CAREER: Search-Based Optimization of Combinatorial Structures via Expensive Experiments

科学和工程应用中的许多设计优化问题都涉及执行耗费资源(计算或物理)代价高昂的实验。这些实验通常由直觉指导，由人类工程师和科学家根据先前实验的结果进行一系列调查。当设计空间是在设计变量(例如，集合、序列和图)中具有丰富结构的组合时，这种实验性设计过程可能是非常具有挑战性的。这个为期五年的项目是一个综合的研究、教育和推广计划，专注于通过为此类实验开发基于人工智能(AI)的新算法来转变优化组合设计空间的做法。本项目的研究目标是开发一种新的基于搜索的学习和优化框架，以解决与优化由离散和混合(离散和连续设计变量的混合)结构组成的组合设计空间相关的挑战。该框架紧密结合了机器学习和人工智能搜索的先进技术，通过对可用资源预算和实验可能提供的潜在信息的有用性进行推理，智能地探索设计空间。基于搜索的框架将扩展到两个新的环境，以实现提高设计优化的资源效率的目标。首先，实验产生的边信息将被适当地建模和利用。其次，在精确度和资源消耗之间权衡的多保真实验将基于其可用性而得到利用。该项目将通过与来自这些应用领域的领域专家的密切合作，将开发的算法应用于电子设计自动化、材料设计和合成微生物设计领域。该项目开发的技术将通过开放源码软件提供给学术界和工业界。研究成果将通过研究论文、会议报告、教程和短期课程广泛传播，以最大限度地造福于科学界。教育和外展活动将包括一项新颖的大使计划，以提高社区大学生(包括未被充分代表的少数族裔)对计算机科学职业的兴趣；让本科生参与研究项目；在华盛顿州立大学为工程师和科学家举办关于数据驱动的设计优化的短期暑期课程；以及通过华盛顿州立大学现有的名为LSAMP的计划，在计算机科学和工程领域招募和指导未被充分代表的少数群体。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Many design-optimization problems in science and engineering applications involve performing experiments that are expensive in terms of the consumed resources (computational or physical). These experiments are often guided by intuition and performed by human engineers and scientists in an series of investigations informed by results of prior experiments. This experimental design process can be very challenging when the design space is combinatorial with rich structure among the design variables (e.g., sets, sequences, and graphs). This five-year project is an integrated research, education, and outreach program focused on transforming the practice of optimizing combinatorial design spaces by developing new artificial intelligence (AI) based algorithms for such experiments. The research goal of this project is to develop a new search-based learning and optimization framework to address the challenges associated with optimizing combinatorial design spaces consisting of discrete and hybrid (mixture of discrete and continuous design variables) structures. This framework tightly integrates advances in machine learning and AI search to intelligently explore the design space by reasoning about the available resource budget and the usefulness of potential information the experiments may provide. The search-based framework will be extended to two novel settings towards the goal of improving the resource-efficiency for design optimization. First, the side-information generated by the experiments will be modeled and exploited appropriately. Second, multi-fidelity experiments that trade off accuracy and consumed resources will be leveraged based on their availability. The project will apply the developed algorithms to revolutionize the areas of electronic design automation, design of materials, and design of synthetic microbiomes via close collaboration with domain experts from these application areas. The techniques developed in this project will be made available to academia and industry through open-source software. Results will be disseminated widely through research papers, conference presentations, tutorials, and short courses to maximize the benefits to the scientific community. Educational and outreach activities will include a novel Ambassador program to improve the interest of community college students including under-represented minorities in computer science careers; involving undergraduate students in research projects; a short summer-course on data-driven design optimization for engineers and scientists at WSU; and recruiting and mentoring under-represented minority groups in computer science and engineering through an existing program called LSAMP at Washington State University.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.