The DeepMind team recently updated their blog on the subject, revealing that the AI crushed its human opponents at Capture the Flag and is now learning the rest of the game. It was decided that 3D FPS games would be the best way to teach the AI “agents” teamwork and cooperation; they were not wrong.
Each AI agent learns through reinforcement with the use of a reward signal, sort of like training a dog with a clicker except your dog isn’t learning how to move efficiently as a member of a death squad. Although, props to you if it is. The AI agents are rewarded when they achieve a goal, which they can set, and for winning. They also learn by playing against each other. Without any instruction, the AI agents figured out how to defend their home base, camp the opponents base, and follow their teammates. It’s just a big ol’ AI assassination training camp. These FTW (For The Win) agents then go on to play humans in procedurally generated environments and take over the world. I’m just assuming that last bit.
According to the science, the AI agents perform better than the humans partly because of their superior reaction times, but even adjusting for this saw the AI win 79% of the time. Apparently, they’re just better at cooperating.
They probably don’t even DM their teammates after a loss, telling them they should die; casuals.