A Comparative Study of Adaptive MCMC Schemes with Stopped and Diminishing Adaptation

Onderzoeksoutput: Conference paper


Markov chain Monte Carlo (MCMC) techniques play a vital role in sampling from complex, high-dimensional target distributions. Ho-wever, the optimization of the proposal distribution for efficient sampling poses a challenging task, which is where adaptation becomes significant. This paper presents a comparative analysis of three adaptive sampling strategies: Metropolis Gaussian Adaptation (MGaA), Metropolis Covariance Matrix Adaptation Evolution Strategy (MCMA), and Adaptive Metr-opolis (AM). It is noteworthy that incorrect implementation of adaptation can compromise the ergodicity of MCMC samplers, which is essential for generating unbiased samples and converging to the target distribution. To address this concern, two strategies, Stopped Adaptation (SA) and Diminishing Adaptation (DA), are introduced within the adaptive sampler framework to uphold ergodicity. Through a comprehensive evaluation across diverse test distributions, this research assesses the performance of MGaA, MCMA, and AM samplers in various scenarios. By comparing their strengths and capabilities, the study provides valuable insights into effective approaches for sampling from complex distributions.
Originele taal-2English
TitelPre-proceedings of the Joint International Scientific Conferences On AI And Machine Learning BNAIC/BeNeLearn 2023
UitgeverijTU Delft Open
Aantal pagina's15
StatusPublished - 8 nov 2023
EvenementBNAIC/BeNeLearn 2023
: Joint International Scientific Conferences on AI and Machine Learning
- TU Delft, Delft, Netherlands
Duur: 8 nov 202310 nov 2023
Congresnummer: 2023


ConferenceBNAIC/BeNeLearn 2023
Internet adres


Duik in de onderzoeksthema's van 'A Comparative Study of Adaptive MCMC Schemes with Stopped and Diminishing Adaptation'. Samen vormen ze een unieke vingerafdruk.

Citeer dit