Non-interactive Remote Coordination

In multi-agent systems coordination is essential to enable autonomous agents to carry out joint tasks. In stochastic environments, this is usually achieved by the agents' actions to approximate a certain desired joint distribution. In this work, we consider coordination among two remote agents, controlled by two controllers through separate rate-limited communication channels. The controllers cannot communicate with each other, but they have access to correlated observations, which can provide a certain level of coordination. Our goal is to explore the effect of limited communication links on the coordination capabilities of the agents. The studied problem can be considered as distributed compression for coordination, with implications in multi-agent reinforcement learning and game theory.