “At Liverpool we knew that we were going to have seven or eight corners per game. “If we can maximize our scoring opportunities, if that gives us one or two wins a season, it can be the difference between coming first or second.”

To address the problem, DeepMind sent Zhe Wang, an engineer with research in robotics who came from teaching humanoids to play soccer, and Petar Velickovic, one of the people who developed the model that allows Google Maps to calculate how long it takes to go from one place to another. “It’s a journey to get out of our research lab and apply these technologies to solve real-world problems. “It’s fascinating,” says Zhe over video call.

The beginnings were somewhat disappointing for Graham: “At first the AI ​​only said pretty obvious things. We said, okay, understand what an open corner and a closed corner are, and that the ball goes away from the goal in the first one. But it wasn’t his fault. If you don’t ask him the right questions, he won’t give you the right answers. Seeing the differences between the corners that open is much more interesting.”

They trained the system until it learned to do three things: review and catalog all the corner videos provided to it, predict what is going to happen in each one depending on the positioning of the players (who will receive the ball, if there may be a shot) and generate alternatives to deal with each case (placing a defender one meter further forward, for example).

One of the examples in which TacticAI suggests adjustments to the position of the players to defend a corner and thus reduces the probability of the attackers receiving the ball.

For Velickovic, the main advantage of TacticAI, whose engine they have released and would be useful for other team sports, is the burden it alleviates: “Video analysts have to spend less time analyzing patterns. We humans can do it well, but not fast.” It also represents a shortcut when looking for options to damage the opponent or defend against him: “Humans can find better adjustments, if they think for longer, but TacticAI can give you something that 90% of the time is better, and it gives it to you in seconds. “This way you have more time for the creative part.”

Graham also sees that as the space of the human mind: “AI learns from the data it has seen, so it will only produce things similar to what it has seen. Not something completely new. “I like crazy corners.” Like that surprise serve with which they eliminated Barça in the second leg of the 2019 Champions League semi-final, after losing 3-0 at the Camp Nou. Alexander-Arnold serves very quickly and Origi scores before the Blaugrana realize it. They followed a plan: James French, the rivals’ analyst, had noticed while watching videos that when a corner was awarded, the Barça players were distracted by protesting to the referee, talking among themselves… And they took advantage of it. TacticAI would not have had that idea.

You can follow The USA Print in Facebook and xor sign up here to receive our weekly newsletter.


Leave a Reply