Theoretical Study and Comparison of SPSA and RDSA Algorithms

Abstract

Stochastic approximation (SA) algorithms are widely used in system optimization problems when only noisy measurements of the system are available. This paper studies two types of SA algorithms in a multivariate Kiefer-Wolfowitz setting: random-direction SA (RDSA) and simultaneous-perturbation SA (SPSA), and then describes the bias term, convergence, and asymptotic normality of RDSA algorithms. The gradient estimations in RDSA and SPSA have different forms and, consequently, use different types of random perturbations. This paper looks at various valid distributions for perturbations in RDSA and SPSA and then compares the two algorithms using mean-square errors computed from asymptotic distribution. From both a theoretical and numerical point of view, we find that SPSA generally outperforms RDSA.