CI-ADDO-NEW: Stan, Scalable Software for Bayesian Modeling

CI-ADDO-NEW：Stan，用于贝叶斯建模的可扩展软件

• 批准号: 1205516
• 负责人: Andrew Gelman
• 金额: $ 49.96万
• 依托单位: Columbia University
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2012
• 资助国家: 美国
• 起止时间: 2012-06-01 至 2015-05-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1205516&HistoricalAwards=false
• 关键词: CI; ADDO; NEW; Stan; Scalable

This award is to design, code, document, test, dissememinate, and maintain Stan,an extensibleopen-source software framework and compiler for efficient and scalable Bayesian statistical modeling.Stan is an extensible, open-source, cross-platform software framework for developing Bayesian statisticalmodels. The first step in Bayesian modeling is setting up a full probability model for all quantities ofinterest. Stan facilitates this process by providing an expressive and extensible domain-specificprogramming language for specifying probabilistic models. By compiling a model specification intoexecutable code, Stan fully automates the second step of Bayesian inference, calculating the probabilitiesof unobserved quantities, such as model parameters and future observations, conditional on observed data.The third step involves evaluating the fit of the model to the data and its predictions for unseen data.When the model is easy to encode and inferences are fast and automatic to compute, it is easy to iteratethe specification, fit and evaluation steps in order to refine the scientific model.Stan improves on the existing state of the art in both algorithmic and implementation details. Rather thanbeing interpreted on the fly like its predecessors, Stan models are compiled to C++ code, whichdramatically improves both scalability and efficiency. Stan provides a full algorithmic differentiation library for the functions required for statistical modeling. This method applies the chain rule from calculus to the program computing the probability function in order to calculate derivatives efficiently and accurately (a small multiple of the time taken to compute thefunction, independently of dimensionality). This allows Stan to fully automate the model fitting stagegiven only a specification of the probability function in Stan's modeling language.To maximize Stan's accessibility to the scientific community, it is being coded using standards-compliantC++, so that it will run under Windows, Macintosh, and Unix/Linux. To make running Stan even easier,it is callable from R, MATLAB, and Python, the three most popular platforms for numerical analysis,including exploration and plotting.

该奖项旨在设计，编码，文档，测试，传播和维护Stan，这是一个可扩展的开源软件框架和编译器，用于高效和可扩展的贝叶斯统计建模。Stan是一个可扩展的，开源的，跨平台的软件框架，用于开发贝叶斯统计模型。贝叶斯建模的第一步是为所有的兴趣量建立一个全概率模型。Stan通过提供一种用于指定概率模型的表达性和可扩展的特定于领域的编程语言来促进这一过程。通过将模型规范编译成可执行代码，Stan完全自动化了贝叶斯推理的第二步，即计算未观测到的数量的概率，例如模型参数和未来观测，以观测数据为条件。第三步涉及评估模型对数据的拟合及其对未观测数据的预测。当模型易于编码且推理快速且自动计算时，它可以很容易地迭代规范，拟合和评估步骤，以细化科学模型。Stan在算法和实现细节方面都改进了现有的技术水平。Stan模型不像它的前辈那样被动态地解释，而是被编译成C++代码，这极大地提高了可伸缩性和效率。Stan为统计建模所需的函数提供了一个完整的算法微分库。这种方法将微积分中的链式法则应用到计算概率函数的程序中，以便有效而准确地计算导数（计算函数所需时间的小倍数，与维数无关）。这使得Stan可以完全自动化模型拟合阶段，只需在Stan的建模语言中指定概率函数。为了最大限度地提高Stan对科学界的可访问性，它使用符合标准的C++进行编码，以便它可以在Windows，Macintosh和Unix/Linux下运行。为了使运行Stan更容易，它可以从R，MATLAB和Python调用，这三个最流行的数值分析平台，包括探索和绘图。