CSE Speaker Series – Dr. Ahmad Rushdi

On Friday, March 11 in Cramer 221 from 11:00-12:00, Dr. Ahmad Rushdi will give a talk on Locally-Accurate Surrogate models for Credible High-Dimensional Data Analysis.

ABSTRACT: Engineering and scientific problems have been rapidly growing in dimensionality and complexity. In turn, high-dimensional objective functions have been too expensive to compute, exhausting the runtime and resources’ budgets. For feasible handling of such functions, global surrogate models were introduced to solve problems in optimization, uncertainty quantification, and sensitivity analysis. To construct a surrogate, a few function evaluations are used to construct a computationally-cheap response surface using some approximation method. Typically, this approach suffered from the curse of dimensionality and numerical instabilities along domain bounds and functional discontinuities. In this talk, we present a new method to construct credible global surrogate models with local accuracy in high-dimensional spaces: Voronoi Piecewise Surrogate (VPS) models. The key component in our method is to decompose the high-dimensional parameter space using an implicit Voronoi tessellation around the known function evaluations as seeds. VPS finds cell neighbors via local hyperplane sampling without constructing an explicit mesh. To avoid the intractable complexity of high-d Voronoi cells, we construct an approximate dual Delaunay graph to establish the neighborhood network between cells. Each cell then uses information at its neighbors to build its own local piece of the global surrogate. Due to its piecewise nature, VPS accurately handles smooth functions with high curvature as well as functions with discontinuities, and can be implemented in parallel.
SPEAKER BIO: Dr. Ahmad Rushdi is a research scientist with the Institute for Computational Engineering and Sciences (ICES) at the University of Texas, Austin, currently visiting the Center for Computing Research (CCR) at Sandia National Laboratories, in Albuquerque NM. His collaborative research between UT and Sandia focuses on novel solutions for uncertainty quantification and optimization problems with different engineering, visualization, medical, and computer graphics applications. Dr. Rushdi has received his Ph.D. degree in Electrical and Computer Engineering from the University of California, Davis in 2008, and his MSc and BSc degrees in Electrical Engineering from Cairo University, Cairo, Egypt, in 2004 and 2002. His research interests include signal processing, scientific computing, and uncertainty quantification. He is an active member of the scientific community, a senior member of IEEE, and an ACM member.

 

 

Posted on .