An Experimental Study of Noisy Optimization with RPS Over Different Distributions

Noisy optimization arises in problems where objective function evaluations are distorted by random noise from sources like measurement errors, stochastic processes, or simulation inaccuracies, making it difficult to accurately locate optima. Given its prevalence in real-world scenarios, effective optimization methods are essential. This study explores the Robust Parameter Searcher (RPS), a recently proposed extension of the Nelder-Mead Simplex algorithm that incorporates non-linearly increasing reevaluation limits and statistical tests for robust solution comparison. In this work, different RPS configurations are evaluated on noisy unimodal functions with Gaussian, Uniform, and Exponential noise distributions, comparing their performance against the canonical Nelder-Mead Simplex. Using graphical analysis and non-parametric statistical tests within a fixed computational budget in a ten- and twenty-dimensional space, the results demonstrate that RPS effectively improves optimization in noisy environments, making it a valuable approach for real-valued problems with box constraints.