marsplus / bMSGN

bMSGN

This is the code to replicate the results reported in the paper "Learning Binary Multi-Scale Games on Networks". The codebase now supports the replication of Figure 1 and 2 (i.e., results on synthetic data); we will add the code to replicate other figures soon.

1. Preparation

• Unzip the files data/BTER.zip and result/BTER.zip. They are the BTER network structures used in the experiments.
• Creat three folders: result/LikRatio, result/synthetic and result/figures, which will be used to store experimental results.
• Install the needed packages with conda create -n bMSGN python=3.7 --file requirements.txt -c mosek -c conda-forge -c pytorch.
• We used the Mosek solver to solve the maximum likelihood estimation (MLE) problems in the paper. A license is needed to use the solver; free license is available for students on their website. The solver is called via the CVXPY API as in here.

2. Generate Figure 1

• Under the src/ folder, run the following to generate the synthetic data and then estimate the game parameters (WARNING: the estimation is memory intensive):

    ./exp.sh simulation
    ./exp.sh estimation
  

• Plot the results by

    python plot.py --figure=1
  

3. Generate Figure 2

• Under the src/ folder, run the following to generate the synthetic data and then estimate the game parameters (WARNING: the estimation is memory intensive):

    ./liktest.sh simulation
    ./liktest.sh liktest
  

• Plot the results by

    python plot.py --figure=2