On the Role of Interaction Between Sensors in the CEO Problem

The CEO problem has recently received attention as an abstraction of remote estimation in sensor networks. It models a group of sensors who observe independently corrupted versions of a hidden source which a central unit (CEO) is interested in estimating. The agents encode their observations without cooperating with one another and communicate the encoded bits through rate constrained noiseless channels to the CEO. In this talk we explore the possibility of co-operation between sensors. In particular, we consider two forms of co-operation. In one scenario, the sensors can communicate with each other over rate constrained noiseless channels, and in the other, the sensors broadcast their messages to all the other sensors and the CEO. In both scenarios, the sensors take turns sending messages over multiple rounds of communication. We show that, while in general both setups may lead to a reduction in the sum-rate of communication over the CEO problem, for the Gaussian problem there is no advantage. We also derive a single-letter lower bound for the sum-rate in the broadcast scenario.