DABS: New deep Bayesian screening method cuts experiment budgets drastically

Deep Adaptive Bayesian Screening (DABS) is a novel method for adaptive factorial screening in high-dimensional discrete design spaces. Developed by Lejeune Herman, Strouwen, Suykens, and Goos, DABS trains a policy network offline to sequentially choose the most informative experiments, effectively amortizing Bayesian optimal experimental design. It handles binary designs and incorporates sparsity and interactions via a spike-and-slab prior with strong heredity constraints. The model uses a contrastive lower bound on information about factor activity, with nuisance effect sizes and noise variance integrated out analytically. At deployment, DABS integrates Gibbs posterior inference, providing posterior probabilities of factor activity and credible intervals on effect sizes.

DABS is tested on screening problems calibrated to real-world benchmarks and outperforms both classical and Bayesian baselines in accuracy and scalability, especially under tight experimental budgets. By automating the selection of experiments, DABS dramatically reduces the number of trials needed to identify active factors and interactions, making it ideal for fields like drug discovery, materials science, and industrial process optimization. The method's ability to provide uncertainty quantification (credible intervals) adds practical value for decision-making. This work represents a significant advance in Bayesian optimal experimental design, enabling more efficient and reliable screening in high-dimensional spaces.