Variance-Reduced Gradient Estimator for Nonconvex Zeroth-Order Distributed Optimization

TL;DR

This paper introduces a variance-reduced gradient estimator for distributed zeroth-order nonconvex optimization, improving convergence and reducing sampling costs.

Contribution

It proposes a novel variance reduction technique combining orthogonal direction renovation and gradient estimation across all dimensions, integrated with gradient tracking.

Findings

Oracle complexity is bounded by O(d/ε) for smooth nonconvex functions.

Oracle complexity is bounded by O(dκ ln(1/ε)) for gradient dominated nonconvex functions.

Numerical simulations demonstrate improved efficiency over existing methods.

Abstract

This paper investigates distributed zeroth-order optimization for smooth nonconvex problems, targeting the trade-off between convergence rate and sampling cost per zeroth-order gradient estimation in current algorithms that use either the $2$-point or $2d$-point gradient estimators. We propose a novel variance-reduced gradient estimator that either randomly renovates a single orthogonal direction of the true gradient or calculates the gradient estimation across all dimensions for variance correction, based on a Bernoulli distribution. Integrating this estimator with gradient tracking mechanism allows us to address the trade-off. We show that the oracle complexity of our proposed algorithm is upper bounded by $O(d/\epsilon)$ for smooth nonconvex functions and by $O(d\kappa\ln (1/\epsilon))$ for smooth and gradient dominated nonconvex functions, where $d$ denotes the problem dimension and $\kappa$ is the condition number. Numerical simulations comparing our algorithm with existing methods confirm the effectiveness and efficiency of the proposed gradient estimator.