Research map / Label-efficient and robust learning / Self-supervised and contrastive
Neural Compression of Long ADMM Trajectory for Multiparametric Quadratic Program
Liang Wu, Bo Yang, Xu Yang, Honghui Zheng, Yilin Mo, Jan Drgona
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TQPNet, a self-supervised neural compression framework that compresses long ADMM optimization trajectories for multiparametric quadratic programs into a compact neural warm start followed by a few ADMM iterations.
Problem
Solving large-scale multiparametric quadratic programs (mpQPs) in real time requires thousands of iterative ADMM updates, a computational bottleneck for model predictive control and learning-enabled systems.
Approach
Instead of predicting only a solution, a neural network is trained self-supervised to compress an entire long ADMM trajectory into a compact warm start, after which only a few ADMM iterations are run to refine it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add MLP(ReLU) learned warm start before ADMM(ReLU) refinement (ablation) | low-dim ADMM(ReLU) · random mpQP n=100, m=200 | iterations to residual 1e-7: 4729 -> 2778 | Improved |
| add 50 ADMM(ReLU) refinement layers after the MLP instead of MLP only (ablation) | TQPNet · random mpQP n=100, m=200 | wall-clock to residual 1e-7: 7.777 ms -> 6.034 ms | Improved |
| add only 1 to 5 ADMM(ReLU) refinement layers after the MLP (ablation) | TQPNet · random mpQP n=100, m=200 | wall-clock to residual 1e-7: 7.777 ms -> 8.032 ms (1 layer), 7.751 ms (5 layers) | No clear effect |
| use low-dimensional ADMM-ReLU instead of full-dimensional ADMM | ADMM for QP · n=100, m=200 random mpQP | convergence to 1e-7: full-dim >10^4 iterations -> 4729 | Improved |
| increase unrolling depth K_train from 10 to 25 (ablation) | TQPNet training · random mpQP n=100, m=200 | median test rel error (K_eval=10): 5.8e-01 -> 5.7e-02 | Improved |
| increase unrolling depth K_train beyond 25 (to 50) (ablation) | TQPNet training · random mpQP n=100, m=200 | median test rel error (K_eval=10): 5.7e-02 -> 5.6e-02 | No clear effect |
| increase unrolling depth K_train from 10 to 50 on small problem (ablation) | TQPNet training · random mpQP n=10, m=20 | median test rel error (K_eval=10): 5.3e-03 -> 3.6e-03 (K=25) -> 1.4e-02 (K=50) | Mixed |
| use self-supervised KKT-residual training instead of label-supervised OSQP regression (TQPNet vs LQPNet) (ablation) | MLP(ReLU) warm start · in-distribution n=100, m=200 | median relative error: 1.5e-02 -> 5.8e-02 | Hurt |
| use self-supervised KKT-residual training instead of label-supervised under distribution shift (ablation) | MLP(ReLU) warm start · OOD p in U[-3,3] | TQP-LQP error gap: 3.9x -> 1.8x gap | Improved |
| use learned warm start instead of cold-start ADMM(ReLU) (K=50) (ablation) | MPC closed loop · random MPC scales | tracking performance | Improved |
| use TQPNet instead of OSQP as problem size grows | MPC QP solving · random MPC, n+m from 60 to 2800 | online latency: OSQP 0.03 ms -> 13.06 ms vs TQPNet 1.31 ms -> 1.88 ms | Mixed |
| use label-free TQPNet training instead of OSQP-labelled LQPNet training | offline training · random MPC n+m=2800 | offline time: 590 s -> 334 s | Improved |
Experimental setup
Builds on: ADMM
Closest papers by meaning
- Deep Distributed Optimization for Large-Scale Quadratic Programming (ICLR 2025 · similarity 0.85)
- Expressive Power of Graph Neural Networks for (Mixed-Integer) Quadratic Programs (ICML 2025 · similarity 0.84)
- NeuralSolver: Learning Algorithms For Consistent and Efficient Extrapolation Across General Tasks (NeurIPS 2024 · similarity 0.84)
- Optimizing ADMM and Over-Relaxed ADMM Parameters for Linear Quadratic Problems (AAAI 2024 · similarity 0.84)
- Learning Provably Improves the Convergence of Gradient Descent (NeurIPS 2025 · similarity 0.83)
- Principled Data Augmentation for Learning to Solve Quadratic Programming Problems (NeurIPS 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
