CAREER: Learning for Real-Time Embedded Optimization

在当今不断变化的世界中，网络物理系统必须能够通过实时做出高质量的决策，对不断变化的条件和意外的中断做出安全的反应。数学优化为制定和解决决策问题提供了一种强大而灵活的工具，在工程、计算机科学和运筹学中有着广泛的应用。然而，将优化应用于实时决策一直受到两个基本挑战的限制：确保对不确定问题参数(例如，能量需求、障碍位置)的健壮性；以及确保在计算资源有限的嵌入式平台(例如，嵌入式微控制器)上的实时算法收敛。该项目寻求使用机器学习来开发下一代实时优化工具来安全地控制网络物理系统来解决这些挑战。这些结果将导致用于嵌入式优化的开源软件，这将使其在广泛的工程学科(例如，能源、交通和自主系统)中的应用变得更加容易，在这些领域，实时优化被认为计算过于昂贵和不可靠。该项目的研究成果将被整合到协同教育计划中，通过促进优化和计算思维来培养未来的工程师。这些举措包括为本科生和研究生开发优化核心课程的新学习单元；为研究生开设关于学习、决策和控制的新的两年一次的研讨会；以及一个由学生创建的新播客。为了实现这些结果，我们将利用实时决策的重复性质。在实际场景中，我们用不同的数据重复求解相同的优化问题。例如，随着传感器信号(例如，系统状态)和目标(例如，期望轨迹)的变化，控制系统更新输入动作。该项目的关键洞察力是利用这种结构来开发可靠和高效的专门的数据驱动优化工具。(I)我们将学习数据驱动的问题公式，并为不确定情况下的决策提供约束满足保证。(Ii)我们将开发数据驱动的工具来设计和数值验证实时优化算法，并提供严格的收敛保证。在这项研究中开发的嵌入式优化技术将通过在用于空间探索和无人机的实时嵌入式平台上进行广泛的真实世界实验来验证。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。

In today’s constantly changing world, cyber-physical systems must be able to safely react to changing conditions and unexpected disruptions by making high-quality decisions in real-time. Mathematical optimization offers a powerful and flexible tool to formulate and solve decision-making problems, with numerous applications in engineering, computer science, and operations research. However, applying optimization to real-time decision-making has always been limited by two fundamental challenges: ensuring robustness to uncertain problem parameters (e.g., energy demand, obstacle positions); and ensuring real-time algorithm convergence on embedded platforms with limited computational resources (e.g., embedded microcontrollers). This project seeks to address these challenges using machine learning to develop the next-generation real-time optimization tools to safely control cyber-physical systems. These results will lead to open-source software for embedded optimization, which will ease their application in a wide range of engineering disciplines (e.g., energy, transportation, and autonomous systems) where real-time optimization is considered too computationally expensive and unreliable. The research outcomes of this project will be integrated into a synergistic education plan to train future engineers by promoting optimization and computational thinking. These initiatives include the development of new learning units for undergraduate and graduate core courses in optimization; a new biennial workshop for graduate students on learning, decision, and control; and a new student-created podcast.To achieve these results, we will exploit the repetitive nature of real-time decision-making. In practical scenarios, we solve the same optimization problem repeatedly with different data. For example, control systems update the input actions as sensor signals (e.g., system state) and goals (e.g., desired trajectory) vary. The key insight of this project is to exploit this structure to develop specialized data-driven optimization tools that are reliable and efficient. (i) We will learn data-driven problem formulations with constraint satisfaction guarantees for decision-making under uncertainty. (ii) We will develop data-driven tools to design and numerically verify real-time optimization algorithms, with tight convergence guarantees. The embedded optimization techniques developed in this research will be validated through extensive real-world experiments on real-time embedded platforms for space exploration and unmanned aerial vehicles.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.