Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
An Accelerated Gradient Method for Convex Smooth Simple Bilevel Optimization
Authors: Jincheng Cao, Ruichen Jiang, Erfan Yazdandoost Hamedani, Aryan Mokhtari
NeurIPS 2024 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we evaluate our AGM-Bi O method on two different bilevel problems using real and synthetic datasets. We compare its runtime and iteration count with other methods, including a-IRG [12], CG-Bi O [6], Bi-SG [13], SEA [14], R-APM [8], PB-APG [16], and Bisec-Bi O [15]. |
| Researcher Affiliation | Academia | Jincheng Cao ECE Department UT Austin EMAIL; Ruichen Jiang ECE Department UT Austin EMAIL; Erfan Yazdandoost Hamedani SIE Department The University of Arizona EMAIL; Aryan Mokhtari ECE Department UT Austin EMAIL |
| Pseudocode | Yes | Algorithm 1 Accelerated Gradient Method for Bilevel Optimization (AGM-Bi O); Algorithm 2 Proximal Accelerated Gradient Method for Bilevel Optimization (P-AGM-Bi O) |
| Open Source Code | Yes | The code and data are attached in the supplementary material. |
| Open Datasets | Yes | We apply the Wikipedia Math Essential dataset [34] which is composed of a data matrix A Rn d with n = 1068 samples and d = 730 features and an output vector b Rn. |
| Dataset Splits | Yes | We use 75% of the dataset as the training set (Atr, btr) and 25% as the validation set (Aval, bval). |
| Hardware Specification | Yes | All simulations are implemented using MATLAB R2022a on a PC running mac OS Sonoma with an Apple M1 Pro chip and 16GB Memory. |
| Software Dependencies | Yes | All simulations are implemented using MATLAB R2022a on a PC running mac OS Sonoma with an Apple M1 Pro chip and 16GB Memory. |
| Experiment Setup | Yes | For our AGM-Bi O method, we set the target tolerances for the absolute suboptimality and infeasibility to ϵf = 10^-4 and ϵg = 10^-4, respectively. We choose the stepsizes as ak = 10^-2(k + 1)/(4Lf). In each iteration, we need to do a projection onto an intersection of a L2-ball and a halfspace, which has a closed-form solution. For a-IRG, we set η0 = 10^-3 and γ0 = 10^-3. For CG-Bi O, we obtain an initial point with FW gap of the lower-level problem less than ϵg/2 = 5 × 10^-5 and choose stepsize γk = 10^-2/(k + 2). |