Multivariate stochastic approximation using a simultaneous perturbation gradient approximation

By James C. Spall. 1992

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Optimization & Training

Citations in the corpus

Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computation (2013, methodology): “Unlike the SPSA (Spall, 1992) estimator, our estimator is unbiased even though the perturbations are not small (0 or 1), and it multiplies by the perturbation rather than dividing by it.”
Evolution Strategies as a Scalable Alternative to Reinforcement Learning (2017, background): “…special case where p ψ is factored Gaussian (as in this work), the resulting gradient estimator is also known as simultaneous perturbation stochastic approximation [Spall, 1992], parameter-exploring policy gradients [Sehnke et al., 2010], or zero-order gradient estimation [Nesterov and Spokoiny,…”
Estimating or Propagating Gradients Through Stochastic Neurons (2013, methodology): “Gradient estimators based on stochastic perturbations have been shown for long (Spall, 1992) to be much more efﬁcient than standard ﬁnite-difference approximations.”
Adversarial Risk and the Dangers of Evaluating Against Weak Attacks (2018, methodology): “We found SPSA to reliably produce adversarial examples, but we believe the speciﬁc choice of gradient-free optimizer is relatively unimportant.”
Evolved Policy Gradients (2018, neutral): “Similar formulations also appear in prior works including (36; 32; 16).”
Model-reduced gradient-based history matching (2011, methodology): “Parameters αk and ck have to be chosen according to the guideline presented by Spall in (Spall (1992), Spall (2000)).”
A review on particle swarm optimization algorithm and its variants to clustering high-dimensional data (2013, methodology): “This used the simultaneous perturbation stochastic approximation (SPSA) (Spall 1992) to drive the multi-dimensional extension of PSO, the MD-PSO (Kiranyaz et al. 2009).”
Online Model Learning Algorithms for Actor-Critic Control (2015, methodology): “The algorithm generates an estimate of the policy gradient using simultaneous perturbation stochastic approximation (SPSA) (Spall, 1992), which has been found to be efficient even in high-dimensional parameter spaces.”
Simulation optimization (2001, background): “Spall (1992) shows in fact that the asymptotic convergence rate using this gradient estimate in an SA algorithm is the same as the naïve method above.”
Risk-Constrained Reinforcement Learning with Percentile Risk Criteria (2015, methodology): “SPSA-based gradient estimate was first proposed in Spall (1992) and has been widely used in various settings, espe- cially those involving a high-dimensional parameter.”
3D-RCNN: Instance-Level 3D Object Reconstruction via Render-and-Compare (2018, methodology): “There exists other schemes like OpenDR [32], SPSA [46] but we found simple central derivatives to be effective and fast, since we avoid all CPUGPU memory transfer by making use of CUDA-OpenGL interop functionality available in all recent GPUs.”
A Stochastic Simplex Approximate Gradient (StoSAG) for optimization under uncertainty (2017, methodology): “Do and Reynolds [15] provided an analysis of the deterministic EnOpt method and demonstrated its close connection to other stochastic gradient estimation techniques such as simultaneous perturbation stochastic approximation (SPSA, [16, 17]) and the simplex gradient method [18].”
Variational quantum algorithms (2020, methodology): “Moreover, it has been shown that for a restricted set of problems, SPSA has a faster theoretical convergence rate (in terms of the number of function calls) than SGD performed with finite differences [81].”
Simulation optimization: a review, new developments, and applications (2005, background): “This can be especially true in the simulation setting, where the user has more control, so for instance can use common random numbers to induce positive correlation between estimates of solution performance to dramatically reduce the number of simulation replications required to make a distinction.”
Multiresolution registration of remote sensing imagery by optimization of mutual information using a stochastic gradient (2003, background): “by Spall [12], which would ensure convergence.”