Non-Bernoulli Perturbation Distributions for Small Samples in Simultaneous Perturbation Stochastic Approximation

TL;DR

This paper explores alternative perturbation distributions, specifically segmented uniform, for small-sample SPSA, showing that Bernoulli may not always be optimal in such settings.

Contribution

It introduces and analyzes segmented uniform perturbations for small-sample SPSA, challenging the traditional use of Bernoulli distributions.

Findings

Segmented uniform can outperform Bernoulli in small-sample SPSA.

Bernoulli distribution's optimality is not guaranteed in small-sample scenarios.

Theoretical analysis supports alternative perturbation choices for improved performance.

Abstract

Simultaneous perturbation stochastic approximation (SPSA) has proven to be efficient for recursive optimization. SPSA uses a centered difference approximation to the gradient based on two function evaluations regardless of the dimension of the problem. Typically, the Bernoulli +-1 distribution is used for perturbation vectors and theory has been established to prove the asymptotic optimality of this distribution. However, optimality of the Bernoulli distribution may not hold for small-sample stochastic approximation (SA) runs. In this paper, we investigate the performance of the segmented uniform as a perturbation distribution for small-sample SPSA. In particular, we conduct a theoretical analysis for one iteration of SA, which is a reasonable starting point and can be used as a basis for generalization to other small-sample SPSA settings with more than one iteration. In this work, we show that the Bernoulli distribution may not be the best choice for perturbation vectors under certain choices of parameters in small-sample SPSA