Collaborative Research: Inference on Expensive, Grey-Box Simulation Models

该拨款将通过允许组织使用随机计算机模拟模型来面对不确定性做出关键的战略、战术和操作决策，从而促进国家的繁荣和福利。供应链物流、运输、医疗保健、国防规划和金融领域的此类决策需要根据对关键绩效指标的估计，从数百万种场景中进行选择。该奖项支持灵活框架的基础研究，利用模拟系统性能的信息，为迄今为止无法解决的大规模问题获得更强的推理和更好的决策。美国领先公司的学术研究人员和工业从业者之间的密切合作将产生快速、可靠的算法，可以针对许多行业和政府机构的各种模拟问题进行定制。所有结果将作为开源软件提供，并以研究生工程、数据科学和商业课程的教程和教学材料的形式传达方法。这项研究的动机是现代大规模模拟问题和最先进的通用方法的缺点，这些方法将模拟模型视为“黑匣子”。这项工作将创造一种方法，通过从现代模拟的“灰盒”性质中提取和利用额外的结构信息，有效地进行统计推断。具体来说，它将利用这些信息来加强交付的推论，并为昂贵的随机模拟模型提供可证明的保证。该方法将考虑多个相互冲突的性能指标，利用结构信息驱动模拟运行的顺序实验设计，并从仿真模型中验证和提取所需的结构信息，以获得可扩展的计算方法。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

This grant will advance the national prosperity and welfare by allowing organizations to use stochastic computer simulation models to make critical strategic, tactical and operational decisions in the face of uncertainty. Such decisions in supply chain logistics, transportation, healthcare, defense planning and finance entail choosing from among millions of scenarios based on estimates of key performance indicators. The award supports fundamental research on a flexible framework for exploiting information about simulated system performance to obtain stronger inference and better decisions for hitherto unsolvable large-scale problems. Close collaboration between academic researchers and industrial practitioners at leading U.S. companies will yield rapid, reliable algorithms that can be tailored to a diverse array of simulation problems across many industries and government agencies. All results will be made available as open-source software and the methods communicated in the form of tutorials and instructional materials for graduate engineering, data science and business classes.The research is motivated by modern large-scale simulation problems and the shortcomings of state-of-the-art general-purpose methods that treat a simulation model as a “black box.” The work will create methods that efficiently carry out statistical inference by extracting and exploiting additional structural information available from the “grey-box” nature of modern simulations. Specifically, it will leverage this information to strengthen the delivered inferences and offer certifiable guarantees for expensive stochastic simulation models. The methods will account for multiple, conflicting performance measures and exploit structural information to drive the sequential experiment design of simulation runs, as well as verify and extract the needed structural information from a simulation model to obtain scalable computational methods.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.