mcmc

Sample from a posterior using Markov chain Monte Carlo

https://github.com/dschrempf/mcmc#readme

Version on this page:0.6.2.4
LTS Haskell 22.39:0.8.3.1
Stackage Nightly 2024-10-31:0.8.3.1
Latest on Hackage:0.8.3.1

See all snapshots mcmc appears in

GPL-3.0-or-later licensed by Dominik Schrempf
Maintained by [email protected]
This version can be pinned in stack with:mcmc-0.6.2.4@sha256:507b2c1a9f73d6eea1ea79fd73291a6c26a4a6b4f606bf133b5bdf39ad7502c9,3127

Module documentation for 0.6.2.4

Markov chain Monte Carlo sampler

Sample from a posterior using Markov chain Monte Carlo (MCMC) algorithms.

At the moment, the following algorithms are available:

  • Metropolis-Hastings-Green 1;
  • Metropolis-coupled Markov chain Monte Carlo (also known as parallel tempering) 2 , 3.
  • Hamilton Monte Carlo proposal 4.

Documentation

The source code contains detailed documentation about general concepts as well as specific functions.

Examples

Example MCMC analyses can be built with cabal-install or Stack and are attached to this repository.

git clone https://github.com/dschrempf/mcmc.git
cd mcmc
stack build

For example, estimate the accuracy of an archer with

stack exec archery

For a more involved example, have a look at the phylogenetic dating project.

Footnotes

1 Geyer, C. J., Introduction to Markov chain Monte Carlo, In Handbook of Markov Chain Monte Carlo (pp. 45) (2011). CRC press.

2 Geyer, C. J., Markov chain monte carlo maximum likelihood, Computing Science and Statistics, Proceedings of the 23rd Symposium on the Interface, (1991).

3 Altekar, G., Dwarkadas, S., Huelsenbeck, J. P., & Ronquist, F., Parallel metropolis coupled markov chain monte carlo for bayesian phylogenetic inference, Bioinformatics, 20(3), 407–415 (2004).

4 Neal, R. M., Mcmc Using Hamiltonian Dynamics, In S. Brooks, A. Gelman, G. Jones, & X. Meng (Eds.), Handbook of Markov Chain Monte Carlo (2011). CRC press.

Changes

Markov chain Monte Carlo sampling - ChangeLog

Unreleased changes

0.6.2.4

  • Specify covariance version bounds. Use covariance-0.2.0.0 (specifically state that sigma is rescaled with rescaleSWith).

0.6.2.3

  • Allow burn in with fast proposals only (BurnInWithCustomAutoTUning). Sometimes it is advantageous to hold back slow proposals initially, especially when the state is so far off that it does not make sense to compute complex proposals.
  • Hamiltonian proposal: Use automatic differentiation specialized to Double (roughly 10 percent faster).

0.6.2.2

  • Remove dependency monad-parallel. Fix stackage build.

0.6.2.1

  • Improve logGammaG. The calculation of the gamma function involves vectors; the generalized version needs boxed vectors, and is slow. Using Typeable, we now check if the type is Double, and use the fast version in this case.

0.6.2.0

  • Improve leapfrog integrator.
  • Update tooling.
  • Cleanup proposals.

0.6.1.0

  • Revamp Hamiltonian proposal (storable vectors).
  • Use mass matrices; allow tuning of all masses (covariance estimation using specialized estimators).

0.6.0.0

  • Improve documentation.
  • Generalized priors allowing automatic differentiation.
  • Hamiltonian proposal.

mcmc-tree

0.5.0.0

  • Marginal likelihood estimation using thermodynamic integration or stepping stone sampling.
  • Various changes of function names (e.g., metropologis -> mhg).
  • Updated examples.
  • Proper but minimal logging framework.
  • Various other changes.

0.4.0.0

  • Greatly improve documentation.
  • Major design change: Introduction of the Algorithm type class; algorithms are data types. See MHG.
  • Metropolic-coupled Markov chain Monte Carlo algorithm (parallel chains).
  • Optimal acceptance rate depends on dimension of proposal.
  • Use a circular trace with constant memory usage (big change).
  • Therefore, batch monitors use vectors now.
  • Always save chain with complete trace (but with sensible length).
  • Determine necessary trace length at initialization.
  • Rename Item to Link.
  • Rename Status to Chain and separate Settings and Environment from the Chain.
  • Many bug fixes.

0.3.0

  • New shorter example/test for dating trees.
  • noData allows running a chain without likelihood function.
  • Give proposal parameters PName, PDescription, and PWeight newtype wrappers.
  • Give Tune a data type.
  • Allow periodical cleansing of state (Cleaner).
  • Add description string to proposals, so that they can be identified in an easier way.
  • Add simplices and proposals on simplices.
  • slideUniform renamed to slideUniformSymmetric.
  • Merge tools into internal.
  • Do not export internal modules.

0.2.4

  • Change order of arguments for proposals.
  • ’slideStem’ was renamed to ’slideBranch’.
  • Change ProposalSimple from newtype to type.
  • Contravariant instances of parameter and batch monitors. Use (>$<) instead of (@.) and (@#).
  • Add gammaDirichlet prior for partitioned dating analyses.

0.2.3

  • Contrary proposals.
  • Change how monitors are lifted (use normal function, not a lens).
  • Priors.
  • Remove concurrent monitors (was slow).
  • Improve MCMC sampler output.

0.2.2

  • Move away from hpack.

0.2.1

  • Consistently use ByteString instead of Text.
  • Verbosity levels.
  • Improved handling of proposals, moves, and monitors.
  • Bactrian moves.
  • Many small changes.

0.1.3

Many changes; notably it is now possible to continue a Markov chain run.