ParticleBayes is an R package that contains a collection of particle filters for a subset of hierarchical bayesian models, with a focus on sequentially learning model parameters alongside state variables. The R code calls an open-source Java library, ParticleLearningModels, to perform calculations, so when/if a practitioner wants to implement a model in production, the underlying, streaming-capable, API is available.
Most models are formulated as dynamic linear models, or mixtures thereof, although some support for hidden markov models exists. The models also implement
- Water-filling resampling
- Parameter learning, via Particle Learning where applicable
Currently, the implemented filters cover two response types:
- Multinomial/Categorical with a logistic link function
- 10-component normal mixture approximation
- TBD
- Multivariate Gaussian
and, naturally, their univariate equivalents.
ParticleBayes depends on the rJava R package, so make sure it's installed and
operational. This also means you'll need to have a functional JRE, so if you
don't you can get one from
here,
install, and then run R CMD javareconf (with sudo if you're using linux)
at the command line. If you're using a Mac, you might need to run something
like R CMD javareconf JAVA_CPPFLAGS=-I/System/Library/Frameworks/JavaVM.framework/Headers
instead. Pay attention for any errors in the output, and look around the internet
for more information.
While ParticleBayes is still in it's beta stages, installation in R requires the devtools package and something like the following:
library(devtools)
install_url("https://bitbucket.org/brandonwillard/particlebayes/get/HEAD.zip", subdir="pkg")
then, as usual, add library(ParticleBayes) to your R source to use.
NOTE: This will pick up the most recently pushed version of the code (i.e. the
jar file in pkg/inst/java). Inclusion of this file in the repo is temporary,
that is, until this package gets submitted to CRAN.
ParticleBayes uses Maven, so you'll need to download that.
Once you have it,
mvn clean package -Dmaven.test.skip=true
should create the necessary jars in pkg/inst/java, so that the code is callable from
R.
The java code is called, within R, through rJava
Similar to the R installation, and from the project's root directory, one possible development setup script could include
library(devtools)
dev_mode(on=T)
system("mvn clean package -Dmaven.test.skip=true")
install_local("./pkg")
options(java.parameters=
c("-Xmx2g",
"-Xdebug",
"-Xrunjdwp:server=y,transport=dt_socket,address=4001,suspend=n"))
library(ParticleBayes)
...
dev_mode(on=F)The script will build the Java code, install the package in an isolated environment (see devtools), and set up the JVM for debugging.