Variational PT plug-in for Blang

This repository contains the Blang code accompanying the paper "Parallel Tempering With a Variational Reference" by Surjanovic, Syed, Bouchard-Côté, and Campbell. NeurIPS (2022). https://arxiv.org/abs/2206.00080

Required software

• bash
• the following should be available in the PATH variable:
  • Java Open SDK, version 11 (use sdkman)
  • Rscript
• the following R packages should be installed:
  • ggplot2
  • dplyr
  • ggridges

Compiling

cd bl-vpt
./gradlew installDist

Running the stabilized moment matching on one model

Still from the bl-vpt folder:

chmod 755 pt-matching.sh
./pt-matching.sh --model ptbm.models.ToyMix

To see available options, use ./pt-matching.sh --model ptbm.models.ToyMix --help

To find other existing available models, use ls src/main/java/ptbm/models/

Adding a new model

• You can base your new model on one of the models in bl-vpt/src/main/java/ptbm/models/
• Implement the model based off the Blang documentation with the following amendments:
  • For each latent random variable that you would like to approach variationally, enclose the distribution declaration with Opt, and pass in the target distribution as argument as in the following example: mu ~ Normal(0.0, 1.0) becomes mu ~ Opt(Normal::distribution(0.0, 1.0))
  • Declare the type of such random variable as VariationalReal instead of RealVar.
• Then follow the compilation instructions above.