ptmc is an R package designed to perform Bayesian inference using a Parallel Tempering Markov Chain Monte Carlo (PT-MCMC) algorithm. This implementation allows efficient exploration of discrete parameter spaces, making it especially useful for models where standard MCMC techniques struggle with multimodal posteriors or slow convergence. This implementation is a generalised version of the algorithm outlined by Miasojedow et al. 2012.
- Implements a Metropolis-Hastings algorithm with parallel tempering for enhanced sampling efficiency.
- Supports arbitrary user-defined likelihood and prior functions , enabling application to a wide range of models.
- Designed to effectively sample from complex posterior distributions with multiple modes.
- Parallelized execution across chains for improved performance.
Standard MCMC techniques can become inefficient when dealing with posterior distributions that:
- Have multiple modes (e.g., distinct peaks in probability).
- Feature complex discrete spaces where local exploration struggles to reach distant areas of high probability.
Parallel tempering mitigates these challenges by running multiple chains at different "temperatures." High-temperature chains explore the parameter space more broadly, while low-temperature chains focus on precise posterior estimation. Occasional exchanges between chains help the algorithm escape local optima and improve convergence.
To install the ptmc package, follow these steps:
Make sure you have R installed on your system. You can download R from https://cran.r-project.org/.
You can install the development version of ptmc from GitHub using the devtools package. If you don't already have devtools installed, you can install it with:
install.packages("devtools")
devtools::install_github("dchodge/ptmc")
After installation, you can load the serojump package into your R session with:
library(ptmc)
See background/ptmc.pdf.
We offer several examples of how to implement this package.
- Case 1: A simple implementation with a continuous space of parameters
- Case 2: Negronis all round: simulation recovery of a discrete model 🍹
We welcome contributions and suggestions! If you'd like to contribute to the ptmc package or report issues, please feel free to:
- Submit a pull request on GitHub.
- Open an issue on the repository.
If you have any questions or feedback, or would like more informative vignettes, you can contact the package maintainer at:
David Hodgson Email: [email protected]
This package is actively maintained and in a stable, usable state. New features and improvements are continually being developed.