INTRODUCTION:

Scythe MCMC is a C++ header library that eases the development of Markov Chain 

Monte Carlo (MCMC).  It is based on the Scythe Statistical Library. Scythe MCMC 

provides an execution framework, including command line and option file 

parsing, that reduces the amount of boilerplate code needed to write custom 

MCMC routines.  It also provides common MCMC step types, including Gibbs, 

Metropolis-Hastings, and Slice sampling.  Users can experiment with which 

sampling steps provide the best results and implement their own sampling steps 

as desired.

Scythe MCMC was motivated with the need to combine simpler samplers, such as 

univariate slice sampling, with more specialized samplers for particular 

problems.  Future versions of Scythe MCMC will include common samplers for 

nonparametric Bayesian models, including a Collapsed Gibbs sampler for the 

Dirichlet Process, and Slice samplers for the Dirichlet Process and Indian 

Buffet Process.

There are many choices for writing MCMC samplers.  Scythe MCMC is similar to, 

and heavily inspired by, MCMC++, but is based on the Scythe, which provides 

convenient matrix types, random number generators,  and probability 

distributions. Scythe MCMC is different in focus from  MCMCPack, which is also 

based on Scythe and interfaces with R.  Users wishing to distribute a specific 

model to R users should consider contributing to MCMCPack directly. Unlike 

MCMCPack, Scythe MCMC makes no attempt to interface with R and assumes samplers 

are executed from the command line (although this is not required). It also 

provides common algorithms (Gibbs, Metropolis-Hastings, Slice Sampling) that 

ease development of samplers for many models.

Even though implementing models in ScytheMCMC requires more work than 

implementing comparable models in Bugs  or JAGS, Scythe MCMC generally leads to 

faster samplers that can exploit the particular structure of a given problem.  

It also allows development of models that cannot be sampled effectively using 

either Bugs or Jags, such nonparametric Bayesian models without truncation 

approximations.

FEATURES: