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 (Geyer, Charles J, 2011);
  • Metropolis-coupled Markov chain Monte Carlo (also known as parallel tempering) (Geyer, Charles J, 1991, Altekar, Gautam and Dwarkadas, Sandhya and Huelsenbeck, John P and Ronquist, Fredrik, 2004);
  • Hamilton Monte Carlo proposal (Neal, Radford M, 2011);
  • No U-Turn Sampler (NUTS) (Matthew D. Hoffman and Andrew Gelman, 2014).

Documentation

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

Examples

The Git repository also includes example MCMC analyses. Build them with cabal-install or Stack.

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 a phylogenetic dating project.

References

Altekar, Gautam and Dwarkadas, Sandhya and Huelsenbeck, John P and Ronquist, Fredrik (2004). Parallel metropolis coupled Markov chain Monte Carlo for Bayesian phylogenetic inference.

Geyer, Charles J (2011). {Introduction to Markov Chain Monte Carlo}, CRC press.

Geyer, Charles J (1991). Markov chain Monte Carlo maximum likelihood.

Matthew D. Hoffman and Andrew Gelman (2014). The No-U-Turn Sampler: Adaptively Setting Path Lengths in Hamiltonian Monte Carlo.

Neal, Radford M (2011). {MCMC Using Hamiltonian Dynamics}, CRC press.

Changes

Markov chain Monte Carlo sampling - ChangeLog

Unreleased changes

0.8.3.1

  • Bug fix MC3: Do not change prir and likelihood values for auto-tuned reciprocals temperatures.

0.8.3.0

  • Fix auto tuning with MC3 algorithm.

0.8.1.0

  • Automatic intermediate tuning for HMC and NUTS.
  • Fix documentation for generic proposals.
  • Tooling updates.
  • Slight runtime improvements (strictness annotations).

0.8.0.1

  • Improve exception handling (also during execution of monitors; also improve the output).
  • Increase maximum tuning parameter.
  • Minor bug fix: acceptance rates are again reported at end of run.

0.8.0.0

  • Use the new bytestring realfloat builder.
  • Fix a small memory leak during burn in with auto tuning.
  • Save chain after burn in.
  • New example (Poisson; see the respective blog post).
  • Simplify prior, likelihood, posterior, and Jacobian types.
  • Do not export grabble.
  • Remove chain index.
  • Remove lengthT.
  • Remove ProposalOrder default instance.
  • Use GHC 9.2.4 by default.
  • Reduce package dependencies as much as possible.

0.7.0.1

  • Use random-1.2. This means, that the random number generator type has changed from GenIO (System.Random.MWC) to IOGenM StdGen (System.Random.Stateful).
  • Always try to save chain, not only on UserInterupt.
  • Hamiltonian proposals:
    • Split HSettings into HParams and HStructure (preparation for NUTS).
    • Rename HTune to HTuningConf.
    • No-U-Turn sampler (NUTS).
  • Significant changes to Proposal and Tuner.
    • PResult.
    • ProposalSimple -> PFunction; Return type is (PResult, Maybe AcceptanceCounts).

0.6.2.5

  • Provide unsafe loader for MHG algorithm (useful for initializing chains with different prior or likelihood functions from saves).
  • Documentation and readme.
  • Log normal prior distribution.
  • Parallel computation of prior and likelihood. Speculative parallelization; this change is not always beneficial, we will see.

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.