Research map / Probabilistic and scientific modeling / Molecules and proteins
Amortized Bayesian Experimental Design for Decision-Making
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 TNDP, a Transformer neural process with dual heads (query policy and predictive distribution) trained to maximize decision utility gain, for amortized decision-aware Bayesian experimental design.
Problem
Standard amortized Bayesian experimental design selects experiments to maximize parameter information rather than the quality of a downstream decision, and needs extra computation to make that decision.
Approach
Introduces Decision Utility Gain (DUG) as the design reward and a Transformer Neural Decision Process with a query head that proposes designs and a prediction head that approximates the predictive distribution, trained jointly with REINFORCE so both design and decision are amortized in one forward pass.
Key results
- On decision-aware active learning, TNDP achieves significantly higher proportion of correct decisions than GP-based baselines (GP-RS, GP-US, GP-DUS, GP-TEIG, GP-DEIG) across acquisition steps.
- TNDP per-design acquisition takes about 0.015 s versus 572 s for D-EIG and 1558 s for T-EIG, with total 10-round time of 0.31 s.
- On top-k HPO over four HPO-B meta-datasets, TNDP achieves the best utility, especially in the first 10 queries, compared with RS, UCB, EI, PI and PFNs4BO.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TNDP's learned query head instead of random sampling (ablation) | TNDP · Decision-aware active learning | decision accuracy | Helped |
| Use a non-myopic objective instead of the myopic objective (α=0) (ablation) | TNDP · Decision-aware active learning | decision accuracy | Helped |
| Use TNDP instead of non-amortized active-learning methods | Decision-aware active learning · Decision-aware active learning | decision accuracy | Helped |
| Use TNDP instead of BO baselines and PFNs4BO | Top-k hyperparameter optimization · HPO-B meta-datasets | utility | Helped |
| Use TNDP instead of random search | Retrosynthesis planning · Retrosynthesis planning | top-k route utility | Helped |
Limitations
- REINFORCE training can be unstable, especially with sparse rewards.
- Transformer quadratic complexity limits very large query sets.
- Assumes fixed-length horizon, fixed design dimensionality, and a well-specified model.
Experimental setup
Datasets: HPO-B synthetic GP functions non-public retrosynthesis meta-dataset
Benchmarks: HPO-B (ranger, svm, rpart, xgboost) synthetic decision-aware active learning toy synthetic regression retrosynthesis top-k routes
Metrics: proportion of correct decisions average utility (top-k sum of accuracies) acquisition time total time
Baselines: GP-RS GP-US GP-DUS GP-TEIG GP-DEIG TNDP-RS RS UCB EI PI PFNs4BO T-EIG
Builds on: Transformer neural process sPCE amortized BED REINFORCE loss-calibrated inference PFNs4BO
Closest papers by meaning
- ALINE: Joint Amortization for Bayesian Inference and Active Data Acquisition (NeurIPS 2025 · similarity 0.86)
- Step-DAD: Semi-Amortized Policy-Based Bayesian Experimental Design (ICML 2025 · similarity 0.86)
- Nesting Particle Filters for Experimental Design in Dynamical Systems (ICML 2024 · similarity 0.86)
- Inverse decision-making using neural amortized Bayesian actors (ICLR 2025 · similarity 0.84)
- Decision Theoretic Foundations for Conformal Prediction: Optimal Uncertainty Quantification for Risk-Averse Agents (ICML 2025 · similarity 0.84)
- Amortized Active Causal Induction with Deep Reinforcement Learning (NeurIPS 2024 · similarity 0.84)
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.