Adaptive and Optimal Second-order Optimistic Methods for Minimax Optimization

Authors: Ruichen Jiang, Ali Kavis, Qiujiang Jin, Sujay Sanghavi, Aryan Mokhtari

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 7 Numerical experiments
Researcher Affiliation Academia Ruichen Jiang ECE department, UT Austin rjiang@utexas.edu Ali Kavis ECE department, UT Austin kavis@austin.utexas.edu Qiujiang Jin ECE department, UT Austin qiujiangjin0@gmail.com Sujay Sanghavi ECE department, UT Austin sanghavi@mail.utexas.edu Aryan Mokhtari ECE department, UT Austin mokhtari@austin.utexas.edu
Pseudocode Yes Algorithm 1 Adaptive Second-order Optimistic Method
Open Source Code Yes We have uploaded our Matlab codes which generate all the empirical results in the numerical experiments. We have also included the instructions to reproduce all the experimental results Section 7 which could be found in Appendix D.
Open Datasets No Synthetic min-max problem: We first consider the min-max problem in [21, 30], given by minx Rn maxy Rn f(x, y) = (Ax b) y + (L2/6) x 3, which satisfies Assumptions 2.1 and 2.2. ... we generate the matrix A Rd d to ensure a condition number of 20. The vector b Rd is generated randomly according to N(0, I).
Dataset Splits No The paper does not explicitly specify training/validation/test dataset splits. It describes how data is generated for synthetic problems and problem formulations, but not how these are split for training, validation, and testing.
Hardware Specification No The paper mentions running experiments 'on our personal computer with normal CPU' but does not provide specific hardware details such as CPU model, GPU model, or memory specifications.
Software Dependencies No The paper mentions using 'MATLAB linear equation solver' and 'Matlab codes' but does not provide specific version numbers for MATLAB or any other software dependencies.
Experiment Setup Yes The hyper-parameters for methods in the prior work are tuned to achieve the best performance per method. Specifically, for the HIPNEX method in [30], it has a hyper-parameter σ (0, 0.5), which we choose in the interval [0.05, 0.1, 0.15, . . . , 0.45] for the best performance. Other hyper-parameters are determined by the formulas from [30]. For the Optimal SOM, the initial step size is set to be 1 as prescribed. Their algorithm has two line-search hyperparameters α and β. Note that their α is the same as ours, and we search for the best choice of α and β for their algorithm from the interval [0.1, 0.2, . . . , 0.9]. We use the combination that achieves the best empirical result. Finally, we initialize all the algorithms at the same point z0 = (x0, y0) Rd, drawn from the multivariate normal distribution.