Comparing effective sampling size vs acceptance rate tuned Random Walk sampling algorithms
Keywords
Loading...
Authors
Issue Date
2020-07-10
Language
en
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Random Walk Metropolis (RWM) is a stochastic approximation method to more efficiently
solve Bayesian Inference problems, whose exact solutions’ computations are intractable.
However, for fast convergence, a good proposal function in RWM needs to be chosen. This is
no easy task, since there is no optimal set of parameters for every type of problem to be
solved. To address this, several self tuning algorithms have been proposed. Traditionally,
acceptance rate tuners are commonly used, which have been shown to struggle with certain
types of probability distributions. We research the effectiveness of an adaptive tuning
algorithm using parameters that optimize the effective sampling size and compare the
algorithm with one that tunes towards an optimal average acceptance rate on those
problematic target probabilities. We find that the effective sampling size tuned RWM
sampler can outperform acceptance rate tuned RWM sampler for rough target densities.
Results for valley targets are ambiguous and further research is required.
Description
Citation
Supervisor
Faculty
Faculteit der Sociale Wetenschappen