Selin Bayramoglu, Georgia Institute of Technology
Nick Sahinidis, Georgia Institute of Technology
Python package to build sparse regression models of strong branching and use them for branching. These models are the lasso, and L0 and L1/L2 penalized linear regression.
Please see COPYRIGHT.md .
Project based on the Computational Molecular Science Python Cookiecutter version 1.1.
This project uses Python3, R and C.
We suggest that users create a new conda environment to install this package.
conda create -n branching-env python=3.9
conda activate branching-env
This project uses the open-source solver SCIP 8.0.0. To install it, run
conda install -c conda-forge scip=8.0.0
For collecting training data via Python, we use a modified version of the Ecole library 0.8.1. To install it with the currently installed SCIP version, run
cd ecole
CMAKE_ARGS="-DSCIP_DIR=/Users/selinbayramoglu/opt/anaconda3/envs/branching-env/lib/cmake/scip -DCMAKE_INSTALL_RPATH_USE_LINK_PATH=ON -DCMAKE_BUILD_TYPE=Release" python -m pip install .
Change SCIP_DIR in the above command to the following path:
/Path/to/branching-env/lib/cmake/scip
We use the R packages glmnet and l0learn for building models. They can be installed as follows:
install.packages("glmnet", repos = "https://cran.us.r-project.org")
install.packages("L0Learn", repos = "http://cran.rstudio.com")
To install this package and its dependencies locally, do
pip install -r requirements.txt
Instances for training, validation and evaluation are in the instances-setcover folder. These are set covering problems generated randomly as in https://github.com/ds4dm/learn2branch-ecole .
Once a sparse branching rule, e.g. glmnet_lasso, is learned, and a model file is created, you can test the model's performance as follows:
In line 5 of branching/test-methods/CMakeLists.txt, set SCIP_DIR to /Path/to/branching-env/lib/cmake/scip
mkdir branching/test-methods/build
cd branching/test-methods/build
cmake .. -DCMAKE_BUILD_TYPE=Release
make
(Note: If you experience a linking problem during compilation, do:
export DYLD_LIBRARY_PATH=/Path/to/branching-env/lib)
This builds an executable called sparse. To run this rule for branching on an instance, e.g. instance_1.lp, with seed 0:
./sparse ../../../instances-setcover/evaluation/instance_1.lp glmnet_lasso 0
The branching rules available in SCIP, such as the default rule relpscost, can be tested similarly:
./sparse ../../../instances-setcover/evaluation/instance_1.lp relpscost 0
First, run example-part1-data.ipynb to collect data and create datasets. Once this runs correctly, do
bash example-part2-model-building.sh
to build sparse models for braching.
Finally, do
bash example-part3-testing.sh
to run the learned models and the default SCIP on the evaluation problems.