Artificial intelligence (AI) specialist DeepMind has already made systems that can excel at games against human and other machine competitors.
The company’s AIs have beaten human champions at board games such as Go, chess and shogi (also known as Japanese chess), as well as complex video games such as StarCraft II.
All these games seem well suited to AI players as they involve deep strategies informed by clear rules in controlled environments.
Now, DeepMind’s AI has learned to play football using decades’ worth of computer simulations, starting almost from scratch.
Researchers at the company taught the algorithm to develop players through a process that took a digital humanoid and grew it from the equivalent of a baby into a fully-fledged player.
The humanoids had realistic body mass and joint movements in the simulations, meaning that the AI was teaching them to play in close to real-world 3D conditions.
They were first taught to walk and then run in a natural fashion using motion-capture footage of humans playing the game.
They can also be taught to shoot and learn virtual dribbling skills using a trial-and-error kind of machine learning that rewards the AI for keeping the ball close to its player.
A year and half of training was sped through in a day
These first two phases of training represented a year and a half of simulation in real-time, which the AI was able to speed through in just 24 hours.
More complex behaviours also started to emerge after exposing the AI to an additional five years’ worth of simulated matches.
A third training phase ran them through two-on-two, and elements of teamwork, such as anticipating where to move to receive a pass, began to emerge after two to three decades of simulated matches – which took the AI around two to three weeks.
The players were able to learn various aspects of coordination but also movements that were not explicitly set in training drills for the machine learning program.
The success of the programme will not immediately lead to the creation of football-playing physical robots, however.
There is still a relatively long learning time, and the digital players used simplified rules where fouls were allowed, elements of the game such as corners and free-kicks were avoided, and the pitch was surrounded by a wall-like boundary.
The AI team has started to work with real robots pushing a ball towards a target though and also intends to see if the AI training techniques have applications beyond football.
Today’s news was brought to you by TD SYNNEX – the UK’s number one solutions distributor.
