Stochastic Activation based Broadcast Push-Sum for Distributed Estimation

Distributed estimation systems enable nodes to estimate a target parameter in a collaborative manner. These systems are useful in sensor networks or distributed machine learning. Here, we explore distributed estimation in graph-connected networks without a fusion center, where nodes exchange information with neighbors to estimate this target parameter synchronously. Due to packet collision, there is a tradeoff between the number of exchanges and the quality of these exchanges. To fix this issue, we propose to activate the nodes randomly. The main contribution of the paper is to determine an activation rate offering a good target estimation quality as fast as possible.