Research map / Probabilistic and scientific modeling / Variational and generative sampling
PASOA- PArticle baSed Bayesian Optimal Adaptive design
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 PASOA, a sequential Bayesian experimental design method that couples stochastic gradient optimization of a prior contrastive estimation EIG bound with adaptive tempered SMC to track posteriors, with a consistency proof for the design estimators.
Problem
Sequential Bayesian optimal experimental design requires sampling successive intractable posteriors, and maximizing expected information gain pushes successive posteriors apart, which degrades standard SMC accuracy.
Approach
Combine stochastic gradient optimization of the PCE contrastive EIG bound with adaptive tempered SMC samplers that recycle particles across design steps; tempering keeps SMC accurate while product-form particle approximations supply contrastive samples.
Key results
- On source location, PASOA improves SPCE by about 30% over RL-BOED and gives much higher SNMC.
- Wasserstein distance to true sources is two orders of magnitude lower than RL-BOED/VPCE.
- Tempering lets PASOA with 5e3 particles outperform naive SMC with 1e6 particles at lower computation time.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PASOA instead of RL-BOED for sequential design | Source location · Source location | SPCE: 30% | Helped |
| Use PASOA instead of RL-BOED for sequential design | Source location · Source location | SNMC | Helped |
| Use PASOA instead of RL-BOED for sequential design | Source location · Source location | L2 Wasserstein distance: two order of magnitude lower | Helped |
| Add tempering to SMC (ablation) | PASOA / SMC · Source location | posterior estimation: 5\mathrm{e}^3 vs 10^6 | Helped |
| Add tempering to SMC (ablation) | PASOA / SMC · Source location | L2 Wasserstein distance | Helped |
| Use PASOA instead of VPCE and RL-BOED | Sequential design · CES | SPCE | Helped |
| Use PASOA instead of VPCE and RL-BOED | Sequential design · CES | L2 Wasserstein distance | Helped |
| Use PASOA instead of the random baseline | Sequential design · Source location | information gain | Helped |
| Replace tempered SMC with non-tempered SMC (ablation) | PASOA · Source location | posterior estimation | Hurt |
Surprising findings
- Greedy PASOA outperforms the non-myopic RL-BOED approach, suggesting accurate posterior estimation matters more than multi-step planning.
Limitations
- Posterior inference may fail in high-dimensional parameter spaces, undermining the design advantage.
- Greedy myopic approach; non-myopic planning not addressed.
- Consistency results rely on simplified assumptions and an unquantified optimization error term.
Experimental setup
Datasets: simulated source location experiments simulated CES basket experiments
Benchmarks: source location CES
Metrics: SPCE lower bound on total EIG SNMC upper bound on total EIG L2 Wasserstein distance to true parameters
Baselines: RL-BOED VPCE naive SMC (non-tempered PASOA) random design
Builds on: prior contrastive estimation (PCE) bound adaptive tempered SMC (ESS-based tempering) variational prior contrastive estimation (VPCE) sequential prior contrastive estimation (SPCE)
Closest papers by meaning
- Bayesian Adaptive Calibration and Optimal Design (NeurIPS 2024 · similarity 0.88)
- On Estimating the Gradient of the Expected Information Gain in Bayesian Experimental Design (AAAI 2024 · similarity 0.88)
- Nesting Particle Filters for Experimental Design in Dynamical Systems (ICML 2024 · similarity 0.87)
- Bayesian Experimental Design Via Contrastive Diffusions (ICLR 2025 · similarity 0.87)
- Maximin Robust Bayesian Experimental Design (NeurIPS 2026 · similarity 0.86)
- Step-DAD: Semi-Amortized Policy-Based Bayesian Experimental Design (ICML 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.