Markov Chain Monte Carlo Sampling Using Direct Search Optimization
Abstract
We introduce population-based Markov chain Monte Carlo sampling algorithms that use proposal densities obtained by a novel method. Direct search optimization techniques (downhill simplex method and differential evolution) operate in real-valued spaces using a population of state vectors and geometric operations to generate proposals. Similar geometric proposals are used here for MCMC sampling but are modified to meet the strict requirements for unbiased sampling of the target density. We compare the new methods with existing population-based sampling approaches, and obtain superior performance, especially in high-dimensional tasks.
Cite
Text
Strens et al. "Markov Chain Monte Carlo Sampling Using Direct Search Optimization." International Conference on Machine Learning, 2002.Markdown
[Strens et al. "Markov Chain Monte Carlo Sampling Using Direct Search Optimization." International Conference on Machine Learning, 2002.](https://mlanthology.org/icml/2002/strens2002icml-markov/)BibTeX
@inproceedings{strens2002icml-markov,
title = {{Markov Chain Monte Carlo Sampling Using Direct Search Optimization}},
author = {Strens, Malcolm J. A. and Bernhardt, Mark and Everett, Nicholas},
booktitle = {International Conference on Machine Learning},
year = {2002},
pages = {602-609},
url = {https://mlanthology.org/icml/2002/strens2002icml-markov/}
}