An explanation of how AI is changing the world through football

Artificial intelligence has opened up a completely new horizon of technological possibilities that day after day are capable of taking our breath away. All these advances and developments are arriving at breakneck speed, and this means that keeping up to date with the full context is far from being a very difficult task, it is downright impossible.

In order to be a little more up to date with what is happening, and where and how it is being applied, throughout this article we will talk about the solutions that have been found and are being used in a field that is well known and followed by millions of people: soccer.

Nike's spot (2014)

Let's take a look at a very popular soccer spot as a guide for this post. Many of you will remember a very famous ad that Nike released when the 2014 World Cup was held in Brazil. In this one, the stars of the moment were replaced by humanoids with perfect skills to play soccer.

The humanoids that I have mentioned we are going to consider, from now on, that they are the AIs of the spot. And this was the first fantasy that, today, we can say for sure has been fulfilled from the advert.

Clip from the Nike spot "The Last Game", in which the creator of the artificial intelligence that replaces real players speaks in statistical and quantitative terms of known players, in the case of the image, Zlatan Ibrahimović.

The second thing that has also been fulfilled is more quantitative; the exhaustive analysis of real players, the extraction of statistics and predictions based on their game.

We are going to dig through the following lines to find out how this point has been reached, which in 2014 (it seems like yesterday!) seemed so futuristic and far away. We are going to analyze how the application of AI has changed the way we enjoy soccer matches and the way the industry works in this sport that is so widely followed throughout the world.

(Real) robots playing soccer (2023)

The first of the surprises with which, in soccer, reality has surpassed fiction, came this year 2023. At the beginning of the year, Google DeepMind published a scientific article with complementary videos in which a huge milestone was reached: the learning by a robot of a complex sport such as soccer.

The company investigated whether Deep Reinforcement Learning can teach a robot complex and safe skills to play soccer. These agents were trained in a simulated self-learning environment, which was then extrapolated to a real environment.

Deep Reinforcement Learning training of a DeepMind AI in a simulated environment.

One of the difficulties of moving from a simulated environment to reality is that... it is impossible to simulate reality as it is 100%. In the simulated environment, the agent learned dynamic movements, basic game strategies and Google DeepMind managed to transfer those skills to real robots, which managed to play soccer.

DeepMind robots playing soccer in a real environment.

Before moving on, we mentioned a type of learning, Deep Reinforcement Learning, but what exactly is it? In short, we could say that it consists of machines learning something specific on their own based on certain rules.

The key elements of Deep Reinforcement Learning include:

• Agent: It is the decision-making entity that interacts with the environment. It can be a computer program, a robot or any other system capable of perceiving its environment and taking actions.
• Environment: It is the context in which the agent operates. It can be real or simulated.
• Actions: Are the decisions that the agent can take at a given time.
• State: It is the current representation of the environment at a specific time.
• Reward: It is the feedback signal that the environment sends to the agent after it performs an action. The agent's goal is to maximize the reward accumulated over time.
• Policy: The strategy that the agent uses to select actions in a state.
• Value function: Evaluates the utility or expected value of a state or an action in terms of the future rewards expected to be received.

In short and very simply explained,

Deep Reinforcement Learning could be said to be the way a child learns.

Starting from an established objective, it seeks to execute actions in an environment that bring the child closer to or further away from the objective. Moving closer to the target is compensated (dopamine in the case of humans, in most learning processes), and moving away from the target is penalized.

✅ In the example of the child, if we analyze the process of learning to walk, we could say that, in an environment such as the living room of a house, a park, etc., a child (agent), receives penalties, pain due to falls, for example, if he/she does not manage to approach the target. If it manages to get closer to the target or is even able to walk, it receives rewards (applause, kisses, encouragement, satisfaction, etc.).

This type of learning, using algorithms, can of course be extrapolated to machines.

In the following table we can see how the rewards are according to the consequences of the actions (controlling, advancing, dribbling, dribbling, shooting, etc.) that the agents (robots) perform in the environment (soccer field). It can be seen very clearly that scoring a goal is the maximum reward.

Fragment of the compensation policy table from the paper Learning Agile Soccer Skills for a Bipedal Robot with Deep Reinforcement Learning by Google DeepMind.

While it is true that Google DeepMind's humanoid robots are a far cry from the AIs in the Nike spot... I wouldn't be so quick to score a goal against the human species! It is worth remembering that in games as complex and famous as chess or the game Go the probability of beating a machine of a professional is, today, very small.

AI also trained with Deep Reinforcement Learning, which beat the world Go champion, also bears DeepMind brand.

The application of advanced analytics to football

At this point we have covered one of the fulfilled prophecies of the announcement, AIs that are able to play soccer, but what about today's numerical deconstruction of players, predictions, statistics, etc.?

Deconstruction into numerical indicators of a player is something that has been done for decades, but the classical methods are far from the way it is done today.

Player statistics have historically been compiled by people by hand and by eye. This process, of high value for the clubs, has nowadays been automated in a paroxysmal way thanks to artificial intelligence, achieving a total and absolute scope of all players, all leagues, etc.

What are the limits to the scope of this process today? As far as a video camera can go, in other words, as far as a smartphone can go.

There are numerous companies that offer services to soccer clubs based on the rich and diverse information they have in their databases. Information gathered by recording hundreds and hundreds of matches, from a variety of leagues and categories, from kids to professional soccer.

Their goal? To find the diamond in the rough, to find the promising young footballer ahead of the competition. Or find a player who can fill in for an injury replacement, sale, etc., who is as similar as possible (mathematical similarity here) to the injured or sold player. Or simply find a player who meets a set of exact characteristics set out by a team's coach. Analysis of strengths or weaknesses of teams (own or opposing), etc.

✅ All this is achieved thanks to the mixture of two booming technological disciplines, Computer Vision and deep learning, which is, in very generic terms, a subdivision of machine learning that uses the power of the neural network algorithm to address predictive tasks mainly related to images or text.

Consequently, a computer can detect that a set of pixels is a number and in turn that this number is a 10, for example (this could be used to identify backbones, right?). Likewise, these algorithms can identify that a set of pixels that make up a round body is a soccer ball, etc.

Deep Learning can identify all the elements of a match for future analysis.

Deep Learning algorithms enable frame-by-frame analysis of a match recording and are able to interpret what is happening at any given moment, for example:

• Which player is who (identifying the numbers of the bibs, for example).
• Who has the ball.
• At what speed each player moves.
• Which are the areas in which players usually have more presence.
• How many passes (effective and ineffective) each player makes.
• ...
• The limit of KPIs lies in creativity.

✅ Many clubs in the main European leagues already rely on this type of software to make data-driven decisions, with excellent results.

Conclusion

There is no doubt that the capacity to surprise in terms of advances in artificial intelligence has no limits. In less than a decade it has gone from the science fiction of an advertising spot to real materialization.

Such rapid and sudden advances, which in many cases can be life-changing, can at first sight be overwhelming and can cause mistrust.

And so as not to panic those who view AI advances with fear and skepticism, here is a little spoiler from the Nike commercial:

Humans end up beating more than 100 artificial intelligences in a soccer match together.