EE Systems Seminar

ABSTRACT:

We consider the coordination of multiagent systems over point-to-point channels and small multiterminal networks. Rather than distributing explicit messages to coordinate the behavior of different agents, we specify coordination by means of an achievable joint distribution between the actions of the agents and communicate only that amount of information needed to achieve a given joint target distribution. In particular, in this talk we address strong coordination where the target joint distribution and the joint distribution induced by a coordination code are statistically indistinguishable.   We first study the problem of coordination in a three-terminal line network, in which agents use common randomness and provide inner and outer bounds to the coordination capacity region. Specifically, we show that common randomness helps to achieve optimal communication rates between agents, and that matching the network topology to the behavior structure may reduce inter-agent communication rates. We then consider a practical coordination code based on polar codes for a two-node network in which the action imposed by nature at the source node is binary and uniform. By exploiting the connection between channel resolvability and strong coordination and the observation that polar codes are useful for channel resolvability over binary symmetric channels, we prove that nested polar codes achieve a subset of the strong coordination capacity region. Therefore these codes provide a constructive and low complexity solution for strong coordination.