Constrained Bi-Level Optimization: Proximal Lagrangian Value Function Approach and Hessian-free Algorithm

Authors: Wei Yao, Chengming Yu, Shangzhi Zeng, Jin Zhang

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate the efficiency of LV-HBA through numerical experiments on synthetic problems, hyperparameter optimization for SVM and federated bilevel learning. Empirical results validate the superior practical performance of LV-HBA.
Researcher Affiliation Collaboration Wei Yao124, Chengming YU1, Shangzhi Zeng31, Jin Zhang21* 1National Center for Applied Mathematics Shenzhen, SUSTech, 2Department of Mathematics, SUSTech, 3Department of Mathematics and Statistics, UVic, 4CETC Key Laboratory of Smart City Modeling Simulation and Intelligent Technology, The Smart City Research Institute of CETC
Pseudocode Yes Algorithm 1 proximal Lagrangian Value function-based Hessian-free Bi-level Algorithm (LV-HBA)
Open Source Code Yes The code is available at https://github.com/SUSTech-Optimization/LV-HBA.
Open Datasets Yes We conduct experiments on the dataset diabetes from Dua et al. (2017) and the dataset fourclass from Ho & Kleinberg (1996). Utilizing dataset gisette (Guyon et al., 2004).
Dataset Splits Yes For dataset diabetes, we randomly partition it into training, validation, and testing subsets containing 500, 150, and 118 examples, respectively. Similarly, dataset fourclass is partitioned into training, validation, and testing subsets with 500, 150, and 212 examples, respectively. For dataset gisette, we segment it into training, validation, and testing subsets, comprising 400, 180, and 5420 examples, respectively.
Hardware Specification Yes All experiments were conducted using Python 3.8 on a computer with an Intel(R) Xeon(R) Gold 5218R CPU @ 2.10GHz CPU and an NVIDIA A100 GPU with 40GB memory GPU.
Software Dependencies Yes All experiments were conducted using Python 3.8... The hyperparameter optimization of SVM and the data hyper-cleaning experiments were performed using qpth version 0.0.11 and cvxpy version 1.2.0. The experiments were executed with opencv-python version 4.6.0.66.
Experiment Setup Yes Detailed experimental settings and parameter configurations can be found in Appendix A.1. In Figure 1, the step sizes are chosen as α = 0.005, β = 0.002, η = 0.03, γ1 = γ2 = 10, r = 1 with parameter ck = (k + 1)0.3. For LV-HBA, the step sizes are chosen as α = 0.02, β = 0.001, η = 0.1, γ1 = γ2 = 1, r = 1000 with parameter ck = (k + 1)0.3.