Cross-Leveraging Computational, System and Data Science in Support of Computational Epidemiology in the Era of Big Data

尽管对ABMs与大数据集成的需求快速增长，并且我们的贡献显示了巨大的潜力，但现有的仿真基础设施对模型与传入数据的集成提供了较差的支持。虽然我们的工作已经证明，在线计算统计技术，如粒子滤波（PF）和粒子马尔可夫链蒙特卡罗（PMCMC）可以形成高效的工具，用于集成仿真建模和传入的大数据——比如来自我们流行的iEpi系统的数据——但这些技术的应用充其量是尴尬的，而且由于高计算成本或涉及的高度实现努力，通常是不可行的。为了解决这个跨学科团队合作至关重要的应用领域，我们已经通过现有的Frabjous领域特定功能反应性ABM编程平台取得了巨大的成功，从而显著提高了ABM的透明度、简明性和模块化。尽管有这些贡献，当前的ABMs通常缺乏可发布的规范，对非技术涉众来说通常是相当不透明的，甚至对技术团队成员来说也常常是困惑的，对政策制定者和分析人员基于场景的探索提出了重大的性能障碍，并且对团队之间的协作交互和共享的支持很差。在这里，我们建议采用多管齐下的策略来解决这些挑战，首先是将Frabjous移植到Scala语言（Frabjous）——一种对模块化、并行化、特定于领域的语言设计和互操作性提供强大支持的语言，我们已经在许多其他工具中使用了这种语言。这个端口之后的工作被分成四个相对独立的流。第一个重点是集成对关键计算统计算法PF、PMCMC和MCMC的语言支持，以及使用流行的Spark数据科学平台的流组件的流接口。第二个目标是通过多级并行（在更细粒度的级别上利用分布式计算、多核和gpu），包括通过与Spark平台的集成，极大地提高ABM的性能。第三个接口使用一元组合来简化常见的建模任务，并授权模型最终用户承担传统上需要程序员支持的分析任务。最后，在与人机交互（HCI）和计算机支持的协同工作（CSCW）领域的领先研究人员的合作中，我们将适应在我们现有的协作模型映射工具中成功使用的技术，以实现FrabjouS的图形规范语言，以及支持多用户探索，运行和修改FrabjouS模型的协作工具。最后，在工作的每个阶段，我们将通过用户研究来评估模型的成功。

Despite the rapid growing demand for -- and our contributions demonstrating the great potential of -- integration of ABMs with big data, existing simulation infrastructures provide poor support for model integration with incoming data. While our work has demonstrated that online computational statistics techniques such as Particle Filtering (PF) and Particle Markov Chain Monte Carlo (PMCMC) can form highly effective tools for integrating simulation modeling and incoming big data -- such as that from our popular iEpi system -- application of such techniques is at best awkward, and is often infeasible because of the high computational costs or high degree of implementation effort involved. To address this application area in which interdisciplinary teamwork is of central importance, we have secured strong success with our existing Frabjous domain-specific functional reactive ABM programming platform to significantly enhance ABM transparency, concision, and modularity. Despite these contributions, current ABMs commonly lack publishable specifications, are typically quite opaque to non-technical stakeholders and often confusing even to technical team members, raise significant performance barriers to scenario-based exploration by policy makers and analysts, and poor support for collaborative interaction and sharing across teams. We propose here to address these challenges using a multi-pronged strategy that starts with a port of Frabjous to the Scala language (FrabjouS) -- a language which offers strong support for modularity, parallelization, domain-specific language design and interoperability, and which we have used for many other tools. Work following this port is divided into four relatively autonomous streams. The first focuses on integrating language support for key computational statistics algorithms PF, PMCMC, and MCMC and for streaming interfaces using the streaming component of the popular Spark data science platform. The second seeks to greatly enhance ABM performance via multi-level parallelization (exploiting both distributing computing, multi-cores and GPUs at a finer-grained level), including via integration with the Spark platform. The third interface uses monadic composition to both ease common modeling tasks, and to empower model end-users to undertake analysis tasks traditionally requiring programmer support. Finally, in collaborations with a leading researcher in this the area of Human Computer Interaction (HCI) and Computer Supported Cooperative Work (CSCW), we will adapt techniques successfully used in our existing collaborative model mapping tools to implement a graphical specification language for FrabjouS as well as a collaborative tool supporting multi-user exploration, running and modification of FrabjouS models. Finally, across each phase of the work, we will evaluate model success with user studies.