The resolution of many important questions in science, engineering, and operations research requires the solution of a global optimization problem. The objective of this research is to apply state-of the art methodology in global optimization to several hard problems of scientific or engineering interest, and to develop, analyze, and implement new methodology which is capable of achieving a new level of performance on these and other complex applications.
This project will focus on three areas: 1) structural biology problems, especially 3D virus reconstructions from x-ray scattering and cryo-electron microscopy data, 2) pattern classification for seismic oil exploration problems, and 3) system design and evaluation for nonlinear satellite communications links. These problems exhibit cost functions with multiple local minima and a wide variety of characteristics which complicate their optimization, including: a) cost functions that are driven by data that has a multi-scale structure, b) cost functions that are computed analytically but at great expense, c) cost functions that are computed by simulation or by a combination of analytical formulas and simulation (so-called semi-analytical simulation), and d) cost functions that are driven by changing data, where the global minimum must be tracked, possibly in real time.
The first part of the work involves applying current global random search algorithms to get a baseline for the performance and computational characteristics of these methods which have not been studied extensively in the problem areas described above. The second part of the work will involve developing and analyzing algorithms with improved efficiency.
Much of the practicality of global optimization stems from the extraordinary (and still growing) computation resources that are now available. The most cost-effective resources are clusters of off-the-shelf computers. This research group has access to a leading edge machine of this type. Because local searches and simulations can be carried out in parallel with minimal communication overhead, the algorithms to be developed here are ideal candidates for implementation on such computers. Implementation efforts will be focused in this direction.