ZO-JADE: Zeroth-order Curvature-Aware Multi-Agent Convex Optimization

Alessio Maritan,Luca Schenato

DOI: https://doi.org/10.1109/LCSYS.2023.3281745

2023-06-13

Abstract:In this work we address the problem of convex optimization in a multi-agent setting where the objective is to minimize the mean of local cost functions whose derivatives are not available (e.g. black-box models). Moreover agents can only communicate with local neighbors according to a connected network topology. Zeroth-order (ZO) optimization has recently gained increasing attention in federated learning and multi-agent scenarios exploiting finite-difference approximations of the gradient using from $2$ (directional gradient) to $2d$ (central difference full gradient) evaluations of the cost functions, where $d$ is the dimension of the problem. The contribution of this work is to extend ZO distributed optimization by estimating the curvature of the local cost functions via finite-difference approximations. In particular, we propose a novel algorithm named ZO-JADE, that by adding just one extra point, i.e. $2d+1$ in total, allows to simultaneously estimate the gradient and the diagonal of the local Hessian, which are then combined via average tracking consensus to obtain an approximated Jacobi descent. Guarantees of semi-global exponential stability are established via separation of time-scales. Extensive numerical experiments on real-world data confirm the efficiency and superiority of our algorithm with respect to several other distributed zeroth-order methods available in the literature based on only gradient estimates.

Optimization and Control,Dynamical Systems

What problem does this paper attempt to address?

The problem that this paper attempts to solve is convex optimization in a multi - agent environment, especially when the derivative of the objective function cannot be obtained (such as in black - box models). In addition, each agent can only communicate with local neighbors according to the network topology. Specifically, the paper focuses on the problem of minimizing the mean of local cost functions in a multi - agent setting, where the derivatives of these local cost functions are unavailable. The paper proposes a new algorithm - ZO - JADE. It estimates the curvature of local cost functions by finite - difference approximation, thereby accelerating the convergence speed, and proves the exponential stability of this algorithm in a semi - global range. Experiments show that the ZO - JADE algorithm exhibits higher efficiency and superiority among multiple distributed zero - order optimization methods.