Solving stochastic weak Minty variational inequalities without increasing batch size
Authors: Thomas Pethick, Olivier Fercoq, Puya Latafat, Panagiotis Patrinos, Volkan Cevher
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 9 Experiments We compare BC-SEG+ and BC-PSEG+ against (EG+) using stochastic feedback (which we refer to as (SF-EG+)) and (SEG) in both an unconstrained setting and a constrained setting introduced in Pethick et al. (2022). See Appendix H.2 for the precise formulation of the projected variants which we denote (SF-PEG+) and (PSEG) respectively. In the unconstrained example we control all problem constant and set ρ = 1/10LF, while the constrained example is a specific minimax problem where ρ > 1/2LF holds within the constrained set for a Lipschitz constant LF restricted to the same constrained set. |
| Researcher Affiliation | Academia | Laboratory for Information and Inference Systems (LIONS), EPFL (thomas.pethick@epfl.ch) Laboratoire Traitement et Communication d Information, Télécom Paris, Institut Polytechnique de Paris Department of Electrical Engineering (ESAT-STADIUS), KU Leuven |
| Pseudocode | Yes | Algorithm 1 (BC-SEG+) Stochastic algorithm for problem (3.1) when A 0 |
| Open Source Code | No | The paper does not provide any statement or link indicating that open-source code for the described methodology is available. |
| Open Datasets | No | The paper uses synthetic examples (Example 2 and Example 3 from cited prior work) rather than a publicly available dataset with concrete access information. |
| Dataset Splits | No | The paper uses synthetic examples and does not specify training, validation, and test splits (e.g., percentages or sample counts). |
| Hardware Specification | No | The paper does not specify any hardware used for the experiments, such as particular CPU or GPU models. |
| Software Dependencies | No | The paper does not provide specific software names with version numbers required to reproduce the experiments. |
| Experiment Setup | Yes | The default configuration is γ = 1/2LF with αk = 1/18 (k/c+1), c = 100 and βk = αk for diminishing stepsize schemes and α = 1/18 for fixed stepsize schemes. We make two exceptions: Figure 1 uses the slower decay c = 1000 when γ = 0.1 and Figure 3 uses c = 5000 for γ = 0.01 (and otherwise c = 1000) to ensure fast enough convergence. When the aggressive stepsize schedule is used then αk = 1/18 k/100+1. Further details can be found in Appendix H. We choose σ = 0.1 and initialize with z0 = 1 if not specified otherwise. |