Speaker
Description
Sequential experiments that adapt to incoming data are more efficient than static designs, but the required posterior inference and design optimization between steps are usually too expensive to run online. Deep Adaptive Design [DAD, Foster et al., 2021] sidesteps this by training a neural network policy onine that maps any experimental history to the next design point in a single forward pass. We apply DAD to three problems central to industrial statistics: model-based design of experiments, factorial screening, and Bayesian optimization. On a Monod bioreactor, the adaptive policy reduces posterior RMSE by about 30% relative to the best static design [Strouwen and Micluµa-Câmpeanu, 2026]. We describe ongoing work on adaptive screening. On the multimodal 2D Rastrigin benchmark, the learned policy achieves roughly an order-of-magnitude lower simple regret than the best tested classical baseline.