Escaping limit cycles: Global convergence for constrained nonconvex-nonconcave minimax problems

Authors: Thomas Pethick, Puya Latafat, Panos Patrinos, Olivier Fercoq, Volkan Cevher

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The algorithms considered in the experiments include the adaptive Algorithm 1, (Curvature EG+), and constant stepsize methods that can be seen as instances of (CEG+) for various choices of γk and αk. ... We test the algorithms on the constructed examples and confirm their convergence guarantees. ... Results for the deterministic setting and stochastic setting can be found in Fig. 3 and Fig. 4 respectively.
Researcher Affiliation Academia Thomas Pethick Puya Latafat: Panagiotis Patrinos: Olivier Fercoq; Volkan Cevher Laboratory for Information and Inference Systems (LIONS), EPFL (thomas.pethick@epfl.ch) :Department of Electrical Engineering (ESAT-STADIUS), KU Leuven ;Laboratoire Traitement et Communication d Information, Télécom Paris, Institut Polytechnique de Paris
Pseudocode Yes Algorithm 1 (Adaptive EG+) Deterministic algorithm for problem (2.1) ... Algorithm 2 Lipschitz constant backtracking
Open Source Code Yes The supplementary code can be found at https://github.com/LIONS-EPFL/weak-minty-code/.
Open Datasets No The paper tests algorithms on
Dataset Splits No The paper does not specify training, validation, or test dataset splits. It evaluates algorithms on constructed examples and toy problems, not traditional datasets with predefined splits.
Hardware Specification No The paper does not provide specific details about the hardware used to run the experiments, such as GPU or CPU models.
Software Dependencies No The paper mentions using "Mathematica code" for computations in the appendices, but does not specify the version number of Mathematica or any other software dependencies crucial for reproducibility.
Experiment Setup Yes When γk 1{L and αk 1 we recover a constrained variant of extragradient, which we denote CEG. When αk 1{2 we denote the scheme CEG+, which is the direct generalization to the constraint setting of the (EG+) scheme studied in Diakonikolas et al. (2021, Thm. 3.2). ... For Example 5 we choose the problem parameters such that ρ 1{3L according to (B.13), and additionally add an ℓ8-ball constraint to keep the iterates bounded.