CAREER: Orchestrating Edge Infrastructures and Mobile Devices under Uncertainty to Provision Edge AI as a Service

人工智能(AI)通常首先由数据训练，然后用来对可能从未见过的数据进行推理。如今的人工智能服务大多在远程数据中心接受培训，并通过互联网访问。然而，随着最终用户产生大量的训练数据，推理任务也在驻地执行，人工智能需要移到离用户更近的网络边缘。该项目的目标是通过协调分布式计算和网络基础设施并将人工智能作为服务从网络边缘提供给不同位置的大量用户来应对这种新兴的范式转变。该项目将定义优化问题，设计控制算法，并开发跨云边缘网络和移动设备将人工智能作为服务运营的可部署实现。该项目以系统动力学和不确定性为中心，由多个研究推动力组成。首先，该项目将通过在参与者选择、模型放置、聚合控制和请求调度方面做出明智决策，优化人工智能的训练和推理，并将产生在线运行的基于时间的算法，并提供可证明的性能保证。其次，该项目将从经济学的角度调查人工智能服务生态系统中不同各方之间的互动，并将设计激励机制，以实现更好的资源利用和最优的社会福利，并具有所需的经济特性。第三，该项目将结合真实世界的数据跟踪和现实的实验设置进行模拟、测量和实施，以全面评估和验证模型、算法和系统。该项目将通过为电信运营商、网络运营商和服务提供商提供一套关键技术，使他们能够在分布式基础设施上提供和运营专用或增值边缘AI服务，从而影响行业。此外，该项目将基于优化、控制、学习和机制设计的数学设计新的理论和算法，并可能具有独立的兴趣，并将应用扩展到其他相关领域的问题，这些领域也面临动态和不确定的输入。最后，除了提供本科和研究生课程材料外，该项目还将执行教育计划，重点是通过培训教师和以道德的方式向学生提供人工智能知识和经验，促进K-12教育。该项目的交付成果，包括但不限于论文、数据和代码，将在以下网站公开提供：https://github.com/ai-at-edge.本网站将在本项目期间及时更新，并在此后在线维护。该奖项反映了NSF的法定使命，并已通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Artificial Intelligence (AI) is often trained by data first, and then used to perform inference upon data which may have never been seen. Today's AI services are mostly trained in remote data centers and accessed via the Internet. However, increasingly, as large volumes of training data are generated by end users and inference tasks are also performed on premises, AI needs to be moved to the network edge in closer proximity to users. The objective of this project is to address this emerging paradigm shift by orchestrating the distributed computing and networking infrastructures and provisioning AI as a service from the network edge to serve large numbers of users at different locations.This project will define the optimization problems, design the control algorithms, and develop the deployable implementations for operating AI as a service across cloud-edge networks and mobile devices. Centered on system dynamics and uncertainty, the project consists of multiple research thrusts. First, the project will optimize the training and the inference of AI through making smart decisions on participant selection, model placement, aggregation control, and request dispatching, and will produce time-based algorithms running in an online manner with provable performance guarantees. Second, the project will investigate the interactions among different parties in the AI service ecosystem from an economics perspective, and will devise incentive mechanisms towards superior resource utilization and optimal social welfare with desired economic properties. Third, the project will conduct a combination of simulations, measurements, and implementations with real-world data traces and realistic experimental settings to comprehensively evaluate and validate the models, algorithms, and systems.This project will impact the industry by providing telecom carriers, network operators, and service providers with a critical set of techniques to enable them to provision and operate dedicated or value-added edge-AI services upon distributed infrastructures. Further, the project will devise novel theories and algorithms based on the mathematics of optimization, control, learning, and mechanism design, and could be of independent interest and have extended applications to problems in other related fields that also face dynamic and uncertain inputs. Finally, besides contributing to undergraduate and graduate course materials, the project will execute the education plan that focuses on facilitating K-12 education via training teachers and also equipping students with AI knowledge and experiences in an ethical manner.The deliverables of this project, which include but are not limited to papers, data, and codes, will be made publicly available at the following website: https://github.com/ai-at-edge. This website will be updated in time for the duration of this project and maintained online thereafter.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.