Generalization Bound and Learning Methods for Data-Driven Projections in Linear Programming

Authors: Shinsaku Sakaue, Taihei Oki

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments demonstrate that learning projection matrices from data is indeed beneficial: it leads to significantly higher solution quality than the existing random projection while greatly reducing the time for solving LPs.
Researcher Affiliation Academia Shinsaku Sakaue The University of Tokyo and RIKEN AIP Tokyo, Japan sakaue@mist.i.u-tokyo.ac.jp Taihei Oki Hokkaido University Hokkaido, Japan oki@icredd.hokudai.ac.jp
Pseudocode No The paper describes algorithms and methods but does not provide structured pseudocode or algorithm blocks.
Open Source Code Yes The source code is available at https://github.com/ssakaue/data-driven-projection-lp-code.
Open Datasets Yes We used five LPs in Netlib [15]... The dataset used in this study is freely available at https://www.netlib.org/lp/data/.
Dataset Splits No each of which consists of 300 LP instances (200 for training and 100 for testing).
Hardware Specification Yes We used Mac Book Air with Apple M2 chip, 24 GB of memory, and mac OS Sonoma 14.1.
Software Dependencies Yes We implemented algorithms in Python 3.9.7 with Num Py 1.23.2. We used Gurobi 10.0.1 [26] for solving LPs and computing projection in (6).
Experiment Setup Yes We initialized P with that obtained by PCA and conducted a single epoch of training, setting the learning rate to 0.01. For Col Rand, PCA, and SGA, we used increasing values of the reduced dimensionality, k = n/100, . . . , until it reached the maximum value no more than n/10, i.e., up to 10% of the original dimensionality. PCA and SGA learned projection matrices P from N = 200 training instances