Research map / Learning theory and algorithms / Prediction and decision losses
Incorporating Surrogate Gradient Norm to Improve Offline Optimization Techniques
Read the paperCheck your idea against 108,967 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 IGNITE, a model-agnostic regularizer for offline optimization that constrains surrogate sharpness (approximated via surrogate gradient norm) during surrogate training, with a PAC-Bayes-based theory linking empirical to generalized sharpness.
Problem
Offline optimization learns a surrogate from a static dataset, but surrogates predict erratically out-of-distribution, misleading the search toward suboptimal designs; existing conditioning tricks are specific to particular surrogate or search models.
Approach
Define generalized surrogate sharpness (max change of surrogate's expected prediction under norm-bounded parameter perturbation), approximate it by rho times the gradient norm of the expected surrogate, and enforce it as a constraint via a Lagrangian (BDMM), applied as a plug-in regularizer to existing offline optimizers; theory extends PAC-Bayes bounds from loss sharpness to surrogate sharpness.
Key results
- IGNITE improved baseline performance in 35 of 44 cases (79.55%), with a peak improvement of 9.6%.
- IGNITE did not degrade performance in 40 of 44 cases (91%).
- On TF-Bind-10, IGNITE raised ENS-MIN SOTA from 68.1% to 70.5%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add IGNITE to offline optimizers | 11 baseline algorithms · Four Design-Bench tasks | performance: 35 out of 44 cases; average improvement of approximately 1.91%; peak improvement of 9.6% | Helped |
| Add IGNITE to offline optimizers | 11 baseline algorithms · Four Design-Bench tasks | performance: 4 out of 44 cases; average degradation of approximately 0.3%; peak degradation of 0.7% | Mixed |
| Add IGNITE to CMA-ES | CMA-ES · Ant Morphology | performance: 195.5% -> 195.7% | Helped |
| Add IGNITE to ICT | ICT · D'Kitty Morphology | performance: 96.2% | Helped |
| Add IGNITE to ENS-MEAN | ENS-MEAN · TF-BIND-8 | performance: 98.5% -> 98.7% | Helped |
| Add IGNITE to ENS-MIN | ENS-MIN · TF-BIND-10 | performance: 68.1% -> 70.5% | Helped |
| Add IGNITE to REINFORCE | REINFORCE · Superconductor | performance: +2.1% | Helped |
| Add IGNITE to GA | GA · Superconductor | performance: +0.3% | Helped |
| Add IGNITE to REINFORCE | REINFORCE · Chembl | performance: +0.2% | Helped |
| Add IGNITE to GA | GA · Chembl | performance: +0.5% | Helped |
| Set threshold epsilon to 0.01 or 0.05 instead of the selected 0.1 (ablation) | IGNITE with COMs and GA · Ant Morphology and TF-BIND-10 | performance improvement over baseline | Hurt |
| Use excessively high epsilon values (ablation) | IGNITE with COMs and GA · Ant Morphology and TF-BIND-10 | performance improvement over baseline | Hurt |
Surprising findings
- SAM (loss sharpness) outperformed IGNITE with GA on TF-Bind-10 and gave notable gains with REINFORCE on D'Kitty, though its GA integration dropped performance.
- L1 and L2 regularizers degraded performance with GA despite helping REINFORCE.
Experimental setup
Datasets: TF-Bind-8 TF-Bind-10 Ant Morphology D'Kitty Morphology Superconductor ChEMBL
Benchmarks: TF-Bind-8 TF-Bind-10 Ant Morphology D'Kitty Morphology Superconductor ChEMBL
Metrics: 100th percentile oracle score percentage improvement over baseline surrogate sharpness on unseen data
Baselines: BO-qEI CbAS RoMA ICT CMA-ES COMs MINs REINFORCE GA L1 L2 SAM
Builds on: sharpness-aware minimization PAC-Bayes bounds basic differential multiplier method COMs Design-Bench
Closest papers by meaning
- Boosting Offline Optimizers with Surrogate Sensitivity (ICML 2024 · similarity 0.92)
- Learning Surrogates for Offline Black-Box Optimization via Gradient Matching (ICML 2024 · similarity 0.91)
- Offline Model-Based Optimization via Policy-Guided Gradient Search (AAAI 2024 · similarity 0.91)
- ROOT: Rethinking Offline Optimization as Distributional Translation via Probabilistic Bridge (NeurIPS 2025 · similarity 0.89)
- Offline Model-Based Optimization by Learning to Rank (ICLR 2025 · similarity 0.88)
- A Universal Class of Sharpness-Aware Minimization Algorithms (ICML 2024 · similarity 0.86)
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.