Electronic Arts has begun testing self-teaching bots in Battlefield 1 in an attempt to find the next big leap forward in gaming AI. Partly inspired by Google's Deepmind AI, the project involved EA's SEED division showing the 'agents' 30 minutes of human play (which they could imitate later on) before setting them loose in the game for six days of self-teaching.
The results, shown in the video above, are interesting. The agents sometimes look like a co-ordinated unit, and have learned to restock on ammo if they're running low. However, you'll also see them spinning randomly around in circles. "The agents aren’t very good at planning ahead. If an agent spots an objective, like an enemy player, it will act. But if there’s nothing in sight, it will eventually start to spin around to look for something to do," says SEED's technical director Magnus Nordin on EA's website.
The point of the exercise is, eventually, to help developers build AI that is more satisfying to play against. In tests, the agents were beaten by human opponents in a game mode restricted to hand guns, but "it wasn’t a complete blowout", which the team thinks is encouraging.
This kind of thing has been tested in other games before, but EA reckons it's the first time it's happening in a shooter as complex as Battlefield 1. The full piece on EA's site is actually a pretty good read: Nordin reckons that self-learning AI will make its way into games sooner rather than later, and that eventually a self-taught agent will be able to beat a professional FPS player in a firefight. That'll be a big moment.
And if this sort of thing interests you, here's Alex's brief history of AI beating humans.
