Distributed Estimation in Multi-Agent Networks

TL;DR

This paper explores distributed state estimation among interconnected agents with privacy constraints, demonstrating that distributed protocols approach centralized performance as the number of agents grows, especially for Gaussian models.

Contribution

It establishes a theoretical lower bound on sum-rate for distributed estimation and proves the sufficiency of local measurements encoding in such protocols.

Findings

Distributed protocols approach centralized sum-rate with many agents.

Sufficiency of local measurement encoding is proven.

Lower bound on sum-rate is established for Gaussian models.

Abstract

A problem of distributed state estimation at multiple agents that are physically connected and have competitive interests is mapped to a distributed source coding problem with additional privacy constraints. The agents interact to estimate their own states to a desired fidelity from their (sensor) measurements which are functions of both the local state and the states at the other agents. For a Gaussian state and measurement model, it is shown that the sum-rate achieved by a distributed protocol in which the agents broadcast to one another is a lower bound on that of a centralized protocol in which the agents broadcast as if to a virtual CEO converging only in the limit of a large number of agents. The sufficiency of encoding using local measurements is also proved for both protocols.