Research map / Probabilistic and scientific modeling / Bayesian optimization
Offline Model-Based Optimization by Learning to Rank
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 RaM, a ranking-based surrogate for offline model-based optimization that trains MLPs with learning-to-rank losses (RankCosine, ListNet) on augmented random design lists, and argues MSE correlates poorly with final design quality while a new AUPCC metric correlates better.
Problem
Offline model-based optimization relies on regression surrogates trained with MSE, which overestimate scores out of distribution and mislead gradient ascent; MSE may be a poor proxy for selecting good designs.
Approach
Replace MSE with learning-to-rank losses trained on randomly sampled design lists (data augmentation), evaluate surrogates with a new ranking metric AUPCC, and normalize predictions (output adaptation) so gradient ascent hyperparameters transfer from regression models.
Key results
- RaM-ListNet achieves the best average rank of 2.2 across 22 methods and RaM-RankCosine 2.7, versus 5.9 for the next best (BDI).
- OOD-AUPCC correlates more strongly with final score than OOD-MSE, e.g., Spearman 0.503 vs 0.243 on D'Kitty and 0.520 vs -0.239 on TF-Bind-8, with average improvement of 0.364.
- Replacing MSE with ListNet and output adaptation improves most combined methods, e.g., BO-qEI on Superconductor +33.2%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RaM-ListNet instead of the strongest listed baseline on each task | Offline MBO on Design-Bench · Ant, D'Kitty, Superconductor, TF-Bind-8, and TF-Bind-10 | 100th percentile normalized score | Mixed |
| Replace MSE with ListNet | Existing offline MBO methods · Design-Bench tasks | 100th percentile normalized score: +0.0% to +33.2%; -0.4% and -0.3% | Mixed |
| Use data augmentation to construct ranking lists instead of treating a batch as one list (ablation) | RaM · Design-Bench tasks | 100th percentile normalized score | Helped |
| Add output adaptation by normalizing predictions using in-distribution mean and standard deviation (ablation) | RaM · Design-Bench tasks | 100th percentile normalized score | Helped |
| Use OOD-AUPCC instead of OOD-MSE to assess surrogate models | Offline MBO surrogate evaluation · Ant, D'Kitty, Superconductor, TF-Bind-8, and TF-Bind-10 | Spearman correlation with 100th percentile score: +0.096; +0.260; +0.217; +0.759; +0.486 | Helped |
| Train with RankCosine rather than MSE | Linear ranking model under heavy-tailed noise · Quadratic objective with noise scale α∈{10, 15, 20, 50, 100} | learned weight w: w_MSE: -1.68, -0.75, -2.98, -14.99, -10.05; w_RankCosine: 0.88, 0.90, 0.74, 0.94, 0.98 | Helped |
| Train with RankCosine rather than MSE | Linear ranking model under heavy-tailed noise · Noise probability p∈{0.1, 0.2, ..., 1.0} | learned weight w: w_MSE: 1.61, -0.75, -4.53, -8.46, -7.76, -10.73, -12.87, -17.28, -20.21, -22.66; w_RankCosine: 0.86, 0.90, 1.01, 0.88, 1 | Helped |
Surprising findings
- MSE has weak or negative correlation with final design quality in OOD regions, despite being the standard training loss.
- In a toy heavy-tailed noise setting, linear models trained with MSE yield negative slopes while RankCosine yields positive slopes.
Limitations
- Theoretical generalization bound is for linear function classes and i.i.d. data, not OOD regions; the paper says OOD ranking analysis is limited.
- Benchmarks restricted to Design-Bench tasks; three tasks excluded following prior work.
- Gains on some tasks are small or negative (e.g., Tri-Mentoring with ListNet on D'Kitty -0.4%).
Experimental setup
Datasets: Ant Morphology D'Kitty Morphology Superconductor TF-Bind-8 TF-Bind-10
Benchmarks: Design-Bench Ant Design-Bench D'Kitty Design-Bench Superconductor Design-Bench TF-Bind-8 Design-Bench TF-Bind-10
Metrics: 100th percentile normalized score mean rank Spearman correlation OOD-MSE OOD-AUPCC
Baselines: BO-qEI CMA-ES REINFORCE Gradient Ascent CbAS MINs DDOM BONET GTG COMs RoMA IOM
Builds on: RankCosine ListNet COMs Tri-Mentoring Design-Bench
Closest papers by meaning
- Learning Surrogates for Offline Black-Box Optimization via Gradient Matching (ICML 2024 · similarity 0.89)
- Incorporating Surrogate Gradient Norm to Improve Offline Optimization Techniques (NeurIPS 2024 · similarity 0.88)
- Offline Model-Based Optimization via Policy-Guided Gradient Search (AAAI 2024 · similarity 0.88)
- Boosting Offline Optimizers with Surrogate Sensitivity (ICML 2024 · similarity 0.87)
- Offline Multi-Objective Optimization (ICML 2024 · similarity 0.87)
- ROOT: Rethinking Offline Optimization as Distributional Translation via Probabilistic Bridge (NeurIPS 2025 · similarity 0.87)
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.