Game publishing giant Electronic Arts (EA) has created a self-learning AI agent that can successfully engage in multiplayer matches.
At the recent Game Developers Conference in San Francisco, EA’s game tech research division SEED presented footage of AI agents that had been trained to play a basic mode of the publisher’s popular World War I first-person shooter Battlefield 1.
“All of the agents you see are controlled by a single neural network that’s been trained to play the game from scratch through trial and error,” EA explains in the video. “Everything the agents do is the result of previous gameplay experience. We only give them encouragement for playing the objective.”
In the footage, the agents can be seen running around and shooting one another, even managing to kill enemies that quickly run in front of them from hidden alleyways.
To help the agents adapt to combat, EA placed boxes that can restore health and ammo into the map, teaching the AI to pick them up when needed.
EA acknowledges that it’s AI-controlled characters aren’t perfect — at certain points in the match, they can be seen running around in circles.
Still, the addition of AI that is trained to shoot and adapt like a human player would is a marked evolution of the non-player character bots that populate some online games.
Image credit: EA