Bouncy Particle Sampler

\[ \newcommand{\ud}{\text{d}} \newcommand{\as}{\text{a.s.}} \newcommand{\Bern}{\text{Bern}} \newcommand{\Norm}{\text{N}} \newcommand{\Poi}{\text{Poi}} \newcommand{\PP}{\text{PP}} \newcommand{\1}{\mathbb{1}} \newcommand{\E}{\mathbb{E}} \newcommand{\P}{\mathbb{P}} \DeclareMathOperator{\Var}{Var} \newcommand{\R}{\mathbb{R}} \]

Summary
Download
Installation
Usage
- Reproducing results
- Creating a new model

Summary

Implementation of the Bouncy Particle Sampler (BPS) algorithm (focussing on the priority queue-based local BPS algorithm).

Please include the following citation:

@article{bps,
    Author = {Bouchard-C\^{o}t\'{e}, A. and Vollmer, S.J. and Doucet, A.},
    Title = {The Bouncy particle sampler: a non-reversible rejection-free {M}arkov chain {M}onte {C}arlo method},
    Note = {Technical report arxiv:1510.02451},
    Year = {2015}
}

Download

Installation

There are several options available to install the package:

Integrate to a gradle script

Simply add the following lines (replacing 1.0.0 by the current version (see git tags)):

repositories {
 mavenCentral()
 jcenter()
 maven {
    url "http://www.stat.ubc.ca/~bouchard/maven/"
  }
}

dependencies {
  compile group: 'ca.ubc.stat', name: 'rejfree', version: '1.0.0'
}

Compile using the provided gradle script

Check out the source git clone git@github.com:alexandrebouchard/rejfree.git
Compile using gradle installApp
Add the jars in build/install/rejfree/lib/ into your classpath

Use in eclipse

Check out the source git clone git@github.com:alexandrebouchard/rejfree.git
Type gradle eclipse from the root of the repository
From eclipse:
Import in File menu
Import existing projects into workspace
Select the root
Deselect Copy projects into workspace to avoid having duplicates

Usage

We document here two usage scenarios: reproducing results from our paper, and creation of new models.

Reproducing results

Complete reproducibility information is stored in separate git repos (containing precise git commits, plotting scripts, etc), but for most purpose the following information should be sufficient (those not listed here are implemented in Julia, and documented elsewhere; replace @@{x,y,z} by the cross product of all argument assignments or use westrun:

Figure 4:
- main class: rejfree.models.normal.CompareStanRFOnNormalModel
- arguments (note: ignore the restrictVelocityNorm option as the organization of this command line argument has been changed since the time this experiment was ran):

-useLocal @@{true,false} \
-offDiag @@{0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9} \
-nPairs 1000 \
-timeMilli 60000 \
-collectRate 0 \
-refreshRate @@{0.01,0.1,1,10} \
-restrictVelocityNorm @@{true,false}

Figure 5:
- main class: rejfree.models.normal.CompareStanRFOnNormalModel
- arguments:

-useLocal true \
-nPairs @@{100,1000} \
-timeMilli 30000 \
-collectRate 0 \
-refreshRate @@{0.01,0.1,1,10} \
-refreshmentMethod @@{GLOBAL,LOCAL,RESTRICTED,PARTIAL}

Figure 6:
- main class: rejfree.models.normal.CompareStanRFOnRN
- arguments:

++plans/singleRunOnRNOptions/options-@@{1,2,3,4}.map \
-stanHome /home/bouchard/bin/cmdstan-2.7.0 \
-restrictVelocityNorm @@{true,false} \
-usePartialRefreshment false

Figure 7 and 8:
- main class: rejfree.models.normal.CompareStanRFOnNormalModel
- arguments:

-recordPartialSums true \
-stanHome /Users/bouchard/Documents/workspace-cpp/cmdstan-2.7.0 \
-nPairs @@{10,100,1000} \
-nRepeats 40 \
-variableMonitorInterval 100

Figure 9:
- main class: rejfree.models.expfam.MRFMain
- arguments:

-stanHome /home/bouchard/bin/cmdstan-2.7.0 \
-nStanIters @@{2^[4--12]} \
-randomForBothAlgorithms @@{1--10} \
-collectRate 0.0 \
-useLocalRefreshment true \
-restrictVelocityNorm false

Creating a new model

Note: we are in the process of making this more user-friendly and better documented.

Create the factor graph
- See rejfree.models.expfam.MRF for an example, in particular the inner class rejfree.models.expfam.MRF$ModelSpec
- The main step is to create an object where the factors (modelled by classes implementing rejfree.local.CollisionFactor) are accessible in fields with the annotation @DefineFactor. The structure of the factor graph is built using reflection by looking at which factors have shared accessibility of some objects.
To create and run the local BPS from the factor graph:
- See rejfree.models.expfam.MRFMain for an example
- The key snippet there is:

ModelSpec modelSpec = mrf.newModelSpecFromGenerativeProcess(new Random(generateRandom));
LocalRFRunner rfRunner = new LocalRFRunner(localRFRunnerOption);
rfRunner.init(modelSpec);
rfRunner.addMomentRayProcessor();
rfRunner.run();

$\;$