DOI
https://doi.org/10.25772/GG9E-SF44
Defense Date
2022
Document Type
Dissertation
Degree Name
Doctor of Philosophy
Department
Systems Modeling and Analysis
First Advisor
Dr. David Edwards
Second Advisor
Dr. Edward Boone
Third Advisor
Dr. D'arcy Mays
Fourth Advisor
Dr. Paul Brooks
Abstract
Fully sequential optimal Bayesian experimentation can offer greater utility than both traditional Bayesian designs and greedy sequential methods, but practically cannot be solved due to numerical complexity and continuous outcome spaces. Approximate solutions can be found via approximate dynamic programming, but rely on surrogate models of the expected utility at each trial of the experiment with hand-chosen features or use methods which ignore the underlying geometry of the space of probability distributions. We propose the use of Gaussian process models indexed on the belief states visited in experimentation to provide utility-agnostic surrogate models for approximating Bayesian optimal sequential designs which require no feature engineering. This novel methodology for approximating Bayesian optimal sequential designs is then applied to conjugate models and to particle approximations for different batch sizes.
Rights
© The Author
Is Part Of
VCU University Archives
Is Part Of
VCU Theses and Dissertations
Date of Submission
8-12-2022