GOALI: Quantifying Input Uncertainty in Stochastic Simulation

该奖项的目的是建立一个框架，用于量化输入不确定性，该框架是严格合理的，但也是有用的。 随机模拟最有价值的方面之一是其表征风险的能力，其中“风险”是系统行为的可变性。 不幸的是，有一个隐藏的，往往是大量的错误，在模拟中存在，即使采用最佳实践建模，实验设计和输出分析：输入不确定性误差。这项研究提供了严格合理的方法，考虑到输入的不确定性，也可以在仿真软件中实现。该研究通过量化误差的置信区间和量化风险的预测区间来显示输入不确定性的影响。计划的方法是有效地和准确地传播输入的不确定性从输入模型的模拟输出使用自举响应和模拟元建模的最新进展。 其结果将是一个量化输入不确定性的框架，这是严格合理的，但也是有用的。Simio LLC的GOALI合作伙伴将增加利用云计算的模拟专业知识，并为本研究所研究的方法提供软件测试平台。输入模型是模拟实验的驱动过程。它们代表的不确定性在一定程度上抵制更详细的建模。服务和制造模拟中的到达流程;供应链模拟中的需求流程;医院模拟中的患者占用时间都是输入的示例。输入模型基于观察到的真实世界数据，因此它们容易出错。这种错误可能会压倒其他来源的模拟错误，使用户面临做出关键和昂贵的决策的风险，对（似乎是）高度精确的模拟评估没有根据的信心。这项研究将通过量化这一风险为决策者提供信息。由于输入不确定性即使在学术界也不为人所知或理解，该项目计划通过冬季模拟会议和其他地方的介绍性、高级和供应商教程“教授教师”。

The objective of this Grant Opportunity for Academic Liaison with Industry (GOALI) award is to develop a framework for quantifying input uncertainty that is rigorously justified, but also useful. One of the most valuable aspects of stochastic simulation is its ability to characterize risk, where "risk" is the variability in a system's behavior. Unfortunately, there is a hidden and often substantial error in simulation that is present even if best practices for modeling, experimental design and output analysis are employed: input-uncertainty error. This research provides rigorously justified methods that account for input uncertainty that are also implementable within simulation software. The research displays the impact of input uncertainty via confidence intervals that quantify error and prediction intervals that quantify risk. The planned approach is to efficiently and accurately propagate input uncertainty from the input models to the simulation output using bootstrap resampling and recent advances in simulation metamodeling. The result will be a framework for quantifying input uncertainty that is rigorously justified, but also useful. The GOALI partner at Simio LLC will add expertise on simulation that exploits cloud computing, as well as providing a software test bed for the methods investigated by this research.Input models are the driving processes in simulation experiments. They represent uncertainty at a level that resists more detailed modeling. Arrival processes in service and manufacturing simulations; demand processes in supply chain simulations; and patient occupancy times in hospital simulations are examples of inputs. Input models are based on observed real-world data, so they are subject to error. This error can overwhelm other sources of simulation error, placing users at risk of making critical and expensive decisions with unfounded confidence in (what appear to be) highly precise simulation assessments. This research will inform decision makers by quantifying this risk. Since input uncertainty is not well known or understood even in the academic community, the project plans ``teaching the teachers'' via introductory, advanced and vendor tutorials at the Winter Simulation Conference and elsewhere.