Analysis of distributed optimization algorithms

Details

Supervisors

Hendrickx

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en mathématiques appliquées, à finalité spécialisée

Abstract

This thesis explores the intricate dynamics of decentralized optimization algorithms, with a focus on consensus matrices and their impact on algorithm performance. Emphasizing key algorithms such as Distributed Gradient Descent (DGD) and the DIGing algorithm, the study addresses fundamental questions related to achieving consensus among computing units and minimizing global objective functions in a decentralized setting. Analyzing spectral properties, convexity, and the implications of consensus matrix weights, the thesis challenges the conventional wisdom of favoring fast-converging matrices in practical scenarios. Comparative studies of consensus matrix weight choices under varying conditions provide valuable insights, contributing to a nuanced understanding of decentralized optimization algorithms and their applicability in distributed computing environments.