Research map / Learning theory and algorithms / Combinatorial and search optimization
Deep Statistical Solvers
Balthazar Donon, Zhengying Liu, Wenzhuo LIU, Isabelle Guyon, Antoine Marot, Marc Schoenauer
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DSS instead of iterative BGS | DSS · Poisson linear systems | time per instance (ms): 2.3 -> 1.8 | Improved |
| use DSS instead of iterative BGS | DSS · Poisson linear systems | loss 50th percentile: 1.7e-2 -> 1.2e-3 | Improved |
| train with the problem loss (DSS) instead of distance to ground truth (proxy) | DSS · Poisson linear systems | loss 50th percentile: 1.6e-2 -> 1.2e-3 | Improved |
| train with the problem loss (DSS) instead of distance to ground truth (proxy) | DSS · Poisson linear systems | NRMSE w/ LU: 1.1e-3 -> 1.6e-3 | Hurt |
| use DSS instead of Newton-Raphson | DSS · AC power flow IEEE 118 | time per instance (ms): 2e1 -> 2e-1 | Improved |
| use DSS instead of Newton-Raphson | DSS · AC power flow IEEE 14 | time per instance (ms): 2e1 -> 1e-2 | Improved |
| use DSS instead of Newton-Raphson | DSS · AC power flow IEEE 118 | loss 50th percentile: 4.2e-14 -> 1.7e-6 | Hurt |
| train with the problem loss (DSS) instead of proxy | DSS · AC power flow IEEE 118 | correlation of theta_i with NR: 99.99% vs 81.31% | Hurt |
| train with the problem loss (DSS) instead of proxy | DSS · AC power flow IEEE 14 | NRMSE of theta_i: 1.7e-3 -> 7.1e-3 | Hurt |
| train with the problem loss (DSS) instead of proxy | DSS · AC power flow | correlation of Pij and Qij with NR: 99.99% | No clear effect |
| train with the problem loss (DSS) instead of proxy | DSS · AC power flow IEEE 118 | loss 50th percentile: 8.3e-6 -> 1.7e-6 | Improved |
| test on meshes of different sizes than training (100 to 1000 nodes vs ~500) | DSS · Poisson super-generalization | correlation with LU: 99.99% | No clear effect |
Closest papers by meaning
- DOGE-Train: Discrete Optimization on GPU with End-to-End Training (AAAI 2024 · similarity 0.86)
- Learning to Relax: Setting Solver Parameters Across a Sequence of Linear System Instances (ICLR 2024 · similarity 0.86)
- Graph Neural Preconditioners for Iterative Solutions of Sparse Linear Systems (ICLR 2025 · similarity 0.85)
- On the Power of Small-size Graph Neural Networks for Linear Programming (NeurIPS 2024 · similarity 0.85)
- Graph Neural PDE Solvers with Conservation and Similarity-Equivariance (ICML 2024 · similarity 0.85)
- Solver-Free Decision-Focused Learning for Linear Optimization Problems (NeurIPS 2025 · similarity 0.85)
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.
