Online distributed optimization algorithm with dynamic regret analysis under unbalanced graphs

We consider online distributed convex optimization problems with a sum of locally dynamic loss functions under unbalanced graphs. When optimizing the dynamic local loss function, every node tracks the time-varying global optimal solution by communicating with neighboring nodes via communication netw...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:Automatica (Oxford) Ročník 174; s. 112116
Hlavní autoři: Yao, Songquan, Xie, Siyu, Li, Tao
Médium: Journal Article
Jazyk:angličtina
Vydáno: Elsevier Ltd 01.04.2025
Témata:
Convex optimization
Distributed optimization
Dynamic regret
Online learning
Online push–pull algorithm
Dynamic regret
Distributed optimization
Online push–pull algorithm
Convex optimization
Online learning
ISSN:0005-1098
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:We consider online distributed convex optimization problems with a sum of locally dynamic loss functions under unbalanced graphs. When optimizing the dynamic local loss function, every node tracks the time-varying global optimal solution by communicating with neighboring nodes via communication networks in a cooperative way. We propose a novel online distributed Push–Pull algorithm and present that the proposed online optimization algorithm can track the dynamic optimal solution with proper step sizes. We analyze the dynamic regret of the proposed algorithm in two cases where the global loss function is strongly convex and smooth, or is convex, smooth and Lipschitz. Our results illustrate that the dynamic regret of the proposed online optimization algorithm can be sublinear, if the path length and the gradient variance are sublinear. At last, we demonstrate the property of the online distributed optimization algorithm by two simulation examples.
ISSN:0005-1098
DOI:10.1016/j.automatica.2025.112116