KDI: Global Adaptive Optimization for Structural Biology anand Other Complex Signal Reconstruction, Pattern Recognition and System Design Problems

在科学、工程和运筹学中，许多重要问题的解决都需要解决全局优化问题。本研究的目的是将最先进的全局优化方法应用于科学或工程领域的一些难题，并开发、分析和实施能够在这些和其他复杂应用中实现新性能水平的新方法。该项目将重点关注三个领域：1)结构生物学问题，特别是基于x射线散射和低温电子显微镜数据的三维病毒重建；2)地震石油勘探问题的模式分类；3)非线性卫星通信链路的系统设计和评估。这些问题表现出具有多个局部极小值的成本函数和各种各样的特征，这些特征使其优化复杂化，包括：A)由具有多尺度结构的数据驱动的成本函数，b)通过分析计算但成本很高的成本函数，c)通过模拟或分析公式与模拟（所谓的半分析模拟）相结合计算的成本函数，以及d)由变化数据驱动的成本函数，其中必须跟踪全局最小值，可能是实时的。第一部分工作涉及应用当前的全局随机搜索算法来获得这些方法的性能和计算特性的基线，这些方法在上述问题领域尚未得到广泛的研究。第二部分的工作将涉及开发和分析提高效率的算法。全局优化的许多实用性源于现在可用的非凡（且仍在增长的）计算资源。最具成本效益的资源是现成的计算机集群。这个研究小组可以使用这种类型的最先进的机器。由于局部搜索和模拟可以以最小的通信开销并行执行，因此这里开发的算法是在此类计算机上实现的理想候选算法。执行工作将集中在这个方向。

DoerschuckThe 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.