Principal Component Hierarchy for Sparse Quadratic Programs

Authors: Robbie Vreugdenhil, Viet Anh Nguyen, Armin Eftekhari, Peyman Mohajerin Esfahani

ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We benchmark different approaches to solve problem (P) in the sparse linear regression setting. All experiments are run on a laptop with Intel(R) Core(TM) i7-8750 CPU and 16GB RAM using MATLAB 2020b. The optimization problems are modeled using YALMIP (L ofberg, 2004), as the interface for the mixed-integer solver (MOSEK Ap S, 2019). Codes are available at: https://github.com/ RVreugdenhil/sparse QP
Researcher Affiliation Collaboration 1Delft Center for Systems and Control, Delft University of Technology 2Department of Management Science and Engineering, Stanford University 3Vin AI Research, Vietnam 4Department of Mathematics and Mathematical Statistics, Umea University.
Pseudocode No The paper describes algorithms (e.g., best response, dual program) in text and mathematical equations, but does not present them in a formalized pseudocode or algorithm block.
Open Source Code Yes Codes are available at: https://github.com/ RVreugdenhil/sparse QP
Open Datasets Yes for the UCI Superconductivity dataset1, we randomly select 70% of the dataset as training samples, and then compute the matrix Q, as specified in (2). 1Available at https://archive.ics.uci.edu/ml/ datasets/Superconductivty+Data. We benchmark different methods using real data sets (Dua & Graff, 2017); the details are listed in Table 3.
Dataset Splits No For each independent replication, we randomly sample 70% of the data as training data and 30% as test data. No explicit mention of a separate validation split or set was found.
Hardware Specification Yes All experiments are run on a laptop with Intel(R) Core(TM) i7-8750 CPU and 16GB RAM using MATLAB 2020b.
Software Dependencies Yes All experiments are run on a laptop with Intel(R) Core(TM) i7-8750 CPU and 16GB RAM using MATLAB 2020b. The optimization problems are modeled using YALMIP (L ofberg, 2004), as the interface for the mixed-integer solver (MOSEK Ap S, 2019).
Experiment Setup Yes To avoid a complicated terminating criterion, we run the BR method for TBR = 20 iterations, and we run the DP method for TDP = 500 iterations. The stepsize constant for DP (see Footnote 2) is set to a = 4 10 3. Regarding the post-processing step in Section 3.3, we fix p BR = 6 and p DP = 50 as the number of terminating solutions that are used to estimate the support Z. For the experiment on the synthetic data, the big-M constant is set to 4, because xtrue = 1. We choose the dimension of subspace k = 400. We fix the regularisation term η = 10. The number of iterations for the BR and DP are chosen to be TBR = 40 and TDP = 5000, respectively. The number of terminating solutions that are used to estimate the support Z in the postprocessing step is fixed to p BR = 10 and p DP = 100 for the two methods. The stepsize constant (see Footnote 2) is set a = 2 10 3. Moreover, we fix the big-M constant using the procedure described in Section 3.3. For the real datasets, the number of samples N is sufficiently bigger than the dimension n, thus we set the ridge regularization parameter to η = Ntrain so that the effect of regularization diminishes as N increases. We choose the sparsity level s = 10, similar to Section 4.1, and set the time limit for MOSEK to 300 seconds.