Infinite-Dimensional Optimization for Zero-Sum Games via Variational Transport

Authors: Lewis Liu, Yufeng Zhang, Zhuoran Yang, Reza Babanezhad, Zhaoran Wang

ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section we report some results for a toy experiment with a Gaussian mixture model with 8 Gaussian distributions. For simplicity, we drop the regularizer terms from WGAN loss and consider a mixture of 8 generators and discriminators corresponding to the particles for parameters of the generator and the discriminator of WGAN. Both generators and discriminators are MLP with 3 layers. We also don t tune the learning rate and set it to be 10 4. We run the model for 20000 iterations which is small compared to the typical number of iterations used in practice to train a WGAN model. In our experiment we reused the code provided by (Hsieh et al., 2018) with some simple modification. We present some samples generated from trained generators in Figure 2. The blue dots are generated from real mixture models and the red ones are generated from generators. We observe that the distribution generated by our generator matches the groundtruth after a short training period, and the sampling procedure is faster than the SGLD-based methd.
Researcher Affiliation Collaboration 1Universit e de Montr eal, Canada 2Northwestern University, United States 3Princeton University, United States 4Samsung SAIT AI Lab, Canada.
Pseudocode Yes Algorithm 1 Multi-Step Variational Transport Algorithm for Infinite-Dimensional Games (VTIG)
Open Source Code No The paper does not provide an explicit statement or link indicating that the source code for the described methodology is open-source or publicly available.
Open Datasets No In this section we report some results for a toy experiment with a Gaussian mixture model with 8 Gaussian distributions.
Dataset Splits No The paper does not explicitly provide details about training, validation, or test dataset splits.
Hardware Specification No The paper does not provide specific details about the hardware used to run the experiments.
Software Dependencies No The paper does not list specific software dependencies with version numbers (e.g., Python 3.x, PyTorch 1.x) that would be needed to replicate the experiment.
Experiment Setup Yes Both generators and discriminators are MLP with 3 layers. We also don t tune the learning rate and set it to be 10 4. We run the model for 20000 iterations