Reproduce “Multiplayer Federated Learning: Reaching Equilibrium with Less Communication” on OpenResearch

About this paper

TaeHo Yoon, Sayantan Choudhury, Nicolas Loizou

This work introduces Multiplayer Federated Learning (MpFL), a framework that models federated learning clients as rational, self-interested players in a game-theoretic setting. It presents PEARL-SGD, an algorithm that enables communication-efficient convergence to a Nash equilibrium or its neighborhood in both synthetic and multi-robot control environments, reducing total communication rounds to "u0398"("u221a"T) in stochastic settings with decreasing step-sizes.

Key insights

• Modeling FL clients as strategic players aiming for a Nash equilibrium expands FL's applicability to competitive multi-agent systems beyond traditional cooperative paradigms.
• Performing multiple local updates (",u03c4" > 1) between communication rounds can significantly reduce communication overhead, especially in stochastic environments.
• The theoretical analysis demonstrates that PEARL-SGD can achieve linear convergence to a neighborhood in stochastic settings, and to an exact equilibrium with decreasing step-sizes, with controlled communication costs.

Summary by alphaXiv.