EXPLORATION FLOORS AND ALMOST SURE CONVERGENCE OF ADAPTIVE GAUSSIAN MIXTURE AND PARZEN-TYPE SAMPLERS IN GLOBAL OPTIMIZATION
Global and Stochastic Analysis · 2026
DOI: 10.64837/gsa.13.5.1 ↗Kurz gefasst
Fachartikel in der Zeitschrift Global and Stochastic Analysis, 2026. Verfasst von SOONHUI LEE.
Abstract (Originalsprache)
The adaptive Gaussian mixture (AGM) algorithm samples candidate solutions from a Gaussian mixture centered at the M best solutions found so far. We develop its convergence theory in a form that covers modern practice. Isolating the property that drives the argument of Lee (2018), a possibly vanishing lower bound on the sampling density that we call an exploration floor, we prove almost sure convergence for every elite kernel sampler whose floor sequence is non-summable. Three consequences follow: an explicit threshold for the logarithmically decaying bandwidth schedule σj = c √ M/(log j) g : convergence for all c, M when g < 1/2, at g = 1/2 if and only if c 2M ≥ 2n, and provable trapping for g > 1/2; a shorter proof of the original theorems without the submartingale step or the positivity assumption; and, to our knowledge, the first almost sure convergence statement for unmodified prior-mixed samplers of the tree-structured Parzen estimator (TPE) family on continuous domains, which recovers the known ϵ-greedy guarantee as its constant-floor special case. Empirically, AGM dominates random search at every budget tested, matches TPE statistically on two real hyperparameter-tuning tasks at constant per-iteration cost, and a factorial sensitivity study together with a finite-budget simulation confirms the theoretical threshold exactly at g = 1/2.
Zitieren
APA
LEE, S. (2026). EXPLORATION FLOORS AND ALMOST SURE CONVERGENCE OF ADAPTIVE GAUSSIAN MIXTURE AND PARZEN-TYPE SAMPLERS IN GLOBAL OPTIMIZATION. Global and Stochastic Analysis. https://doi.org/10.64837/gsa.13.5.1
Vancouver
LEE S. EXPLORATION FLOORS AND ALMOST SURE CONVERGENCE OF ADAPTIVE GAUSSIAN MIXTURE AND PARZEN-TYPE SAMPLERS IN GLOBAL OPTIMIZATION. Global and Stochastic Analysis. 2026. doi:10.64837/gsa.13.5.1
BibTeX
@article{LEE2026,
author = {LEE, SOONHUI},
title = {EXPLORATION FLOORS AND ALMOST SURE CONVERGENCE OF ADAPTIVE GAUSSIAN MIXTURE AND PARZEN-TYPE SAMPLERS IN GLOBAL OPTIMIZATION},
journal = {Global and Stochastic Analysis},
year = {2026},
doi = {10.64837/gsa.13.5.1}
}
EUVIMED