Collaborative Research: DESC: Type I: FLEX: Building Future-proof Learning-Enabled Cyber-Physical Systems with Cross-Layer Extensible and Adaptive Design

当电子设备不符合系统要求时，它可能成为电气和电子设备(WEEE)的废物。全球每年产生约5000万公吨的废电子电气产品，预计到2050年将增加到每年1.11亿吨，造成巨大的环境、人类健康和社会经济破坏。改装和重复使用这些设备可能会延长它们的寿命，减少这种负面影响。例如，以智能手机为例，仅将其使用时间延长一年，就可以将其对环境的二氧化碳影响减少31%。该项目增强了新出现的学习型网络物理系统(LE-CPSS)的改装能力，以适应现有硬件的变化/更新，从而延长设备寿命、减少电子垃圾和提高可持续性。该项目的成功可能会促进一个更可持续的未来，其深远的影响将跨越环境、经济和社会层面。通过加强改装能力，可以有效地延长电子设备的使用寿命，从而减少对新产品的需求，并减少产生的电子废物的累积数量。改进的翻新还可以带来更节能的系统，减少温室气体排放和减缓气候变化。此外，改造电子系统可以为企业和消费者节省成本，增加获得负担得起的技术的机会，特别是对经济困难的社区。该项目促进了几个领域的研究：设计自动化、计算机物理系统、机器学习和多代理系统应用领域的专家。该项目还通过课程开发、扩大对计算、K-12推广活动和国际设计竞赛的参与来产生更广泛的影响。改装可能非常具有挑战性，特别是对于新兴的LE-CPSS，如自动车辆、医疗设备和机器人，它们通常在不断变化的物理环境中运行，具有有限的资源和严格的时间要求，并使用复杂且耗费资源的机器学习技术。为了在改造过程中弥合功能和体系结构之间的差距并延长系统生命周期，该项目开发了跨层框架Flex，一方面使LE-CPSS的体系结构更具可扩展性，以便于适应现有硬件的变化，另一方面使其功能更适应资源限制和环境变化。该框架包括新颖的(1)系统级可扩展性驱动的设计和改装方法，其在设计时探索设计空间并权衡LE-CPSS与其他系统目标的未来可扩展性，并且在改装时利用LE-CPSS中现有功能的健壮性以基于弱硬范例进一步扩展调度松弛并适应改装需求；(2)资源适应性驱动的神经体系结构模型和设计方法，其提供神经网络的多种设计(例如，具有多个存在)以使得能够在改装期间进行资源感知配置；以及(3)持续和设备上的学习方法，使LE-CPSS能够有效地适应不断变化的物理环境和系统输入，而几乎没有监督来延长系统寿命。该奖项反映了NSF的法定使命，并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。

When an electronic device does not meet system requirements, it could become a waste of electrical and electronic equipment (WEEE). About 50 million metric tonnes of WEEE are generated annually worldwide, and it is estimated to increase to 111 million tonnes per year by 2050, causing immense environmental, human health, and socioeconomic damage. Retrofitting and reusing these devices may prolong their lifespan and reduce such negative impacts. For instance, in the case of a smartphone, extending its usage by just one year can cut its CO2 impact on the environment by 31%. This project enhances the retrofitting capability for emerging learning-enabled cyber-physical systems (LE-CPSs) to accommodate changes/updates with existing hardware, so as to increase device lifespan, reduce e-waste and improve sustainability. The success of this project could foster a more sustainable future, with far-reaching impacts spanning across environmental, economic, and societal dimensions. By enhancing retrofitting capabilities, the lifespan of electronic devices can be effectively extended, consequently mitigating the necessity for new products and reducing the cumulative quantity of electronic waste generated. Improved retrofitting can also lead to more energy-efficient systems, reducing greenhouse gas emissions and mitigating climate change. Moreover, retrofitting electronic systems can lead to cost savings for both businesses and consumers, increasing access to affordable technology, especially for economically disadvantaged communities. The project catalyzes research in several communities: design automation, cyber-physical systems, machine learning, and domain experts in multi-agent system applications. This project also generates broader impacts through curriculum development, broadening participation in computing, K-12 outreach activities, and international design contests.Retrofitting could be quite challenging, especially for emerging LE-CPSs such as autonomous vehicles, medical devices, and robots, which often operate within a constantly-changing physical environment, have limited resources and stringent timing requirements, and employ complex and resource-consuming machine learning techniques. To bridge the gap between functionality and architecture during retrofitting and prolong system lifetime, this project develops FLEX, a cross-layer framework that on the one hand makes the architecture of LE-CPSs more extensible to facilitate accommodating changes with existing hardware, and on the other hand makes their functionality more adaptive with respect to resource limitations and environment changes. The framework includes novel (1) system-level extensibility-driven design and retrofitting methods that at the design time explore the design space and trades off future extensibility of LE-CPSs with other system objectives and at the retrofitting time leverage the robustness of existing functionality in LE-CPSs to further expand scheduling slack and accommodate retrofitting needs based on a weakly-hard paradigm; (2) resource adaptability-driven neural architecture model and design methods that provide multiple designs of neural networks (e.g., with multiple exists) to enable resource-aware configuration during retrofitting; and (3) continual and on-device learning methods that enable LE-CPSs to effectively adapt to the changing physical environment and system input with little supervision for increasing system lifetime.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.