Johann Brehmer, Sebastian Macaluso, Duccio Pappadopulo, and Kyle Cranmer

Particle physics experiments often require the reconstruction of decay patterns through a hierarchical clustering of the observed final-state particles. We show that this task can be phrased as a Markov Decision Process and adapt reinforcement learning algorithms to solve it. In particular, we show that Monte-Carlo Tree Search guided by a neural policy can construct high-quality hierarchical clusterings and outperform established greedy and beam search baselines.

Please see our paper for more details.

• A conda environment with all package dependencies can be installed with conda env create -f environment.yml
• Ideally, also get a MongoDB to run and use OmniBoard to monitor experiments (though this is optional)
• For the Ginkgo simulator, install ToyJetsShower (pip install -e . works, it's missing the pyro-ppl dependency though)
• For beam search and MLE estimates through the trellis, clone ReclusterTreeAlgorithms and hierarchical-trellis, and adapt the paths hard-coded in evaluator.py

To run individual experiments:

cd experiments

./experiment.py with truth         # Ground truth
./experiment.py with mle           # Trellis MLE
./experiment.py with greedy_s      # Greedy baseline
./experiment.py with beamsearch_s  # Beam search baseline
./experiment.py with mcts_s        # MCTS
./experiment.py with lfd_s         # Pure learning from demonstration
./experiment.py with lfd_mcts_s    # BC policy in MCTS algorithm

For a full list of configurations, see the experiments/config.py.

To automate the whole process on a SLURM HPC system:

cd experiments/hpc
sbatch --array 0-59 run.sh

You can monitor the training live with Omniboard, and plot the final results with experiments/plot_results.ipynb.