Last-Iterate Convergence of Optimistic Gradient Method for Monotone Variational Inequalities

Authors: Eduard Gorbunov, Adrien Taylor, Gauthier Gidel

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

Reproducibility Variable Result LLM Response
Research Type Experimental By solving (10) numerically, we empirically observe that e GPEG(γ, L, N) = O(1/N) for different choices of γ, see Fig. 1a. ... Codes for verifying the potentials and convergence rates are publicly available: https://github.com/eduardgorbunov/potentials_ and_last_iter_convergence_for_VIPs, the codes rely on the PEP packages [Taylor et al., 2017b, Goujaud et al., 2022] as well as on YALMIP [Lofberg, 2004].
Researcher Affiliation Academia Eduard Gorbunov MIPT, Russia Mila & Ude M, Canada MBZUAI, UAE eduard.gorbunov@mbzuai.ac.ae; Adrien Taylor INRIA & École Normale Supérieure, CNRS & PSL Research University, France adrien.taylor@inria.fr; Gauthier Gidel Mila & Ude M, Canada Canada CIFAR AI Chair gauthier.gidel@umontreal.ca
Pseudocode No The paper describes the methods using mathematical recursions and equations (e.g., Proj-EG, Proj-PEG, OG) but does not provide a formally labeled pseudocode or algorithm block.
Open Source Code Yes Codes for verifying the potentials and convergence rates are publicly available: https://github.com/eduardgorbunov/potentials_ and_last_iter_convergence_for_VIPs
Open Datasets No The paper focuses on theoretical analysis and numerical verification using performance estimation problems, not on traditional machine learning experiments that involve training on a specific dataset. Therefore, no public dataset is referenced for training.
Dataset Splits No The paper focuses on theoretical analysis and numerical verification using performance estimation problems, not on traditional machine learning experiments that involve training/validation splits.
Hardware Specification No The paper mentions using SDP solvers and YALMIP for numerical experiments but does not specify any hardware details like CPU/GPU models or memory.
Software Dependencies No The paper mentions using 'the PEP packages' and 'YALMIP' but does not specify their version numbers, which is necessary for reproducibility.
Experiment Setup Yes Under Assumption 1, for all N 0 and γ = 1/3L, we have F(x N) 2... Under Assumption 1, for all N 2 the iterates of Proj-PEG with γ = 1/4L satisfy...