Collaborative Research: Enabling Hybrid Methods in the NIMBLE Hierarchical Statistical Modeling Platform

协作研究：在 NIMBLE 分层统计建模平台中启用混合方法

• 批准号: 2332442
• 负责人: Daniel Turek
• 金额: $ 6.99万
• 依托单位: Lafayette College
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2023
• 资助国家: 美国
• 起止时间: 2023-05-15 至 2025-06-30
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2332442&HistoricalAwards=false
• 关键词: Collaborative; Research; Enabling; Hybrid; Methods

This project will enable researchers in many fields of science to harness advanced computer algorithms to analyze complex data sets. In many fields, researchers seek to determine what hypotheses are supported by data collected in complex study designs. Data may be complex because they are collected in many locations, at many points in time, from related sampling units, under different sampling conditions, with different sample sizes, and/or with imperfect measurements. Such complexities arise in research fields such as biology, astronomy, education, environmental science, political science, and psychology, among others. When analyzing complex data, it can be difficult for researchers to determine which potential patterns are real and which are spurious. To solve this problem, researchers utilize computer algorithms to thoroughly explore all possible underlying relationships among variables that might explain the observed data. Such algorithms can be slow, costly, and difficult to create, so it is important to make them faster and easier for researchers to use. The investigators of this project have previously created a software package called NIMBLE (Numerical Inference for statistical Models using Bayesian and Likelihood Estimation) for this purpose. NIMBLE has been successfully used for many complex data analysis problems. Compared to other relevant software, NIMBLE enables researchers to use a wider range of algorithms and to customize algorithms to each research problem. This has allowed much faster performance in some cases, which in turn allows more comprehensive analysis of complex data. In the current project, the investigators will extend NIMBLE’s capabilities. They will make it possible to use some kind of accurate mathematical approximations for statistical calculations in combination with existing algorithms, which in turn will allow researchers to create new kinds of hybrid algorithms for data analysis. They will also make it possible to use certain kinds of very efficient calculations in some problems, which will greatly improve performance. The investigators will also provide support and training to users of the software as well as creating educational modules to help the next generation of undergraduate and graduate students learn to use these methods.NIMBLE is unique among hierarchical statistical modeling software because it combines a language for statistical models, a language for model-generic algorithms, and a compiler to generate and use C++ source code for models and algorithms. In the current project, NIMBLE will be extended to support hybrid methods by enabling algorithms to be nested within models. This will allow methods such as sparse grid quadrature to integrate over one set of model dimensions to achieve the calculations needed by another algorithm such as Markov chain Monte Carlo. In turn, this capability will allow composition of methods such as Laplace approximation and methods that use it. This project will also extend NIMBLE’s algorithm language to support sparse matrix algebra methods, allowing this efficient approach to be used by algorithm developers to enhance computational efficiency. Together, the advances in this project will enhance statistical research by enabling NIMBLE to serve as a hub for composition of models and methods, whereby a data analyst can create one statistical model and use many different methods with it. Finally, this project will include training and support for new and existing users.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.

该项目将使许多科学领域的研究人员能够利用先进的计算机算法来分析复杂的数据集。在许多领域，研究人员试图确定在复杂的研究设计中收集的数据支持哪些假设。数据可能很复杂，因为这些数据是在许多地点、许多时间点、从相关的采样单位、在不同的采样条件下、不同的样本量和/或不完美的测量中收集的。这种复杂性出现在生物学、天文学、教育、环境科学、政治学和心理学等研究领域。在分析复杂数据时，研究人员可能很难确定哪些潜在模式是真实的，哪些是虚假的。为了解决这个问题，研究人员利用计算机算法彻底探索变量之间所有可能的潜在关系，这些变量可能解释了观察到的数据。这样的算法可能速度慢、成本高，而且很难创建，因此让研究人员更快、更容易地使用它们是很重要的。该项目的研究人员此前为此目的开发了一个名为Nimble的软件包(使用贝叶斯和似然估计对统计模型进行数值推断)。Nimble已经成功地用于许多复杂的数据分析问题。与其他相关软件相比，Nimble使研究人员能够使用更广泛的算法，并为每个研究问题定制算法。这在某些情况下实现了更快的性能，进而允许对复杂数据进行更全面的分析。在目前的项目中，调查人员将扩展Nimble的能力。它们将使使用某种精确的数学近似与现有算法相结合进行统计计算成为可能，这反过来又将允许研究人员创建用于数据分析的新型混合算法。它们还将使在某些问题中使用某些非常高效的计算成为可能，这将大大提高性能。研究人员还将为该软件的用户提供支持和培训，并创建教育模块，帮助下一代本科生和研究生学习使用这些方法。NIMBLE在分层统计建模软件中是独一无二的，因为它结合了统计模型语言、模型通用算法语言以及编译器来生成和使用模型和算法的C++源代码。在当前的项目中，通过支持将算法嵌套在模型中，Nimble将被扩展为支持混合方法。这将允许稀疏网格求积等方法在一组模型维度上进行积分，以实现另一种算法(如马尔科夫链蒙特卡罗)所需的计算。反过来，这种能力将允许组合拉普拉斯近似等方法和使用它的方法。该项目还将扩展Nimble的算法语言以支持稀疏矩阵代数方法，使算法开发人员可以使用这种有效的方法来提高计算效率。总而言之，该项目的进展将加强统计研究，使Nimble能够成为组成模型和方法的中心，从而数据分析师可以创建一个统计模型并使用多种不同的方法。最后，该项目将包括对新用户和现有用户的培训和支持。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。