CAREER: Uncertainty-aware sensing and management for IoT

在大量无处不在的智能设备的支持下，物联网（IoT）的出现通过大量应用为我们的日常生活带来了极大的便利，其中许多应用对安全至关重要，包括医疗保健、监控和自动驾驶等。对于这样的安全关键领域，目前的工具包通常在不确定性量化方面存在不足，而不确定性量化是知情决策所必需的关键特征。考虑到物联网设备在运行中收集的大量数据，可扩展性是实现低延迟实时物联网传感和管理的另一个关键因素。此外，如何赋予学习对物联网中不可预测动态的适应性和鲁棒性是至关重要的，特别是在人在环的情况下。在充分发挥安全关键型物联网的全部潜力之前，必须开发新的工具来应对这些重大挑战。为了实现这一目标，本CAREER提案倡导基础研究，旨在推进当前实时物联网传感和管理工具，直接影响许多安全关键领域，包括医疗保健、交通和环境传感。利用PI的机构资源，PI将通过i)指导研究生和本科生，特别是那些来自代表性不足群体的学生，将拟议的研究目标转化为教育活动；Ii)使机器学习、通信、信号处理和网络等领域相互融合的课程开发；以及iii)与佐治亚大学信息物理系统中心的跨学科合作。这种研究和教育的无缝整合是中央的PI的职业道路，并与UGA的使命“教，服务，并探究到事物的本质。”为了进一步促进社会对新兴物联网技术的接受，PI致力于通过短期课程，在线视频和研讨会向公众，特别是K-12学生传播研究成果。本提案提出了一个雄心勃勃的计划，通过剪裁当代贝叶斯机器学习工具的进展，即贝叶斯函数近似，贝叶斯强盗优化和贝叶斯强化学习，来解决上述挑战。这种新鲜的贝叶斯风格自然地创新了现有的工具包，具有不确定性量化和鲁棒性，对安全关键型物联网至关重要。由此产生的方法不仅有利于关键的物联网任务，而且通过纳入物联网驱动的约束，显著推动了这些学科的发展。具体而言，将进行三个互补和相互交织的研究重点。Thrust 1 （T1）提出了一个基本的不确定性感知函数学习框架，该框架不仅直接有利于T1中面向预测的物联网感知任务，而且有助于Thrust 2 （T2）中开环盲物联网管理的贝叶斯优化，即物联网控制器的决策不影响物联网状态。Thrust 3进一步建立在T1和T2的基础上，通过物联网状态和物联网控制器之间的充分交互，扩展贝叶斯强化学习，实现实时闭环物联网管理。最终追求的是一个整体框架，该框架集成了新颖的算法，具有不确定性意识、可扩展性和实时物联网传感和管理的自适应能力，以及对不可预测动态的鲁棒性的相关严格分析，以及部署到真正的安全关键型物联网应用。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Bolstered by a massive scale of ubiquitously connected smart devices, the emergence of Internet-of-Things (IoT) has brought about substantial conveniences to our daily life through a plethora of applications, of which many are safety-critical, including healthcare, surveillance, and autonomous driving, to name a few. For such safety-critical domains, the current toolkits usually fall short in uncertainty quantification, a key feature that is necessitated for informed decision-making. Given the enormous data collected by IoT devices on-the-go, scalability is another key enabler of real-time IoT sensing and management with low latency. Further, how to endow learning with adaptivity and robustness to unpredictable dynamics in IoT is of utmost importance, especially with humans-in-the-loop. Before embracing the full potential of safety-critical IoT, novel tools have to be developed to address these major challenges. Towards this goal, this CAREER proposal advocates fundamental research that aspires to advance the current tools for real-time IoT sensing and management, with direct impact on a number of safety-critical domains, including healthcare, transportation, and environmental sensing. Leveraging the PI's institutional resources, the PI will transform the proposed research goals into educational activities, through i) mentoring graduate and undergraduate students, especially those from the underrepresented groups; ii) curriculum development that cross-fertilizes the fields of machine learning, communications, signal processing and networking; as well as iii) interdisciplinary collaboration with UGA's Center of Cyber-Physical Systems. This seamless integration of research and education is central to the PI's career path and is well aligned with UGA's mission ``to teach, to serve, and to inquire into the nature of things." To further promote the societal embracing of the emergent IoT technologies, the PI is committed to disseminate the research outcomes to the general public, in particular K-12 students, through short courses, online videos, and workshops.This proposal puts forth an ambitious plan by tailoring advances in contemporary Bayesian machine learning tools, namely, Bayesian function approximation, Bayesian bandit optimization, and Bayesian reinforcement learning, to address the aforementioned challenges. This fresh Bayesian flavor naturally innovates existing toolkits with uncertainty quantification and robustness, essential to safety-critical IoT. The resultant approaches will not only benefit key IoT-enabled tasks, but also markedly push the envelope of these disciplines by incorporating IoT-driven constraints. Specifically, three complementary and intertwined research thrusts will be pursued. Thrust 1 (T1) puts forth a fundamental uncertainty-aware function learning framework, which not only directly benefits the prediction-oriented IoT sensing task in T1, but also contributes to Bayesian optimization for open-loop blind IoT management in Thrust 2 (T2), where the decisions made by the IoT controller do not affect the IoT state. Thrust 3 further builds on T1 and T2 to scale up Bayesian RL for real-time closed-loop IoT management with full interaction between the IoT state and the IoT controller. The ultimate pursuit is a holistic framework that integrates novel algorithms with uncertainty awareness, scalability, and adaptivity for real-time IoT sensing and management, the associated rigorous analyses for robustness to unpredictable dynamics, and the deployment to real safety-critical IoT applications.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.