CAREER: Design and Application of Scalable Hierarchical Optimization Algorithms by Combining Evolutionary Computation, Machine Learning and Statistics

摘要提案编号：0547013提案题目：职业：结合进化计算、机器学习和统计学的可扩展分层优化算法的设计与应用[ei]名称：Pelikan， MartinPI机构：密苏里大学圣路易斯分校非凸优化的挑战仍然是在工程的所有分支中不断出现的最基本的挑战之一。理解哺乳动物大脑和各种复杂系统中的创造力机制（打破局部最小值或常规的能力）至关重要。基于遗传算法、运筹学等的现有方法被广泛使用，但在扩展到大型复杂问题时存在困难，部分原因是它们甚至没有解决在任何设计领域获得更多经验后如何更好地学习搜索的问题。这个PI是为数不多的采用新方法来完成这项任务的研究人员之一，称为分布算法估计（EDA）。他开发了新的、有原则的方法，能够从一个搜索到另一个搜索进行学习，用于所有设计选择本质上是离散的情况。在这里，他将把工作扩展到连续变量和网络设计的情况下，使用现实世界的测试平台，并探索扩展到更大问题的方法。更广泛的利益：PI将展示他的新方法在药物设计、生物信息学和医学诊断方面的价值。他还将展示新的扩展如何提高由USC/ISI、Cycorp、英特尔、洛克希德、马丁MS2、麻省理工学院、格鲁曼和斯坦福大学组织的search项目的收益，该项目已经将离散版本作为国防部应用的智能认知引擎的一部分。新的跨学科课程将建立在这种新方法所提供的统一知识的基础上。将会有新的K-12课程竞赛和推广，并在这所大学建立一个新的实验室。

AbstractProposal Number: 0547013Proposal Title: CAREER: Design and Application of Scalable Hierarchical Optimization Algorithms by Combining Evolutionary Computation, Machine Learning and StatisticsPI Name: Pelikan, MartinPI Institution: University of Missouri Saint LouisIntellectual Merit. The challenge of nonconvex optimization remains one of the most fundamental challenges occurring again and again in all branches of engineering. It is crucial to understanding the mechanisms for creativity (ability to break out of a local minimum or rut) in the mammalian brain, and in complex systems of all kinds. Existing methods based on genetic algorithms, operations research and the like are widely used, but have difficulties in scaling up to large, complex problems in part because they do not even address how it is possible to learn to search better as one acquires more experience in any design domain. This PI is one of the few researchers taking a new approach to this task, called Estimation of Distribution Algorithms (EDA). He has developed new, principled methods capable of learning from search to search, for the case where all the design choices are discrete in nature. Here he will extend the work to the case of continuous variables and network design, with real-world testbeds, and explore ways to scale up to larger problems.Broader Benefits: The PI will demonstrate the value of his new methods to drug design, bioinformatics and medical diagnostics. He will also show how the new extensions improve the benefits to the CEARCH project organized by USC/ISI, Cycorp, Intel, Lockheed, Martin MS2, MIT, Grumman and Stanford, which is already using the discrete versions as part of an Intelligent Cognitive Engine for DOD applications. New cross-disciplinary courses will be built, building on the unification of knowledge offered by this new approach. There will be new competitions for and outreach to K-12 programs, and a new laboratory created at this university.