EAGER-Dynamic Data: A New Scalable Paradigm for Optimal Resource Allocation in Dynamic Data Systems via Multi-Scale and Multi-Fidelity Simulation and Optimization

在工程和自然系统中，无处不在的传感和控制的基本目标是理解、分析和优化这些系统的运行条件。尽管经典的反馈控制理论为满足基于评估系统状态的运行目标和约束奠定了坚实的基础，但传统的控制范式对现代动态大数据和复杂系统的适用性有限。最根本的挑战是处理、融合和计算来自多个异构和分布式数据源的数据的效率，以达到及时和最佳的决策。该项目将开发创新方法，以实现无缝和有效地集成多个实体、多模式和多保真度数据收集活动的前所未有的动态交互，以及不同级别和规模的系统运行条件的计算。技术的发展将使系统中多个实体以高效和稳健的方式评估一个非常大的决策空间，并扩展到无处不在的传感带来的大数据环境。该研究成果有可能显著推动动态数据系统、仿真和优化研究的最新进展，有可能为改进各种应用系统的执行开辟一条新的途径。这项研究虽然具有通用性，也适用于其他动态数据工程系统，但它特别受到半导体行业大数据问题的推动。半导体行业是美国乃至世界经济的一个关键部门。研究成果将通过技术出版物和演讲以及课堂教学传播，在GMU将开发一门新的多学科课程，并向广泛的学生提供。本探索性研究的目的是为动态数据系统中的实时优化资源分配开发一种新的可扩展计算范式。这项研究的变革之处在于认识到动态数据系统的成功执行依赖于实时的全球态势感知和将感知及时转化为（接近）最佳资源分配决策的能力。基础技术突破是一种新的多尺度、多保真度仿真与优化框架，该框架以自适应、高效和鲁棒的方式集成了多尺度、多保真度的数据收集和决策。多尺度模拟和优化允许使用局部数据识别有前途的局部尺度资源分配决策。然后通过全局尺度的多保真度模拟和优化来评估局部尺度的资源分配决策，以寻找最优的全系统决策。这种集成的多尺度、多保真度模式利用了局部尺度的响应能力和全球尺度的全局态势感知能力，从而有可能在实时决策过程中实现效率和鲁棒性。通过对多模态、多保真度数据的高效、可扩展融合，系统中的分布式实体监控动态数据系统的运行状态，并根据系统运行状态的扰动自主控制仪器仪表和数据采集过程。有了信息的价值，决策模型可以对数据收集、计算和资源分配决策进行联合评估，从而动态地调度和优先处理对成功执行动态数据系统最有贡献的任务。

The fundamental objective of ubiquitous sensing and control in engineered and natural systems is to understand, analyze, and optimize operational conditions of these systems. Although the classical feedback control theories lay a solid foundation to enable meeting operating goals and constraints based on the assessed system states, the traditional control paradigm has limited applicability to the modern dynamic big data and complex systems. The fundamental challenge is the efficiency of processing, fusing, and computing of data from multiple heterogeneous and distributed sources to arrive at a timely and optimal decision. This project will develop innovative approaches to enable seamlessly and efficiently integrating unprecedented dynamic interactions of multiple entities multimodal and multi-fidelity data collection activities, and the computing of systems operational conditions at different levels and scales. The technological developments will enable the evaluation of a very large decision space for multiple entities in the system in an efficient and robust manner, and scale up to the big data environment brought forth by ubiquitous sensing. The research outcome has a potential to significantly advance the state of the art in dynamic data system, simulation, and optimization research, potentially opening a new avenue to improve the execution of a large variety of application systems. The research, while generic and applicable to other dynamic data engineered systems, is specifically motivated by the big data problem in semiconductor industry, a crucial sector of the US and world economy. Research findings will be disseminated through technical publications and presentations as well as classroom teaching where a new multi-disciplinary course will be developed at GMU and offered to a wide-range of students. The objective of this exploratory research is to develop a new scalable computational paradigm for real-time optimal resource allocation in dynamic data systems. The transformative aspect of this research is the recognition that the successful execution of a dynamic data system relies on real-time global situational awareness and the capability to translate awareness into (near) optimal resource allocation decisions in a timely manner. The fundamental technical breakthrough is a new multi-scale and multi-fidelity simulation and optimization framework that integrates data collection and decision making at multiple scales and multiple fidelity levels in an adaptive, efficient, and robust manner. Multi-scale simulation and optimization allows identifying promising local scale resource allocation decision using localized data. Local scale resource allocation decisions are then evaluated by global scale multi-fidelity simulation and optimization in search of the optimal system-wide decision. Such an integrated multi-scale and multi-fidelity paradigm exploits the responsiveness at local scale and the global situational awareness at the global scale, and thus has the potential to attain both efficiency and robustness in the real-time decision making process. Through efficient and scalable fusion of multi-modal and multi-fidelity data, distributed entities in the system monitor the operating conditions of the dynamic data system and autonomously control the instrumentation and data collection process in response to perturbations in system operating conditions. With value of information, the decision model enables joint evaluation of data collection, computing, and resource allocation decisions to dynamically schedule and prioritize tasks that would contribute most to the successful execution of dynamic data systems.