This is the implementation of "Coordinating Followers to Reach Better Equilibria: End-to-End Gradient Descent for Stackelberg Games" by Kai Wang, Lily Xu, Andrew Perrault, Michael K. Reiter, and Milind Tambe published at AAAI 2022. In this repository, you can find three different subfolders, qre, ssg, and cyber, which refer to three examples presented in the paper.

All methods are implemented in Python3.6, where a list of the dependency can be found in package-list.txt.

model.py contains the implementation of an iterative Nash equilibrium oracle and a pytorch implementation of the stochastic differentiable Nash equilibria layer. Within each folder, main.py is the major implementation of the corresponding domain. You can simply run python main.py --method=$METHOD with various args to run the experiments. For reproducibility, you can specify the random seed for running experiments.

Our method is called diffmulti. Additionally, there are three different baselines: SLSQP, trust, and knitro, where the first two use scipy.optimize.minimize to run blackbox optimization, and the last one is the implementation of variational inequality reformulation using knitro solver. You can obtain a trial license from Knitro website to run the experiment.