Date: April 16th (Wednesday)

Time: 9 am – 11 am EST

Location: Groseclose 402

Shangkun Wang

Industrial Engineering PhD Student

H. Milton Stewart School of Industrial and Systems Engineering (ISyE)
Georgia Institute of Technology

Committee

1 Dr. Roshan Joseph (Advisor, ISyE)

2 Dr. Weijun Xie (ISyE)

3 Dr. Nick Sahinidis (ISyE)

4 Dr. Enlu Zhou (ISyE)

5 Dr. Peng Chen (CSE)

Abstract

Active learning plays a crucial role in computer experiments, particularly when the simulation models are computationally expensive. In emulation, active learning sequentially selects input points to improve the surrogate model’s accuracy in regions of high uncertainty, enabling efficient approximation of complex simulators. For inverse design, where the goal is to identify inputs that produce desired outputs, active learning focuses sampling in regions of the input space that are likely to yield responses that cover the output space. This thesis presents new methods of active learning to tackle challenges in emulation and inverse design problems.

Chapter 1 proposes a new experimental design framework called output space-filling design (OSFD) that aims to sequentially fill the space of the outputs (responses or features). Chapter 2 extends the OSFD framework and proposes a new criterion for sequential design. The new criterion enables explicit balancing of the input and output points with a user-specified weight. Chapter 3 deals with the active learning problem for emulation by proposing a new non-stationary model called heteroskedastic rational kriging. Chapter 4 develops the SFDesign package which offers a comprehensive suite of functions to construct various types of space-filling designs , which can be used as initial designs in active learning.