Caltech SURF 2018: Evolutionary Pythia Tune with Distributed Computing

This is an improvement using genetic algorithms and fitting experimental data (instead of simulated data) compared to the original Bayesian optimization framework for tuning a Monte Carlo event generator, as described in the article below:

Event generator tuning using Bayesian optimization
Philip Ilten, Mike Williams, Yunjie Yang
arXiv:1610.08328

Set up and run the TuneMC framework

STEP 1: Install Spearmint and PYTHIA

1. Follow the instructions on PYTHIA (a popular Monte Carlo event generator widely used in High Energy Physics event simulation, e.g. LHC physics) for installation. (Note: no special package flags are needed during Pythia installation.)
2. Git clone the repository code.
3. In ./pythia_space, complile pythia_gen.cc with a PYTHIA8_HOME argument set to the top-level directory of your installation, i.e. compile it by typing $ make PYTHIA8_HOME=<path/to/pythia/top-level>

The framework should now be ready to use. If you like, you are welcome to familiarize yourself with basic usage of Spearmint and/or PYTHIA by looking at the examples they provide.

STEP 2: Set up the cluster

Place the files in the directory headnode onto the head node of the cluster. (The only dependency required on the head node is numpy.) Build an image from the Dockerfile, then edit the template.jdl file appropriately to match your system configuration and Docker image path.

STEP 3: Run the tune

Run python headrunner.py on the head node for the optimization to proceed. It will output fitness histories and parameter histories as .txt files in the headnode directory.