Code repository for Approximating Nash Equilibria for Black-Box Games: A Bayesian Optimization Approach

conda install nb_conda
conda env create --file environment.yml

This will create an environment called ne.

To activate this environment, execute:

source activate ne

The demo.ipynb demonstrates the algorithm along with the algorithms considered in the paper. It also demonstrates other utils/plots that can be found in the repo.

jupyter notebook

The script toy_experiments.py performs experiments on SADDLE and MOP problems. The experiment can be configured according to a configuration files, as follows. cd to the repo directory and do the following:

export PYTHONPATH=.
python ne/experiments/toy_experiments.py -f ne/experiments/configs/saddle_config.yml

As the experiment is running, results of different runs/algorithms/problems will be stored in ne/experiments/res as {experiment_name}_{alg_name}_{dimension}_{run_number}.json, these files are helpful for backup/monitoring purposes. At the end of the experiment a json file {experiment_name}.json will be generated which essentially concatentates all {experiment_name}_*.json

The demo.ipynb demonstrates how the results can be plotted. Moreover, a json file whose format is similar to that created by toy_experiments.py can be passed to the plot_regret_trace function under ne/utils/plots.py as demonstrated in the main block of ne/utils/plots.py. The results of the saddle_config.yml experiment is stored in ne/experiments/res/saddle_res.json

For the experiment defined above, the plots.py script is set to display its result:

python ne/utils/plots.py

To install from Python

from rpy2.robjects.packages import importr
utils = importr('utils')
utils.install_packages('GPGame')

There might be some difficulty in installing the "GPGame" package and interfacing it with Python. Make sure you install the package (and all the required packages) with sudo. Then copy it to the environment's R/library

sudo R
>> install.packages("GPGame")
>> quit()
sudo cp -a ~/R/x86_64-pc-linux-gnu-library/3.2/. ~/anaconda2/envs/ne/lib/R/library/

There is unintended multithreading with numpy ( https://stackoverflow.com/questions/19257070/unintented-multithreading-in-python-scikit-learn )

Check the blas/lapack library used and set the number of threads. E.g.

export OPENBLAS_NUM_THREADS=1

If you make use of this code and you'd like to cite us, please consider the following:

@article{al2018approximating,
  title={Approximating Nash Equilibria for Black-Box Games: A Bayesian Optimization Approach},
  author={Al-Dujaili, Abdullah and Hemberg, Erik and O'Reilly, Una-May},
  journal={arXiv preprint arXiv:1804.10586},
  year={2018}
}