Research map / Probabilistic and scientific modeling / Dynamical systems
Nesting Particle Filters for Experimental Design in Dynamical Systems
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 Inside-Out SMC² (nested particle filters with IBIS inside a particle filter) to compute risk-sensitive EIG objectives for non-exchangeable sequential Bayesian experimental design, and amortizes the design policy via Markovian score climbing with a conditional SMC kernel.
Problem
Sequential Bayesian experimental design for non-exchangeable dynamical systems requires maximizing expected information gain over long horizons; existing amortized approaches rely on sPCE bounds that are biased and need exponentially many contrastive samples in high-EIG regimes.
Approach
Casts sequential BED as risk-sensitive policy optimization equivalent to smoothing in a non-Markovian state-space model, and estimates the marginal dynamics with an inner IBIS parameter filter nested inside an outer particle filter (Inside-Out SMC²), trained by Markovian score climbing with a Rao-Blackwellized conditional SMC kernel instead of contrastive estimators.
Key results
- On the conditionally linear pendulum, IO-SMC² reaches an EIG estimate of 3.53 ± 0.15 versus 2.58 ± 0.17 for iDAD.
- On the stochastic cart-pole, IO-SMC² achieves an EIG estimate of 21.23 ± 0.62, above iDAD's 18.99 ± 0.68, while sPCE bounds saturate at log(10^6).
- On the stochastic double-link, iDAD (11.73 ± 0.45) slightly outperforms IO-SMC² (11.53 ± 0.49) in EIG estimate.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use IO-SMC² instead of iDAD | Sequential design policy · Conditionally linear pendulum | EIG estimate: 2.58 ± 0.17 -> 3.53 ± 0.15 | Helped |
| use IO-SMC² instead of iDAD | Sequential design policy · Nonlinear pendulum | EIG estimate: 3.01 ± 0.29 -> 3.72 ± 0.17 | Helped |
| use IO-SMC² instead of iDAD | Sequential design policy · Stochastic cart-pole | EIG estimate: 18.99 ± 0.68 -> 21.23 ± 0.62 | Helped |
| use IO-SMC² instead of iDAD | Sequential design policy · Stochastic double-link | EIG estimate: 11.73 ± 0.45 -> 11.53 ± 0.49 | Hurt |
| use exact posterior computation instead of IBIS particle approximations (ablation) | IO-SMC² · Conditionally linear pendulum | EIG estimate: 3.53 ± 0.15 -> 3.63 ± 0.18 | No clear effect |
Surprising findings
- On the stochastic double-link, iDAD matches or slightly exceeds IO-SMC² in EIG estimate.
- sPCE bounds for all cart-pole policies hit the upper bound log(10^6) ≈ 13.82, so the bound is uninformative there.
Limitations
- Requires conditional transition densities to be available in closed form.
- Time complexity O(NMT²), with CSMC requiring N proportional to T for stability.
- Choice of tempering parameter η is an open question balancing optimism and particle weight variance.
Experimental setup
Benchmarks: stochastic pendulum stochastic cart-pole stochastic double-link
Metrics: expected information gain estimate sPCE lower bound
Baselines: Random policy PRBS Myopic IO-SMC² iDAD
Builds on: SMC² iterated batch importance sampling implicit deep adaptive design sequential prior contrastive estimation conditional sequential monte carlo
Closest papers by meaning
- PASOA- PArticle baSed Bayesian Optimal Adaptive design (ICML 2024 · similarity 0.87)
- Step-DAD: Semi-Amortized Policy-Based Bayesian Experimental Design (ICML 2025 · similarity 0.87)
- Bayesian Experimental Design Via Contrastive Diffusions (ICLR 2025 · similarity 0.86)
- Amortized Bayesian Experimental Design for Decision-Making (NeurIPS 2024 · similarity 0.86)
- On Estimating the Gradient of the Expected Information Gain in Bayesian Experimental Design (AAAI 2024 · similarity 0.86)
- Maximin Robust Bayesian Experimental Design (NeurIPS 2026 · 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.