Learning-based Energy Management for Cyber-Physical Systems

随着工程技术和网络技术的进步，许多嵌入式设备现在都连接到网络或互联网上，并正在形成网络物理系统。网络物理系统可以在各种控制和监控应用中找到，例如汽车，航空航天，医疗保健，交通运输，建筑和过程控制以及娱乐。与桌面系统不同，许多网络物理系统使用的是有限能量供应的电池。因此，能源效率是网络物理系统的固有要求之一，因为低能耗产生更好的电池寿命，这对于涉及植入医疗设备的应用尤其重要。随着网络物理系统的日益增多，网络物理系统的低能耗问题日益突出。如何在满足实时系统要求的前提下，将CPU或设备的能耗降至最低，是近十年来备受关注的研究课题。由于系统配置和任务特征的多样性，具有动态电压/频率缩放（DVFS）和/或动态电源管理（DPM）的调度安排对于一种系统配置是节能的，但可能不适用于另一种系统配置。因此，设计能够适应各种系统配置和任务特征的调度算法是非常重要的。该项目的目标是为各种类型的网络物理系统在各种时间和资源约束下的调度协同设计问题开发自适应高效算法。该计划采用基于学习的方法，能够学习基于调度历史的底层网络物理系统的电压选择或调度策略选择的隐式模型。当任务特性和体系结构模型对于DVFS调度器来说是未知的（或者太复杂而无法考虑）时，这种方法特别有用。还将探讨提取良好特征作为基于学习的方法的隐式模型的输入的问题，该方法可以最好地代表底层网络物理系统的模型。目前将探索q学习、双q学习和深度双q学习。对该框架的评价也将进行广泛研究。所设计的框架将用于探索具有各种任务类型（依赖、独立、周期或非周期等）、各种调度策略（最早截止日期优先、固定优先级等）、各种系统配置（单核、多核、GPU、基于noc的、可穿戴设备类型等）的系统。

With the advance of engineering and networking technology, many embedded devices are now connected into a network or with the Internet, and are emerging into cyber-physical systems. A cyber-physical system can be found in various applications for control and monitoring, such as automotive, aerospace, health care, transportation, building and process control, and entertainment. Unlike desktop systems, many cyber-physical systems operate with batteries that have limited energy supplies. Thus, energy efficiency is one of the inherent requirements of cyber-physical systems, as low-energy consumption yields better battery life, which is especially important for applications involving implanted medical devices. With more and more cyber-physical systems around us nowadays, low-energy consumption problem of cyber-physical systems becomes more critical. The research challenge to minimize the energy consumption of CPU or devices, while still meeting the constraints of the real-time systems, has attracted much attention in the past decade. With the variety of system configurations and task characteristics, a scheduling arrangement with Dynamic Voltage/Frequency Scaling (DVFS) and/or Dynamic Power Management (DPM) that is energy efficient for one system configuration might not be appropriate for another. Therefore, it is important to design scheduling algorithms that can be adapted to various system configurations and task characteristics.The objective of this project is to develop adaptive efficient algorithms for scheduling co-design problems for various types of cyber-physical systems subjected to various timing and resource constraints. The plan is to adopt learning-based methods that are able to learn an implicit model for voltage selection or scheduling strategy selection for the underlying cyber-physical system based on scheduling history. This method is especially useful when the task features and architecture model are unknown to (or too complex to be considered by) the DVFS scheduler. The problem of extracting good features to serve as input for an implicit model for the learning-based method, that can best represent the model of an underlying cyber-physical system, will also be explored. Currently, Q-learning, Double Q-learning and Deep Double-Q-learning will be explored. Evaluation of the framework will also be studied extensively. The framework designed will be used to explore systems with various types of tasks (dependent, independent, periodic or non-periodic, etc.), various types of scheduling policy (earliest deadline first, fixed-priority, etc.), various types of system configurations (single-core, multi-core, GPU, NOC-based, wearable type of devices, etc.).