Reproduce “Online Bayesian Experimental Design for Partially Observed Dynamical Systems” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sara Pérez-Vieites, Sahel Iqbal, Simo Särkkä, Dominik Baumann
Researchers at Aalto University developed BAD-PODS, an online sequential Bayesian experimental design method for partially observed dynamical systems with intractable likelihoods due to latent states. The approach demonstrates information efficiency competitive with computationally expensive oracle solutions while offering superior computational and memory scaling for long-horizon applications across diverse models.
Key insights
- Gradient-based optimization of experimental designs is achievable in continuous spaces for partially observed dynamical systems by deriving EIG gradients that explicitly account for and marginalize latent states.
- Nested Particle Filters can effectively approximate intractable likelihoods and evidence terms required for EIG estimation, providing reliable online inference of both states and parameters.
- Maintaining a constant computational cost per step through the NPF's recursive update mechanism is crucial for enabling practical, long-horizon online sequential Bayesian experimental design in complex systems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.