This repository includes new developments with Bayesian Additive Regression Trees and extends the original OpenBT repository created by Matt Pratola (https://bitbucket.org/mpratola/openbt/src/master/). Such extensions include Bayesian Model Mixing and Bayesian Calibration. All of the Bayesian Tree code is written in C++. User interfaces constructed in R and Python allow one to easily run the software. The BART Model Mixing software has been implemented in the Taweret Python package in conjunction with the BAND collaboration.

The heart of OpenBTMixing is a set of C++ tools that can be used directly via the command line or indirectly through the Python and R packages, which wrap them. Typically these tools are built with an implementation of the Message Passing Interface (MPI) such as Open MPI or MPICH to enable distributed parallelization of computations. In particular, the Python wrapper package is always built with MPI support.

The software and its distribution scheme have been developed to allow users to use OpenBTMixing with the MPI installation of their choice. For instance, it can be built with MPI installations on leadership class platforms and clusters that were installed by experts and optimized for their specific platform. As a result, however, the software is not distributed as prebuilt binaries or wheels, but rather must be built for each case with the compiler suite and matching MPI implementation provided by the user.

Before building and installing the bare C++ tools or the Python package, users must provide

• a compiler suite that includes a C++ compiler that supports the C++11 standard,
• an MPI installation that is compatible with the compiler suite,
• the Meson build system and its prerequistes such as Python 3 and ninja, and
• optionally the Eigen software package.

The Meson build system is setup to automatically detect the compiler suite and MPI installation to use. If Eigen already exists in the system and Meson can find it, then Meson will use it for the build. Otherwise, Meson will automatically obtain a copy of Eigen for internal use.

We presently test OpenBTMixing with both Open MPI and MPICH. In addition, we have successfully tested with the Intel MPI implementation and have used the Python package with MPI implementations installed

• via package managers such as Ubuntu's Advanced Packaging Tool (apt) and homebrew on macOS;
• by experts on clusters and that are available as modules; and
• with conda from the prebuilt conda forge openmpi package.

The Meson build system documentation suggests installing Meson via package manager when possible. Please refer to that documentation for detailed and up-to-date installation information.

If Meson cannot be installed by package manager or the manager's version is too old, the following is contrary to Meson suggestions but has been used successfully to install Meson with Python into a dedicated virtual environment as well as to install meson in the PATH for use without needing to activate that virtual environment.

$ /path/to/target/python -m venv ~/local/venv/meson
$ . ~/local/venv/meson/bin/activate
$ which python
$ python -m pip install --upgrade pip
$ python -m pip install meson
$ python -m pip list
$ ln -s ~/local/venv/meson/bin/meson ~/local/bin
<add ~/local/bin to PATH if desired and appropriate>
$ deactivate
$ which meson
$ meson --version

Note that this meson virtual environment is for installing just the Meson build system. Attempts to install openbtmixing into this virtual environment will likely fail with an error that the mesonbuild module cannot be found.

The OpenBTMixing Python package is distributed on PyPI as a source distribution that contains the C++ code and files needed by Meson to build the dedicated, standalone command line tools that the package will use. The tools are built and installed automatically by Meson as part of executing

python -m pip install openbtmixing

By default, pip install does not show any of Meson's progress. Users and developers interested in seeing how Meson satisifies dependencies and reviewing compiler output should pass -v to pip install.

Openbtmixing package installations can be minimally tested with

$ python
>>> import openbtmixing
>>> openbtmixing.__version__
'<version>'
>>> openbtmixing.test()

The package can also be built and installed from a clone of this repository with

$ cd /path/to/OpenBT/openbtmixing_pypkg
$ python -m build --sdist
$ python -m pip install -v dist/openbtmixing-<version>.tar.gz

where we assume that the build package has already been installed.

Developers can setup a virtual environment with a developer/editable mode installation of the package with

$ /path/to/target/python -m venv /path/to/venv/my_venv
$ . /path/to/venv/my_venv/bin/activate
$ cd /path/to/OpenBT/openbtmixing_pypkg
$ python -m pip install -v -e .

In this latter case, the command line tools are built automatically and installed at /path/to/OpenBT/openbtmixing_pypkg/src/openbtmixing/bin. The Python package is hardcoded to use those tools so that the existence of another set of tools in the system and in PATH should not cause issues.

Developers and C++ users can directly build and install the command line tools, an OpenBTMixing library, and library tests with tools/build_openbt_clt.sh. Note that these do not need to be built in order to use the Python package.

TODO: Needs to be written based on current state of affairs.

The examples from the article "Model Mixing Using Bayesian Additive Regression Tress" are reproduced in the jupyter noteboook BART_BMM_Technometrics.ipynb. This notebook can be run locally or in a virtual environment such as google colab.