PhD Seminar • Theoretical Neuroscience — Learning to Cooperate: Emergent Communication in Multi-agent Navigation

Ivana Kajić, PhD candidate
David R. Cheriton School of Computer Science

Emergent communication in multi-agent reinforcement learning has been studied to understand properties of resulting languages, as well as to develop communication systems that are suitable for interaction with humans.

Here, using a cooperative navigation task, we show that agents learn a communication protocol that depends on the features of a gridworld environment they're situated in. We analyze the learned protocol, and demonstrate that it efficiently encodes paths in the environment, while exhibiting basic structure.