Football is a simple thing, as long as you are a human being. Teaching software how to face an opponent, where to position other players, when to pass, sprint or cross is more difficult than teaching it to beat a world chess champion. The machine learning algorithms applied to football simulations have undergone a very rapid refinement process over time, also due to the global success of these video games. Over time titles like Pro Evolution Soccer and Fifa have become privileged playgrounds for training some forms of artificial intelligence. With evident and progressive results from year to year.
In fact, in the nineties playing against the computer was not particularly challenging also because after a couple of games the system was “hackable”. In short, it was possible to find the weak point, either in the balance of the players (some athletes were impregnable) or in the defensive tactics (by making certain movements one could confuse the opponents and the goal was almost guaranteed). Today the “bugs”, the errors, are still there, but it is more difficult to identify them.
Loading…
“I think the biggest challenge is to create players who can act as humans would,” he explains Sole 24 Ore Fabio Zinno, senior software engineer of Ea, the developer of Fifa, the best-selling video game in Italy and the most famous in the world. . “We’ve seen a lot of academic research generating AI players with super human abilities, but they would definitely fail a Turing test. You can train an autonomous agent to play Fifa and win every time, but if his movements are unnatural, irregular and unpredictable, you probably haven’t managed to create a good game AI ». The goal should not be to win at all costs, but to create an engaging, stimulating and overall fun experience for our players ».
Inside Ela, he explains, there are several groups that develop machine learning algorithms, including core technology teams. Several development teams, including FIFA, Madden, UFC and NHL, are also carrying out their own Machine Learning projects, the architects of some solutions already used within the game and others that will soon follow. “In a nutshell – he adds – at EA we believe that Machine Learning should not be treated as a “secret art”, with a group of experts separated from the development process: on the contrary, we promote a culture that strives to make Machine Learning simply another “tool in the toolbox “for all our Software Engineers and Technical Artists”.
What kind of machine learning are they studying?
“We are very committed to creating smarter tools for creating content, including 3D models and textures, terrains, stadiums, cityscapes, face and body animations. A considerable effort is heading towards the motion capture of the face and body movements starting from the videos, a technology that will not only give us a greater volume of animations, but also much more organic and natural results: as much as the animations acquired with the traditional motion capture process in the studio are of a good standard, these are achieved using markers and props which can sometimes hinder natural movement. Being able to capture the movements of athletes “in their natural environment”, especially with regard to sports titles, is a very attractive goal. Machine learning is also making possible more holistic solutions for character performance and we are working on models capable of generating speech from text, along with lip sync and facial animations. The goal is to include body gestures in the future, to provide realistic and convincing behaviors.