Auxiliary States for Decentralized Optimization in Probabilistic Communication Networks

Noah Adhikari

EECS Department
University of California, Berkeley
Technical Report No. UCB/EECS-2025-114
May 16, 2025

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-114.pdf

Large optimization problems often require distribution and communication across several nodes. One class of such problems is consensus optimization, where agents must agree upon an optimal solution to these problems in a decentralized manner. Distributed primal-dual methods such as the consensus alternating direction method of multipliers (C-ADMM) are applicable when the communication network is static. However, dynamic communication is less well-studied; prior work has adapted C-ADMM to only a small class of dynamic networks. We generalize C-ADMM further by introducing auxiliary states, which capture information that may affect communication and model communication probabilities as a function of both endogenous and exogenous factors. In addition, we propose a novel, generalized C-ADMM variant, called ASV-ADMM, designed for dynamic communication graphs with auxiliary state-dependent edge transition probabilities. We evaluate ASV-ADMM on several scenarios with state-dependent network topologies wherein agents distributively optimize a global objective.

Advisor: Joshua Hug

\"Edit"; ?>

BibTeX citation:

@mastersthesis{Adhikari:EECS-2025-114,
    Author = {Adhikari, Noah},
    Editor = {Hug, Joshua and Yan, Lisa},
    Title = {Auxiliary States for Decentralized Optimization in Probabilistic Communication Networks},
    School = {EECS Department, University of California, Berkeley},
    Year = {2025},
    Month = {May},
    URL = {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-114.html},
    Number = {UCB/EECS-2025-114},
    Abstract = {Large optimization problems often require distribution and communication across several nodes. One class of such problems is consensus optimization, where agents must agree upon an optimal solution to these problems in a decentralized manner. Distributed primal-dual methods such as the consensus alternating direction method of multipliers (C-ADMM) are applicable when the communication network is static. However, dynamic communication is less well-studied; prior work has adapted C-ADMM to only a small class of dynamic networks. We generalize C-ADMM further by introducing auxiliary states, which capture information that may affect communication and model communication probabilities as a function of both endogenous and exogenous factors. In addition, we propose a novel, generalized C-ADMM variant, called ASV-ADMM, designed for dynamic communication graphs with auxiliary state-dependent edge transition probabilities. We evaluate ASV-ADMM on several scenarios with state-dependent network topologies wherein agents distributively optimize a global objective.}
}

EndNote citation:

%0 Thesis
%A Adhikari, Noah
%E Hug, Joshua
%E Yan, Lisa
%T Auxiliary States for Decentralized Optimization in Probabilistic Communication Networks
%I EECS Department, University of California, Berkeley
%D 2025
%8 May 16
%@ UCB/EECS-2025-114
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-114.html
%F Adhikari:EECS-2025-114