Artificial Intelligence Google DeepMind will try to play football

DeepMind, a subsidiary of Google, has created a football simulator for testing machine learning algorithms. A video game called the Google Research Football Environment allows researchers to test algorithms in a physics-based and easy-to-use world.   It was no coincidence that the football simulator was chosen as an educational platform for AI. According to an article on the website of the Massachusetts Institute of Technology, one of the tasks for AI researchers is to find new problems for machine learning algorithms that they could solve and thus be trained. Simple video games, such as Pong or Breakout, are sometimes too primitive and predictable for these algorithms. Other games, such as Starcraft, are too complex. Starcraft, a real-time strategy that unfolds in a large online universe, is so vast and confusing that it requires tremendous computing resources to gather relevant data and learn.   Then football simulations come to the rescue. They have certain levels of predictability based on the physics of the game. But they also have many unpredictable moments that arise due to the tactics of the players and which allow the machine to learn something new.   As a base, DeepMind researchers used a game called Gameplay Football, which allows for full matches with fouls, corners, penalties, offsides and so on.   “The game is a physics-based soccer simulator where the agent needs to control his players and learn how to overcome the opponent’s defense in order to score goals,” say Google employees.   Today, DeepMind neural networks can play many games, such as Atari video games such as Pong, Breakout and Space Invaders, as well as more complex ones, including the already mentioned Starcraft. At the beginning of 2019, AlphaStar-developed DeepMind artificial intelligence was able to beat people in Starcraft.   DeepMind has also made significant strides in traditional games. In 2016, DeepMind AlphaGo defeated one of the world’s strongest professional go players. The game ended with the victory of AlphaGo with a margin of half a point.