this is for holding javascript data
Xavier Holt edited Slice_Sampling_Slice_sampling_is__.md
over 8 years ago
Commit id: f3625426a91ea0e945a6b269510ff4359a1b471b
deletions | additions
diff --git a/Slice_Sampling_Slice_sampling_is__.md b/Slice_Sampling_Slice_sampling_is__.md
index a210158..2f3ce32 100644
--- a/Slice_Sampling_Slice_sampling_is__.md
+++ b/Slice_Sampling_Slice_sampling_is__.md
...
Slice sampling in particular possesses several very desirable properties \cite{murray2009elliptical}. It allows us to sample from integrals that are _proportional_ to our underlying distribution. There is no assumption that our probability mass sums to one. Consequently our intractable denominator above is of no concern -- we simply proceed using the scaled posterior.
Additionally, slice sampling has no free parameters. This is useful from a practical perspective, as it removes the need to spend time tuning and updating hyper-parameters for more complex models. It also impacts on performance; the more general Metropolis-Hastings model is highly sensitive to step-size parameters, which means unless we get it right we're going to perform very poorly on evaluation. In contrast, slice sampling adjusts the step-size to match the local shape of the density function.
There are several modified slice-samplers which are heavily parallelisable \cite{Tran_2015}.
SS as parallel MCMC \cite{Tran_2015} In our analysis we use the following slice-sampling algorithm, iteratively tuning on each subsequent weight-parameter.