Distributed Alignment Processes with Samples of Group Average

TL;DR

This paper proves that a distributed weighted-average algorithm optimally minimizes agents' deviation from the average in noisy, multi-agent systems, matching centralized performance despite limited information processing.

Contribution

It introduces and analyzes a distributed weighted-average algorithm that is proven to be optimal for minimizing deviation in noisy multi-agent systems, even matching centralized solutions.

Findings

Weighted-average algorithm minimizes deviation optimally.

Distributed approach matches centralized performance in idealized noise model.

The drift in the distributed setting is close to the best centralized drift, with small overhead.

Abstract

Reaching agreement despite noise in communication is a fundamental problem in multi-agent systems. Here we study this problem under an idealized model, where it is assumed that agents can sense the general tendency in the system. More specifically, we consider $n$ agents, each being associated with a real-valued number. In each round, each agent receives a noisy measurement of the average value, and then updates its value, which is in turn perturbed by random drift. We assume that both noises in measurements and drift follow Gaussian distributions. What should be the updating policy of agents if their goal is to minimize the expected deviation of the agents' values from the average value? We prove that a distributed weighted-average algorithm optimally minimizes this deviation for each agent, and for any round. Interestingly, this optimality holds even in the centralized setting, where a master agent can gather all the agents' measurements and instruct a move to each agent.We find this result surprising since it can be shown that the total measurements obtained by all agents contain strictly more information about the deviation of Agent $i$ from the average value, than the information contained in the measurements obtained by Agent $i$ alone. Although this information is relevant for Agent $i$, it is not processed by it when running a weighted-average algorithm. Nevertheless, the weighted-average algorithm is optimal, since by running it, other agents manage to fully process this information in a way that perfectly benefits Agent $i$.Finally, we also analyze the drift of the center of mass and show that no distributed algorithm can achieve drift that is as small as the one that can be achieved by the best centralized algorithm. In light of this, we also show that the drift associated with our weighted-average algorithm incurs a relatively small overhead over the best possible drift in the centralized setting.