ICLR-2020

code for our ICLR 2020 paper: Convergence Behaviour of Some Gradient-Based Methods on Bilinear Zero-Sum Games

In the directory gmm:

• gmm_"algorithm".py: using the "algorithm" to learn Gaussian mixtures, Gauss-Seidel version
• gmm_"algorithm"J.py: using the "algorithm" to learn Gaussian mixtures, Jacobi version
• model.py: the neural net architecture

In the directory shift:

• shift_"algorithm".py: using the "algorithm" to learn the mean of a Gaussian, Gauss-Seidel version
• shift_"algorithm"J.py: using the "algorithm" to learn the mean of a Gaussian, Jacobi version
• model.py: the discriminator and the generator

Acronyms:

• gd/sgd: gradient descent
• m: momentum / heavy ball
• eg: extra-gradient
• ogd: optimistic gradient descent