Title: "Adaptive Parameter Space Exploration with Gaussian Process Trees"

Speaker: Herbert K. H. Lee, University of California, Santa Cruz

Date/Time: Wednesday, April 6, 2005, 11:00-12:00 (MT)

Location: Building 980, Room 24 (Sandia NM), Building 921, Room 137 (Sandia CA)

Brief Abstract: Traditionally, to obtain even a qualitative understanding of the output of a computer simulation, runs have been made over a complete grid of input parameter configurations.  For large scale simulations, such sweeps can be prohibitively expensive, and fixed designs such as Latin hypercubes can be still difficult yet inefficient.  Thus, there is a need for computationally inexpensive surrogate models that can be used in place of simulation to adaptively select new settings of input parameters and map the response with far fewer simulation runs. We provide a general methodology for modeling and adaptive sampling to greatly speed up parameter sweeps. Binary trees are used to recursively partition the input space, and Gaussian process models are fit within each partition. Trees facilitate non-stationarity and a Bayesian interpretation provides a measure of uncertainty in the sample space which can be used to guide future sampling. Our methods are illustrated on several examples, including the motivating example involving computational fluid dynamics simulation of a NASA reentry vehicle.