We all know that artificial intelligence will eventually take over the world and murder as all, but before that happens, they’re prepared to beat us in our favorite video games. Electronic Arts’ Search for Extraordinary Experiences Division — or SEED — partnered with Battlefield 1 studio Dice to teach A.I. to play the multiplayer shooter, and they aren’t half bad.
A self-learning artificial intelligence system controlled all players in a huge Battlefield 1 multiplayer match, fighting each other and capturing objective points as if they were real humans. When low on ammunition or health, they were able to replenish their supplies by running over colored cubes located over the map. This was all done with trial and error, and machine learning over a neural network, and it’s pretty tough to tell there aren’t real players behind each character. Before jumping into the action themselves, the agents observed others playing the game for 30 minutes. Playing “on several machines in parallel” afterward, the agents each have about 300 days of experience with the game.
One thing you’ll notice from the development video is the A.I. agents don’t like aiming down the sights with their weapons. Instead, they pick off targets with shots from the hip and have remarkable target acquisition, but it isn’t a simple aim-bot took — they also miss several shots before they take down their target.
“While their basic skills are impressive, they get easily confused and have many things left to learn,” the video’s narrator says as we see a group of agents running in a tight circle continually.
Perhaps it’s best that artificial intelligence is still running into some of these problems, because it gives us more time to prepare for their inevitable uprising.
EA and SEED will continue developing their neural network and machine learning systems in order to assist in the development of games. Technical director Magnus Nordin said the technology will benefit the quality assurance departments at studios, and it will eventually result in “truly intelligent” nonplayer characters instead of the relatively static ones we see in games today. Machine learning will also be applied to animations and voice recognition in order to produce more natural procedurally generated content.
