pgoelz / fluid

Simulation code for the paper “The Fluid Mechanics of Liquid Democracy” by Paul Gölz, Anson Kahng, Simon Mackenzie, and Ariel D. Procaccia.

Read me

This repository contains the simulation code for the paper

Paul Gölz, Anson Kahng, Simon Mackenzie and Ariel D. Procaccia: The Fluid Mechanics of Liquid Democracy. 2018.

The paper is freely accessible at http://procaccia.info/papers/fluid.pdf.

Requirements

• Python 3.6 (higher versions might work, but so far Gurobi does not support them)
• Gurobi with gurobipy python bindings (we used version 8.0.1)
• Matplotlib (2.2.2)
• Numpy (1.14.5)
• Mock (2.0.0): only required for unit tests in test_approximate_confluent_flow.py
• Graphviz (2.40.1): the dot and unflatten binaries should be available in PATH to generate example graphs in graph_examples.py.

For academic use, Gurobi provides free licenses at http://www.gurobi.com/academia/for-universities.

Replication of experiments in the paper

The figures from the body of the paper can be reproduced by running the script scripts/generate_figures.sh. The produced plots can be found in data/plots/, named after the figure number, and most experiments also generate a log in data/logs/. For more information on what the individual commands do and the meaning of the parameters, call the individual python scripts with -h. Since script sets the environment variable $GUROBI_SINGLE_THREAD, our code prevents Gurobi from spawning multiple threads. Without this variable, execution times will likely be faster, but less reproducible.

We expect the resulting graphics to be close to the ones found in our publication. In particular, the randomness seed is fixed for all experiments. Nonetheless, details like the iteration order of dictionaries are not guaranteed to be the same across different systems and versions of Python, which can lead to different results. Finally, running times and time-outs heavily depend on the system. Our figures were generated on a MacBook Pro (2017) on MacOS 10.12.6 with a 3.1 GHz Intel Core i5 processor and 16 GB of RAM.

Questions

For questions on the simulations, feel free to contact Paul Gölz or Anson Kahng.