New BED method handles dynamic constraints with online planning

Bayesian experimental design (BED) has long been a principled framework for data-efficient sequential experiments, but existing methods struggle with dynamic constraints such as budget ceilings, changing costs, or physical limits on how designs evolve. In a new paper accepted at ICML 2026, researchers (Guo, Huang, Zhang, Katt, Kaski, Bharti) propose a novel approach that combines offline pre-training of an amortized policy and posterior network with online multi-step lookahead planning using scenario trees. This hybrid framework allows the system to adapt to constraints on the fly during the experiment, optimizing for information gain while respecting limitations.

The authors demonstrate empirically across several constrained BED tasks that their method produces design sequences that are substantially more informative than those from existing baselines. The computational overhead is modest, making the approach practical for real-world applications in fields like materials science, clinical trials, and robotics. By enabling BED to handle dynamic constraints, this work bridges the gap between principled experimental design and the messy realities of domain-specific experiments, where budgets and physical limitations are always present.