boundml 0.1.11

BoundML is a toolbox that helps with the development and evaluation of machine learning approach for the branch and bound algorithm.

BoundML

BoundML is a wrapper around a fork of ecole. It allows to easily develop new machine learning based branching strategies based for the Branch and Bound.

Installation

First, it is recommended to install pyscipopt and the different dependencies with conda to have the same scip installation for pyscipopt and ecole-fork (on which is based boundml).

conda install pyscipopt fmt pybind11

export CMAKE_PREFIX_PATH="${CONDA_PREFIX}"      
export CPLUS_INCLUDE_PATH="${CONDA_PREFIX}/include/"
export LIBRARY_PATH=${CONDA_PREFIX}/lib
export LD_LIBRARY_PATH=${CONDA_PREFIX}/lib

pip install boundml

The exports commands allow the compiler to find SCIP.

To install range-v3 for a ubuntu ditribution:

sudo apt install librange-v3-dev librange-v3-doc

Example

The file gnn_pipeline shows how tu use this library to reproduce easily the work of Gasse et al.. It consists of training a GCNN to learn to imitate strong branching.