Quantized Zeroth-Order Gradient Tracking Algorithm for Distributed Nonconvex Optimization Under Polyak-Łojasiewicz Condition

This article focuses on distributed nonconvex optimization by exchanging information between agents to minimize the average of local nonconvex cost functions. The communication channel between agents is normally constrained by limited bandwidth, and the gradient information is typically unavailable....

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:IEEE transactions on cybernetics Ročník 54; číslo 10; s. 5746 - 5758
Hlavní autoři: Xu, Lei, Yi, Xinlei, Deng, Chao, Shi, Yang, Chai, Tianyou, Yang, Tao
Médium: Journal Article
Jazyk:angličtina
Vydáno: United States IEEE 01.10.2024
Témata:
Convergence
Cost function
Deep learning
Distributed algorithms
Gradient methods
Gradient tracking algorithm
linear convergence
nonconvex optimization
Quantization (signal)
uniform quantizer
zeroth-order algorithm
ISSN:2168-2267, 2168-2275, 2168-2275
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:This article focuses on distributed nonconvex optimization by exchanging information between agents to minimize the average of local nonconvex cost functions. The communication channel between agents is normally constrained by limited bandwidth, and the gradient information is typically unavailable. To overcome these limitations, we propose a quantized distributed zeroth-order algorithm, which integrates the deterministic gradient estimator, the standard uniform quantizer, and the distributed gradient tracking algorithm. We establish linear convergence to a global optimal point for the proposed algorithm by assuming Polyak-Łojasiewicz condition for the global cost function and smoothness condition for the local cost functions. Moreover, the proposed algorithm maintains linear convergence at low-data rates with a proper selection of algorithm parameters. Numerical simulations validate the theoretical results.
Bibliografie:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 23
ISSN:2168-2267
2168-2275
2168-2275
DOI:10.1109/TCYB.2024.3384924