[Colloquium] Seminar Announcement: Polynomial Regression with Derivative Information for Uncertainty Analysis of Simulation Models - TODAY!

[Ninfa Mayorga]

Computation Institute- Data Lunch Seminar (DLS)

Speaker: Oleg Roderick, Mathematics and Computer Science Division,  
Argonne National Laboratory
Date: March 11, 2011
Time: 12:00 PM - 1:00 PM
Location: The University of Chicago, Searle 240A, 5735 S. Ellis Avenue

Polynomial Regression with Derivative Information for Uncertainty  
Analysis of Simulation Models

Abstract:
The task of uncertainty quantification of a simulation model is to  
describe the variability of outputs given uncertainty in the inputs.  
For a model that is both structurally complex, and nonlinear, methods  
based on sampling may be too computationally expensive, while local  
approximations/linearizations are not precise. We use a hybrid  
polynomial regression methods that constructs a surrogate uncertainty  
model using the original model outputs and also first derivatives of  
the outputs as fitting conditions. The method, in the basic form, is  
easy to reproduce, and is effective for a wide class of models.

Aspects of our work to date include: automatic differentiation of the  
code to obtain derivatives, choice of best polynomial basis, use of  
dimensionality reduction to deal with large uncertainty spaces. Our  
main area of application is simulation models of nuclear engineering.

Bio:
Oleg Roderick graduated from Portland State University with a Ph.D. in  
mathematics along with a concentration in allied areas of modern  
physics and atmospheric chemistry. Since 2008, he is with Mathematics  
and Computing Science Division at Argonne as a postdoctoral  
researcher. Oleg's current research interests include uncertainty  
quantification, verification and validation, missing data recovery and  
prediction for complex models and massive data sets. He has used these  
techniques in developing simulation models for nuclear engineering.


Information: Lunch will be provided