Research map / Probabilistic and scientific modeling / Bayesian optimization
Generative Adversarial Model-Based Optimization via Source Critic Regularization
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 adaptive source critic regularization (aSCR), a WGAN-style critic penalty with dynamically tuned Lagrange multiplier, added to Bayesian optimization (GABO) and gradient ascent (GAGA) to keep offline MBO designs in-distribution.
Problem
Offline model-based optimization: surrogate models overestimate objective values for out-of-distribution designs, so optimizing against them yields poor oracle scores.
Approach
Constrain surrogate optimization with a Wasserstein source critic that scores generated designs against the offline dataset, cast as a Lagrangian whose multiplier is set adaptively via a dual-based search over alpha on Monte Carlo prior samples, making the method task- and optimizer-agnostic.
Key results
- GABO (BO with aSCR) achieves average rank 3.8 on top-1 evaluation vs 5.5 for next best (COM) across eight tasks.
- GABO achieves average rank 3.0 on top-128 evaluation vs 4.6 for next best (BootGen).
- On LogP, GABO proposes a ring-free candidate with penalized logP 22.1, above the offline dataset best of 11.3.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add adaptive source critic regularization to BO-qEI | Bayesian optimization on eight MBO tasks · Eight MBO tasks, top-1 | average rank (lower is better): 8.8 -> 3.8 | Helped |
| add adaptive source critic regularization to BO-qEI | Bayesian optimization on eight MBO tasks · Eight MBO tasks, top-128 | average rank (lower is better): 6.6 -> 3.0 | Helped |
| add adaptive source critic regularization to gradient ascent | Gradient ascent on eight MBO tasks · Eight MBO tasks, top-1 | average rank (lower is better): 9.0 -> 7.4 | Helped |
| add adaptive source critic regularization to gradient ascent | Gradient ascent on eight MBO tasks · Eight MBO tasks, top-128 | average rank (lower is better): 11.0 -> 7.6 | Helped |
| adapt α rather than fix it to a constant (ablation) | GABO · Eight MBO tasks, top-1 | average rank (lower is better): fixed α ranks: 4.5, 3.9, 3.6, 2.4, 4.8; aSCR: 1.9 | Helped |
| adapt α rather than fix it to a constant (ablation) | GABO · Eight MBO tasks, top-128 | average rank (lower is better): fixed α ranks: 4.3, 4.8, 2.9, 3.4, 3.4; aSCR: 2.4 | Mixed |
| adapt α rather than fix it to a constant (ablation) | GAGA · MBO tasks, top-1 and top-128 | average rank (lower is better) | Helped |
| constrain the objective with adaptive source critic regularization instead of unconstrained BO-qEI | LogP optimization · LogP | distance covariance with oracle scores: 0.392 ± 0.040 -> 0.535 ± 0.067 | Helped |
Surprising findings
- Constant alpha values give no consistently best setting across tasks, so a fixed regularization strength is task-dependent.
- GABO does not improve the 90th-percentile suboptimal-design metric over several baselines, which the authors attribute to BO's high variance.
Limitations
- Requires prior knowledge of the input distribution (latent prior) to make the dual computation tractable.
- Only evaluated with BO-qEI and gradient ascent as parent optimizers.
- Poor performance on the 90th-percentile suboptimal-design metric relative to some baselines.
Experimental setup
Datasets: Branin (synthetic) Guacamol Design-Bench TF-Bind-8 Design-Bench GFP Design-Bench UTR Design-Bench ChEMBL D'Kitty Warfarin pharmacogenetic dataset
Benchmarks: Branin LogP TF-Bind-8 GFP UTR ChEMBL D'Kitty Warfarin
Metrics: top-1 oracle score (100th percentile) top-128 oracle score 90th percentile oracle score average rank
Baselines: BO-qEI TuRBO L-BFGS CMA-ES simulated annealing COM RoMA BDI DDOM BONET ExPT ROMO
Builds on: Wasserstein GAN conditioning by adaptive sampling conservative objective models Robust Model Adaptation variational autoencoder
Closest papers by meaning
- Offline Model-Based Optimization via Policy-Guided Gradient Search (AAAI 2024 · similarity 0.88)
- Robust Model-Based Optimization for Challenging Fitness Landscapes (ICLR 2024 · similarity 0.86)
- Flow Density Control: Generative Optimization Beyond Entropy-Regularized Fine-Tuning (NeurIPS 2025 · similarity 0.86)
- Variational Search Distributions (ICLR 2025 · similarity 0.85)
- Adaptive Divergence Regularized Policy Optimization for Fine-tuning Generative Models (NeurIPS 2025 · similarity 0.85)
- Learning Surrogates for Offline Black-Box Optimization via Gradient Matching (ICML 2024 · 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.