B1494
Title: Bayesian sequential experimental design for Gaussian-process-based partially linear model
Authors: Shunsuke Horii - Waseda University (Japan) [presenting]
Abstract: The problem of designing experiments in a sequential manner to accurately estimate the parameters of a partially linear model that employs a Gaussian process prior is investigated. In an active learning context, the experimenter adaptively selects the data to be collected to meet their objectives efficiently. These objectives can differ, ranging from minimizing the probability of classification errors to enhancing the precision of parameter estimation for the underlying data-generating process. The primary objective of this research is to refine the estimation accuracy of the parametric component of a partially linear model. Given certain conditions, this parametric component can be viewed as a causal parameter, such as the average treatment effect (ATE) or the average causal effect (ACE). A Bayesian algorithm is introduced for sequential experimental design specifically tailored for partially linear models with a Gaussian process prior. The efficacy of the proposed approach is demonstrated through computational experiments using both synthetic and semi-synthetic datasets.
