Two-Timescale Gradient Descent Ascent Algorithms for Nonconvex Minimax Optimization

Authors: Tianyi Lin, Chi Jin, Michael I. Jordan

JMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We discuss our evaluation of TTGDA and TTSGDA in two domains robust regression and Wasserstein generative adversarial networks (WGANs). All of algorithms were implemented on a Mac Book Pro with an Intel Core i9 2.4GHz and 16GB memory. Figure 1: Performance of all the algorithms with 6 LIBSVM datasets. The numerical results are presented in terms of epoch count where the evaluation metric is the gradient norm of the function Φ( ) = maxy Y f( , y).
Researcher Affiliation Academia Tianyi Lin EMAIL Department of Industrial Engineering and Operations Research Columbia University New York, NY 10027, USA Chi Jin EMAIL Department of Electrical and Computer Engineering Princeton University Princeton, NJ 08544, USA Michael I. Jordan EMAIL Department of Electrical Engineering and Computer Science and Department of Statistics University of California, Berkeley Berkeley, CA 94720-1776, USA
Pseudocode Yes We provide detailed pseudocode for TTGDA and its stochastic counterpart (TTSGDA) in Algorithms 1 and 2. Algorithm 1 TTGDA. Algorithm 2 TTSGDA.
Open Source Code No The paper does not provide an explicit statement about open-sourcing code for the methodology described, nor does it provide a direct link to a code repository. It mentions implementation details but not code availability.
Open Datasets Yes We will consider the problem of robust logistic regression with nonconvex penalty functions. [...] Here, we compare TTGDA and TTSGDA with GDmax (Jin et al., 2020) on 6 LIBSVM datasets3 [...] 3. https://www.csie.ntu.edu.tw/~cjlin/libsvm/
Dataset Splits No The paper mentions using "6 LIBSVM datasets" and generating data for WGAN experiments (normal distribution with ˆµ = 0 and ˆσ = 0.1). However, it does not provide specific training/validation/test splits (e.g., percentages, sample counts, or predefined splits) for these datasets.
Hardware Specification Yes All of algorithms were implemented on a Mac Book Pro with an Intel Core i9 2.4GHz and 16GB memory.
Software Dependencies No The paper mentions comparing their algorithms against "Adam (Kingma and Ba, 2015)" and "RMSprop (Tieleman and Hinton, 2012)", which are optimization algorithms. However, it does not specify any software libraries or packages with version numbers used for implementing their own TTGDA/TTSGDA algorithms or experiments.
Experiment Setup Yes The algorithmic parameters are tuned as follows: we consider different pairs of (ηx, ηy) where ηy ∈ {10−1, 1} and the ratio ηy/ηx ∈ {10, 102, 103}, and different sizes M ∈ {10, 100, 200} for TTSGDA. [...] to set λ1 = 1/n2, λ2 = 10−2 and α = 10 for our experiments. [...] We fix the batch size M = 100 in TTSGDA and tune the parameters (ηx, ηy) as we have done in Section 6.1.