Deep Adaptive Bayesian Screening

2026-07-18

AI Summary

This work addresses adaptive factor screening in high-dimensional discrete design spaces under limited experimental budgets. It proposes a novel approach that integrates amortized Bayesian optimal experimental design with Gibbs posterior inference. The method employs an offline-trained policy network to sequentially select the most informative experiments and incorporates a spike-and-slab prior with strong heredity constraints to model factor sparsity and interaction effects. Crucially, it analytically marginalizes out nuisance parameters and noise variance. By uniquely combining amortized experimental design with Gibbs posterior inference, the proposed framework achieves substantially higher screening accuracy and scalability compared to both classical and Bayesian baselines on real-world benchmark tasks, particularly under tight budget constraints.

Abstract

We introduce Deep Adaptive Bayesian Screening (DABS), a method for performing adaptive factorial screening in high-dimensional discrete design spaces. DABS learns a policy network offline to sequentially select informative experiments, amortizing Bayesian Optimal Experimental Design. It handles binary designs, incorporates sparsity and interactions via a spike-and-slab prior with strong heredity. The model is trained using a contrastive lower bound on information about factor activity with nuisance effect sizes and noise variance analytically integrated out. Unlike prior amortized Bayesian design approaches, DABS also integrates Gibbs posterior inference at deployment, yielding posterior probabilities of factor activity and credible intervals on effect sizes. We demonstrate DABS on screening problems calibrated to real-world benchmarks and show it achieves superior accuracy and scalability over classical and Bayesian baselines under tight experimental budgets.

Problem

• adaptive factorial screening
• high-dimensional discrete design
• Bayesian experimental design
• factor activity
• sparsity and interactions

Innovation

• Deep Adaptive Bayesian Screening
• Amortized Bayesian Experimental Design
• Spike-and-Slab Prior
• Strong Heredity
• Gibbs Posterior Inference