Scalable Algorithms for Deterministic Global Optimization With Parallel Architectures

使用并行架构实现确定性全局优化的可扩展算法

• 批准号: 2330054
• 负责人: Matthew Stuber
• 金额: $ 34.7万
• 依托单位: University of Connecticut
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2024
• 资助国家: 美国
• 起止时间: 2024-03-01 至 2027-02-28
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2330054&HistoricalAwards=false
• 关键词: Scalable; Algorithms; Deterministic; Global; Optimization

Complex systems are everywhere: from agricultural supply chains, wastewater treatment and public water systems, to the energy/power infrastructure that heats, cools, and light residential and industrial buildings. Decarbonizing process industries, especially as they relate to food, energy, and water, is of particular and timely importance. Engineers have an ever-constant mission of improving the health, safety, and robustness of these complex systems that support and improve society. However, innovation inherently increases complexity and, therefore, the efforts to solve complex and interconnected challenges, which include designing new systems and improving existing systems, rely heavily on computational modeling, simulation, and optimization-based approaches. There are two major challenges that this proposal aims to address: (1) the current performance of computational optimization approaches limits their applicability to simplified, lower-complexity problems, and (2) university engineering programs often lack cohesive computational-thinking activities throughout their curricula. Solving (1) will alleviate the current computational bottlenecks and broaden the scale and scope of complex problems that can be solved. Solving (2) will not only help train the next generation of engineers on computational modeling approaches but improve their overall problem-solving skills.The research objective of this project is to develop scalable deterministic global optimization (DGO) algorithms and open-source software implementations by exploiting alternative stream computing architectures for parallelization, to enable the solution of higher complexity models that include first-principles models and machine learning elements involving nonlinear (partial) differential and algebraic equations. The significance of the proposed work lies in unlocking the massive parallel computing performance of graphical processing units (GPUs) for DGO with the development of a new branch-and-bound deterministic search algorithm. The result will be a significant speedup over the current state of the art, which will enable the solution of larger-scale higher-complexity problems that arise in food-energy-water (FEW) applications, among others. The first major innovation is a method for automatically generating source code representations of convex/concave relaxations of nonconvex functions in the optimization formulation, and subgradients thereof, on arbitrary domains of interest. The second major innovation is a scalable GPU-compatible parallel DGO algorithm and open-source software implementation for the guaranteed solution of nonconvex programs. This project will align the proposed research with educational activities aimed at transforming a diverse cohort of students into skilled computational thinkers. This project will support the training of students to understand the complexity of systems models from an optimization context to better understand the practicality of optimization-based approaches. This project will deliver methods, tools, and training modules to serve the immediate and future technology workforce training needs of engineering fields that will increasingly depend on optimization for innovation.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.

复杂的系统无处不在：从农业供应链、废水处理和公共供水系统，到住宅和工业建筑的供暖、制冷和照明的能源/电力基础设施。加工工业脱碳，特别是与食品、能源和水相关的工业，具有特别和及时的重要性。工程师始终肩负着提高这些支持和改善社会的复杂系统的健康、安全和稳健性的使命。然而，创新本质上会增加复杂性，因此，解决复杂且相互关联的挑战（包括设计新系统和改进现有系统）的努力在很大程度上依赖于计算建模、模拟和基于优化的方法。该提案旨在解决两个主要挑战：（1）计算优化方法的当前性能限制了它们对简化的、复杂性较低的问题的适用性，以及（2）大学工程项目在整个课程中通常缺乏有凝聚力的计算思维活动。解决（1）将缓解当前的计算瓶颈，并扩大可以解决的复杂问题的规模和范围。解决（2）不仅有助于培训下一代工程师计算建模方法，还可以提高他们解决问题的整体技能。该项目的研究目标是开发可扩展的确定性全局优化（DGO）算法和开源软件实现通过利用替代流计算架构进行并行化，能够解决更高复杂度的模型，包括涉及非线性（偏）微分和代数方程的第一原理模型和机器学习元素。这项工作的意义在于通过开发一种新的分支定界确定性搜索算法来释放 DGO 图形处理单元 (GPU) 的大规模并行计算性能。其结果将是比当前技术水平显着加速，这将能够解决食品-能源-水（FEW）应用等中出现的更大规模、更复杂的问题。第一个主要创新是一种在任意感兴趣域上自动生成优化公式中非凸函数的凸/凹松弛的源代码表示及其次梯度的方法。第二个重大创新是可扩展的 GPU 兼容的并行 DGO 算法和开源软件实现，以保证非凸程序的解决方案。该项目将把拟议的研究与教育活动结合起来，旨在将多元化的学生转变为熟练的计算思想家。该项目将支持培训学生从优化环境中理解系统模型的复杂性，从而更好地理解基于优化的方法的实用性。该项目将提供方法、工具和培训模块，以满足工程领域当前和未来的技术劳动力培训需求，这些需求将越来越依赖于创新优化。该奖项反映了 NSF 的法定使命，并通过使用基金会的智力价值和更广泛的影响审查标准。