AI has trained a computer to play football, using digital players who started out able to only make goofy, random movements, but learned to run and kick the ball into the goal.
Researchers at UK-based tech firm DeepMind, owned by the same parent company as Google, used a series of AI methods to build up the prowess of two two-player teams.
Before training, the virtual players, which look similar to those in football video games, would writhe around on the pitch making wiggly, undirected movements. An AI taught the players how to get up and run by mimicking motion-capture data gathered from real footballers.
Advertisement
The digital players improved their abilities on specific training tasks 鈥� such as dribbling a ball to follow a target, or kicking a ball to a target 鈥� through reinforcement learning, which works by rewarding desired actions.
They were then let loose in a series of 45-second matches. Players were rewarded for scoring goals. After 24 hours of training, players learned speedy running, how to keep possession of the ball and how to get up after they had been knocked down.
However, like many young footballers, the trained players all still scrambled after the ball without really thinking ahead. They also tended to dribble and shoot for themselves, rather than working as a team by passing. After further training 鈥� up to 10 days 鈥� the players learned to consider future actions and work as a team. For instance, they learned the benefit of passing the ball to where a teammate was positioned.
Overall, it took three days for the computer to learn the basic rules and mechanics of football. But there are shortcomings: the matches only use two players per team instead of the usual 11, and they don鈥檛 use all the standard rules. Fouls don鈥檛 exist and the pitch has invisible walls around it.
鈥淲hile impressive, the approach is far from automated,鈥� says at IT University Copenhagen, Denmark, adding that it required a carefully designed training regimen and motion-capture demonstrations.
However, Risi says it is 鈥渁n exciting open challenge how we can learn complex tasks such as football end-to-end through more open-ended approaches that would discover the necessary stepping stones by themselves鈥�.
Reference:
