Gibbs race MEV repo

Prerequisite

• Install Julia 1.10
• Install Stan 2.35.0
• Set an environment variable CMDSTAN pointing to the directory where cmdstan is installed
• Install Java 11

Download the data using the following script:

cd data
./download.sh

Replication of results

To replicate an experiment, use for example

./nextflow run data_scale_normal.nf -resume

The scripts correspond to the following figures:

• data_scale_normal.nf and data_scale_HSP.nf are used for Figures 6, 7, 8 and 18,
• dim_scale_ex.nf and dim_scale_ex_HSP.nf are for the Figures where an increasing number of features is subsampled,
• dim_vs_size_scale.nf is for Figure 5.

Function usage

Some key functions in the Julia code:

• CGGibbs interface cggibbs(...).
• Stan NUTS interface nuts_stansample(...).

These two functions perform data preprocessing as well as sampling.

For example, from the root of this repo:

using Pkg 
Pkg.activate("julia-environment")
Pkg.instantiate()
include("julia-environment/src/sampling_functions.jl")

const PRIOR_SIGMA = 10.0
const MINIMUM_REQUIRED_ESS = 100

seed = 1
data_name = "colon"
n_samples = 1000 # number of initial samples
requested_dim = 100
requested_size = 10


# CGGibbs sampling
a_cggibbs = cggibbs(seed,data_name,n_samples,requested_dim,requested_size;
		model="logistic",prior="normal",synthetic=false,return_chain=false)

# Stan NUTS sampling
a_nuts = nuts_stansample(seed,data_name,n_samples,requested_dim,requested_size;
		model="logistic",prior="normal",synthetic=false,return_chain=false)