Seaholm S K, Ackerman E, Wu S C
Division of Health Computer Sciences, University of Minnesota, Minneapolis 55455.
Int J Biomed Comput. 1988 Oct;23(1-2):97-112. doi: 10.1016/0020-7101(88)90067-0.
Discrete, algorithmic simulation and Monte Carlo methodologies are currently used in population biology, connectionist cognitive modeling, and physics. However, little is typically known about the sensitivity of such models to changes in the values of the model features. Traditional methods of sensitivity analysis for systems of differential equations do not apply. Sometimes, one or two parameters are modified at a time in an ad hoc fashion in an attempt to assess sensitivity. To include more model features and their interactions in a sensitivity study, while limiting computer utilization, various sampling methods have been suggested. In this article, a sensitivity study based on a Latin hypercube (LH) sampling design is compared with a similar study using a full factorial (FF), fixed-point sample. A discrete, Monte Carlo model of epidemics of influenzavirus infections in a human community is used for illustrative purposes. Although the FF scheme used over 14 times as many samples as the LH sampling one, both provided comparable predictive ability and comparable information about simulation sensitivity to model features.
离散算法模拟和蒙特卡罗方法目前应用于群体生物学、联结主义认知建模和物理学领域。然而，此类模型对于模型特征值变化的敏感性通常鲜为人知。传统的微分方程系统敏感性分析方法并不适用。有时，人们会临时一次修改一两个参数，试图评估敏感性。为了在敏感性研究中纳入更多模型特征及其相互作用，同时限制计算机的使用，已有人提出了各种抽样方法。在本文中，将基于拉丁超立方（LH）抽样设计的敏感性研究与使用全因子（FF）定点样本的类似研究进行了比较。为了说明问题，使用了一个人类社区中流感病毒感染流行的离散蒙特卡罗模型。尽管FF方案使用的样本数量是LH抽样方案的14倍多，但两者都提供了相当的预测能力以及关于模拟对模型特征敏感性的相当信息。
