Targeted Adaptive Design

Abstract: We present Targeted Adaptive Design (TAD), a new data-driven method that aims at efficiently locating optimal control parameters that would yield a target output design in a multi-dimensional input/multi-dimensional output environment. At each stage of the algorithm, a Vector-Valued Gaussian Process (VVGP) surrogate is placed on the unknown experimental response, and N new sample locations and target control parameters are optimized to maximize a novel acquisition function based on the expected predictive log-likelihood of the target parameters at the target location, conditioned on the acquired data and the N latent future sample. A dynamic model validation based on chi-squared statistics is also integrated at each stage of the algorithm. Our method incorporates robust Uncertainty Quantification (UQ) from the ground up and is well-suited for design problems arising from advanced manufacturing. We showcase the performance of the algorithm on a simulated design problem and compare it to algorithms based on grid and random sampling data acquisition strategies.