This repository contains the code and results presented in arXiv:1903.09644.
GroomRL is a reinforcement learning framework to train jet grooming strategies.
GroomRL is tested and supported on 64-bit systems running Linux.
Install GroomRL with Python's pip package manager:
git clone https://github.com/JetsGames/groomrl.git
cd groomrl
pip install .
To install the package in a specific location, use the "--target=PREFIX_PATH" flag.
This process will copy the groomrl program to your environment python path.
We recommend the installation of the GroomRL package using a miniconda3
environment with the
configuration specified here.
GroomRL requires the following packages:
- python3
- numpy
- fastjet (compiled with --enable-pyext)
- gym
- matplotlib
- pandas
- keras
- keras-rl
- tensorflow
- json
- gzip
- argparse
- hyperopt (optional)
The final models presented in arXiv:1903.09644 are stored in:
- results/groomerW_final: GroomRL model trained on W jets.
- results/groomerTop_final: GroomRL model trained on top jets.
All data used for the final models can be downloaded from the git-lfs repository at https://github.com/JetsGame/data.
In order to launch the code run:
groomrl <runcard.json> [--output <folder>]
This will create a folder containing the result of the fit.
To create diagnostic plots, run with the --plot option or use
groomrl --model <folder> --plot [--nev n]
on the previous output folder.
The groomer can also be exported with the --cpp option, or by running
groomrl --model <folder> --cpp
This will create a cpp/model.nnet file that can be imported in a c++ framework.
To apply an existing grooming model to a new data set, you can use
groomrl --model <folder> --data <datafile> [--nev n]
which will create a new directory in <folder> using the datafile name.
To call models trained with GroomRL in a C++ code, install and use the libGroomRL library.
- S. Carrazza and F. A. Dreyer, "Jet grooming through reinforcement learning," arXiv:1903.09644