This project is focused on the the development of scalable methods for electronic excitation and optical responses of nanostructures. This is a joint BES/OASCR funded project that involves LBNL, UCB, University of Minnesota, University of Texas, and Princeton.
This project is focused on the the development of optimization algorithms for large-scale scientific and engineering design problems. Usually these problems involve computationally expensive objective functions with either no derivative information or inaccurate gradients. We have been concentrating on both derivative-free methods and parallel optimization methods that use a combination of direct search methods and Newton methods.
This is a new area of research that is part of a recently founded reading group on machine learning at LBNL. In addition to the group's web pages , I've also compiled a short overview of ensemble methods for machine learning at Ensemble Methods for Machine Learning . To see some Matlab files for running some simple experiments on reconstructing high-energy physics tracks using Exemplars go to Track Examples .
Another area of research is the development of object-oriented software for scientific computing. This work has led to the development of a package for nonlinear optimization called OPT++ . If you'd like to learn more about this package, the documentation is available at OPT++ Documentation
Optimization in Statistics. These projects involve the use of optimization in statistical problems. One such project involves the use of optimization for computing MLE's to determine the best family of distributions for a given data set.
Molecular Conformation. A past interest of mine has been the use of nonlinear optimization methods for solving the molecular conformation problem. This work led to several papers on the use of direct search methods as competitors to genetic algorithms and simulated annealing.