EA Created An AI That Taught Itself To Play Battlefield

Illustration for article titled EA Created An AI That Taught Itself To Play Battlefield

Electronic Arts’ Search for Extraordinary Experiences (SEED) Division—which sounds like an anime space force—has created a “self-learning AI-agent” that has managed to teach itself how to play Battlefield 1 multiplayer.


In this blog post, Magnus Nordin from SEED details how his team, inspired by Google’s work with old Atari games, wondered “how much effort it would take to have a self-learning agent learn to play a modern and more complex first person AAA game like Battlefield”.

So they tried to find out. The results are an “agent” that, while inferior to human players, “is pretty proficient at the basic Battlefield gameplay”. The agent changes behaviour if it’s low on health or ammo, and while more complex behaviours like knowing the details of each map are beyond it (at the moment), EA has found that “while the human players outperformed the agents, it wasn’t a complete blowout by any stretch.”

You can see some agents in action below.

While there’s some good stuff on show there, you can also see some of the agent’s limitations, especially when it comes to thinking ahead. While they’re capable of acting immediately in response to threats, when left alone they “will eventually start to spin around to look for something to do”.

Which is funny now, but these are self-learning agents. Every time they play, they’re getting better, and it’s only a matter of time before they start reaching the point where it’s tough for a player to determine who’s a real player and who is just some AI running around in a uniform.

The potential for these agents is two-fold: Nordin not only sees them as a superior replacement for bots, but as useful for quality assurance and testing as well.

Luke Plunkett is a Senior Editor based in Canberra, Australia. He has written a book on cosplay, designed a game about airplanes, and also runs cosplay.kotaku.com.


I’d like to see an AI be given actual motor controls, like a mechanical hand, where they then have to learn how to use a controller in order to properly play.

It would be interesting to see them given physical restrictions that human players have to play with as compared to straight software, practice-until-it’s-an-aimbot capabilities.