Distributed Randomized Gradient-Free Convex Optimization With Set Constraints Over Time-Varying Weight-Unbalanced Digraphs

Abstract

This paper explores a class of distributed constrained convex optimization problems where the objective function is a sum of $N$ convex local objective functions. These functions are characterized by local non-smoothness yet adhere to Lipschitz continuity, and the optimization process is further constrained by $N$ distinct closed convex sets. To delineate the structure of information exchange among agents, a series of time-varying weight-unbalance directed graphs are introduced. Furthermore, this study introduces a novel algorithm, distributed randomized gradient-free constrained optimization algorithm. This algorithm marks a significant advancement by substituting the conventional requirement for precise gradient or subgradient information in each iterative update with a random gradient-free oracle, thereby addressing scenarios where accurate gradient information is hard to obtain. A thorough convergence analysis is provided based on the smoothing parameters inherent in the local objective functions, the Lipschitz constants, and a series of standard assumptions. Significantly, the proposed algorithm can converge to an approximate optimal solution within a predetermined error threshold for the consisdered optimization problem, achieving the same convergence rate of ${\mathcal O}(\frac{\ln (k)}{\sqrt{k} })$ as the general randomized gradient-free algorithms when the decay step size is selected appropriately. And when at least one of the local objective functions exhibits strong convexity, the proposed algorithm can achieve a faster convergence rate, ${\mathcal O}(\frac{1}{k})$. Finally, rigorous simulation results verify the correctness of theoretical findings.