Competitive Gradient Descent

Authors: Florian Schaefer, Anima Anandkumar

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Using numerical experiments and rigorous analysis, we provide a detailed comparison to methods based on optimism and consensus and show that our method avoids making any unnecessary changes to the gradient dynamics while achieving exponential (local) convergence for (locally) convex-concave zero sum games.
Researcher Affiliation Academia Florian Schäfer Computing and Mathematical Sciences California Institute of Technology Pasadena, CA 91125 florian.schaefer@caltech.edu Anima Anandkumar Computing and Mathematical Sciences California Institute of Technology Pasadena, CA 91125 anima@caltech.edu
Pseudocode Yes Algorithm 1: Competitive Gradient Descent (CGD) for 0 k N 1 do xk+1 = xk η Id η2D2 xyf D2 yxg 1 xf ηD2 xyf yg ; yk+1 = yk η Id η2D2 yxg D2 xyf 1 yg ηD2 yxg xf ; return (x N, y N);
Open Source Code No The paper mentions third-party libraries like Autograd and Julia Diff frameworks that support automatic differentiation, but it does not provide a link or statement for the authors' own implementation code.
Open Datasets No The paper mentions using a "Gaussian mixture model with two modes, in two dimensions (see supplement for details)" and a "covariance estimation example used by Daskalakis et al. (2017)[Appendix C]". While a citation is provided for the latter, neither is explicitly stated as a publicly available dataset with concrete access information (e.g., URL, DOI, or a clear statement of public access to the specific dataset used).
Dataset Splits No The paper does not provide specific details regarding dataset splits for training, validation, or testing (e.g., percentages, sample counts, or references to predefined splits).
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory, or specific cloud instances) used for running the experiments.
Software Dependencies No The paper mentions "Autograd", "Forward Diff", "Reverse Diff", and "RMSProp (ρ = 0.9)" but does not provide specific version numbers for these software dependencies, which are necessary for full reproducibility.
Experiment Setup Yes We apply SGA, Con Opt (γ = 1.0), OGDA, and CGD for stepsize η {0.4, 0.1, 0.025, 0.005} together with RMSProp (ρ = 0.9).