DC3: A learning method for optimization with hard constraints

Authors: Priya L. Donti, David Rolnick, J Zico Kolter

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

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate DC3 for convex quadratic programs (QPs), a simple family of non-convex optimization problems, and the real-world task of AC optimal power flow (ACOPF). Each experiment is run 5 times for 10,000 examples x (with train/test/validation ratio 10:1:1).
Researcher Affiliation Collaboration 1Carnegie Mellon University, 2Mc Gill University and Mila, 3Bosch Center for AI
Pseudocode Yes A schematic of the DC3 framework is given in Figure 1, and corresponding pseudocode is given in Algorithm 1.
Open Source Code Yes Code for all experiments is available at https://github.com/locuslab/DC3
Open Datasets Yes Each experiment is run 5 times for 10,000 examples x (with train/test/validation ratio 10:1:1). During training, we use examples x with entries drawn i.i.d. from the uniform distribution on [-1, 1]. We assess our method on a 57-node power system test case available via the MATPOWER package.
Dataset Splits Yes Each experiment is run 5 times for 10,000 examples x (with train/test/validation ratio 10:1:1).
Hardware Specification Yes all neural nets and the qpth optimizer were run with full parallelization on a Ge Force GTX 2080 Ti GPU. The OSQP, IPOPT, and PYPOWER optimizers were run sequentially on an Intel Xeon 2.10GHz CPU
Software Dependencies No The paper mentions software like PyTorch, OSQP, IPOPT, and PYPOWER/MATPOWER, but does not provide specific version numbers for these dependencies.
Experiment Setup Yes The following parameters were kept fixed for all neural network-based methods across all experiments... Epochs: 1000 Batch size: 200 Hidden layer size: 200 (for both hidden layers) Correction procedure stopping tolerance: 10^-4 Correction procedure momentum: 0.5