Convex-SKM-Net speeds convex constraint repair up to 144.3×
Original title: Convex-SKM-Net: Low-Latency Block Halfspace Corrections for Neural Convex Constraint Satisfaction
Fixed-depth dual PGD keeps SOC-MPC and power-system tests inside a 10^-3 feasibility tolerance.
Convex-SKM-Net is a convex constraint-satisfaction layer that builds linearization halfspaces from residual and subgradient oracles, then runs fixed-depth dual projected-gradient descent without constraint-specific projectors. The predictor and a positive-definite mobility matrix are trained together. On three second-order-cone model-predictive control horizons and six power-system convex relaxations, every test output met a 10^-3 feasibility tolerance. Reported inference speedups reach 144.3× on SOC-MPC and 7.7× on a QCQP benchmark with joint training.
- QCQP mean objective gap is 0.012%, with full test feasibility at 10^-4.
- Learned mobility cuts the five-seed affine-task mean gap by 22.4%.
- Authors prove outer-iterate convergence at fixed dual depth.
- Linear distance convergence assumes Slater's condition and bounded subgradients.
- AI models
- Convex-SKM-Net
From the original article
1 min read at arxiv.orgFast inference and strict constraint satisfaction are essential for learning-based decision-making. Existing methods often rely on constraint-specific projections or feasible-set parameterizations, while iterative repair can be costly.
By Haoyu Zhu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.