ITR ACTS - Adjoint Complier Technology & Standards

0205590 Wunsch自动程序编译（及其相关的解释程序执行任务）已经成为创建越来越复杂的软件工具的重要组成部分。事实上，现代编译器和解释器已经成为利用数字硬件能力指数级进步的核心。如果没有编译器和解释器来自动化正确转换算法的高级语言描述的任务，现代计算机系统几乎无法存在。借助现代计算能力，许多领域的工作人员使用复杂相互作用元素的大规模数值模拟来了解物理，化学，生物和经济现象。伴随计算器技术标准（Adjoint Optimizer Technology Standards）项目旨在为科学家、工程师、学者和学生提供技术、工具、参考项目和易于获取的专业知识，从而允许采用新方法来研究基于数学精确导数信息的真实世界问题的数值模型。考虑到大规模的应用，激励的自动微分（AD）的选择方法没有合理的替代品。AD将实现原始模型的程序转换为导数代码，方法是用计算偏导数的指令对其进行扩充。例如，研究某些值相对于某些参数的数值灵敏度成为可能。当观测值与模型输出相结合以考虑两者的信息内容时，就有一个非常大规模的优化问题。这个问题的解决方案也在于生成衍生代码的能力。由于应用程序的庞大规模，AD必须处理生成的派生代码的效率是至关重要的。使用AD生成的代码的效率低下通常意味着不可行的硬件要求导致的不适用性。该项目有可能影响从工程设计优化项目到基础研究再到复杂物理系统监测工作的各个领域。该提案汇集了计算机科学家，物理科学家和工程师，共同努力创建一个标准，参考实现的可扩展工具包AD的数字代码。这样一个工具包是一个关键的资产，在提高保真度和生产力的日益复杂的仿真代码开发工作在许多领域。总的来说，这个项目的目标有三个方面1。发展一个可扩展的、基于标准的基础设施，使AD领域的研究蓬勃发展，并使学术发展和新兴的商业努力以互利的方式合作; 2.开发一个开源的、免费的参考编译器和支持材料，它和传统编译器一样容易使用和可靠，并将促进AD作为各种数值模拟的关键元素的广泛采用; 3.在计算机科学中的主流编译器社区与历史上一直处于AD工具开发前沿的数学，物理科学和工程社区之间建立明确的形式联系。对于数学，科学和工程代码，所提出的工作可能会产生深远的影响。正如传统的编译器的发展提供了新一代的仿真工具的基础上，这里提出的项目也将提供新一代的数学上严格和强大的制造分析，诊断和预测软件的基础。

0205590 WunschAutomatic program compilation (and its related task of interpreted program execution) has been a vital ingredient in the creation of ever more complex software tools. Indeed, modern compilers and interpreters have been central to exploiting exponential advances in digital hardware capability. Without compilers and interpreters to automate the task of correctly transforming a high-level language description of an algorithm, modern computer systems could barely exist. With modern computational capabilities, workers in numerous fields use large-scale numerical simulation of complex interacting elements to understand physical, chemical, biological and economical phenomena. The Adjoint Compiler Technology & Standards (ACTS) project aims to provide scientists, engineers, academics, and students with techniques, tools, reference projects, and easily accessible expertise allowing for new approaches to investigating numerical models of real-world problems based on mathematically precise derivative information. Considering the large-scale applications motivating the ACTS project there is not sensible alternative to Automatic Differentiation (AD) as the methods of choice. AD transforms the program implementing the original model into derivative code by augmenting it with instructions for computing partial derivatives. For example, the investigation of numerical sensitivities of certain values with respect to certain parameters becomes possible. Where observations are to be combined with model output so as to account for the information content of both, one has a very large-scale optimization problem. The solution to this problem lies with the ability to generate derivative code as well. Because of the sheer size of the application programs that AD has to deal with efficiency of the generated derivative code is crucial. Inefficiency of the codes generated using AD will often mean inapplicability caused by infeasible hardware requirements. This project has potential to impact fields ranging from engineering design optimization projects to basic research to complex physical system monitoring endeavors. The proposal brings together computer scientists, physical scientists and engineers in a group effort to create a standard, reference implementation of an extensible toolkit for AD of numerical codes. Such a toolkit is a key asset in improving the fidelity and productivity of increasingly complex simulation code development efforts in many fields. In summary the goals of this project are three fold 1. development of an extensible, standards based, infrastructure that will allow research in the field of AD to flourish and that will allow academic developments and emerging commercial efforts to collaborate in a mutually advantageous manner; 2. development of an open-source, freely available reference compiler and support material that is as easy and reliable to use as a traditional compiler and that will nurture widespread adoption of AD as an key element of all manner of numerical simulation; 3. establishment of clear and form connections between the mainstream compiler communities in computer science and the mathematics, physical science and engineering communities that have historically been at the forefront of AD tool development. For mathematical, scientific and engineering codes, the work proposed can have far reaching consequences. Just as conventional compiler developments have provided the foundation for new generations of simulation tools the project proposed here will also provide a foundation for new generations of mathematically rigorous and robustly manufactured analysis, diagnostic and forecast software.