Zeroth-order Gradient and Quasi-Newton Methods for Nonsmooth Nonconvex Stochastic Optimization (2401.08665v1)

Abstract: We consider the minimization of a Lipschitz continuous and expectation-valued function defined as $f(\mathbf{x}) \triangleq \mathbb{E}[{\tilde f}(\mathbf{x}, \boldsymbol{\xi})]$, over a closed and convex set. Our focus lies on obtaining both asymptotics as well as rate and complexity guarantees for computing an approximate stationary point (in a Clarke sense) via zeroth-order schemes. We adopt a smoothing-based approach reliant on minimizing $f_{\eta}$ where $f_{\eta}(\mathbf{x}) = \mathbb{E}_{\mathbf{u}}[f(\mathbf{x}+\eta \mathbf{u})]$, $\mathbf{u}$ is a random variable defined on a unit sphere, and $\eta > 0$. It has been observed that a stationary point of the $\eta$-smoothed problem is a $2\eta$-stationary point for the original problem in the Clarke sense. In such a setting, we develop two sets of schemes with promising empirical behavior. (I) We develop a smoothing-enabled variance-reduced zeroth-order gradient framework (VRG-ZO) and make two sets of contributions for the sequence generated by the proposed zeroth-order gradient scheme. (a) The residual function of the smoothed problem tends to zero almost surely along the generated sequence, allowing for making guarantees for $\eta$-Clarke stationary solutions of the original problem; (b) To compute an $\mathbf{x}$ that ensures that the expected norm of the residual of the $\eta$-smoothed problem is within $\epsilon$ requires no greater than $O(\eta^{-1} \epsilon^{-2})$ projection steps and $ O\left(\eta^{-2} \epsilon^{-4}\right)$ function evaluations. (II) Our second scheme is a zeroth-order stochastic quasi-Newton scheme (VRSQN-ZO) reliant on a combination of randomized and Moreau smoothing; the corresponding iteration and sample complexities for this scheme are $ O\left(\eta^{-5}\epsilon^{-2}\right)$ and $ O\left(\eta^{-7}\epsilon^{-4}\right)$, respectively