CAREER: Data-Driven Control of Dynamical Networks: Robustness, Risk, and Network Architectures

最近惊人的技术进步和前所未有的数据可用性为从根本上重新设想新兴复杂网络系统的控制和信息架构和算法提供了机会，例如用于运输和交付的自主机器人团队，以及具有大规模可再生能源渗透率的能源电网。然而，这些网络对安全、高效和强大的操作和控制构成了重大挑战。它们日益增长的复杂性，特别是来自机器学习组件的快速集成，可能会超出我们对其健壮性属性的理解，并导致严重的故障和安全风险。该项目将为动态网络的稳健、风险约束的数据驱动控制算法和架构的分析和设计建立一个严格的框架。拟议的创新将通过增强对未来社会至关重要的各种新兴复杂网络的安全性、效率和稳健性，产生广泛的社会影响。应用的重点将是自主能源电网和分布式多机器人团队，但基础知识的进步也可能影响其他新兴的关键基础设施网络。此外，综合研究和教育计划具有广泛的活动，将培训科学家和工程师，并促进公众对动态网络中的数据驱动控制的理解，特别是关于稳健性、风险和安全问题。该项目将通过基于模型的控制和基于数据的学习方法的变革性、稳健和风险感知的集成，明确纳入有限数据估计的不确定性，从而显著促进知识的发展。整个方法将结合基于模型的随机控制、动态博弈论和基于风险的分布式稳健优化的性能和健壮性保证，以及现代的非渐近鞅集中界和来自统计学的自举技术。对有限数据模型估计中的不确定性的稳健性将以两种创新的方式明确地结合在一起：(1)丰富的随机动态博弈框架，它结合了对抗性输入和乘性噪声，以提高对高度结构化的参数和非参数不确定性的稳健性；(2)使用公理风险理论和分布稳健优化来保证基于风险的有意义的安全约束。在此基础上，该项目将开发完全数据驱动的动态输出反馈控制、主动探索以及使用迭代随机动态博弈的非线性系统的局部最优、稳健和可证明的收敛算法。拟议的研究还将通过开发创新的自我调整架构和用于网络控制架构设计的正规化策略优化算法来提升知识，并阐明底层网络结构和规模如何对数据驱动的控制和估计施加基本限制。该方法将通过自主电网和分布式异质多机器人团队的数值模拟和硬件试验台实验来说明、改进和验证。作为补充的教育计划将：(1)通过在佩罗自然科学博物馆和德克萨斯大学达拉斯分校广泛的教育推广计划创建互动展览；(2)通过新课程、互动网络技术和基于项目的学习来增强动态网络的课程；(3)指导直接支持拟议研究的高级顶峰设计项目和本科生研究人员。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。

Recent spectacular technological advances and unprecedented availability of data are providing opportunities to fundamentally reimagine the control and information architectures and algorithms for emerging complex networked systems, such as autonomous robot teams for transportand delivery and energy grids with massive renewable penetration. However, these networks are posing major challenges for safe, efficient, and robust operation and control. Their accelerating complexity, especially from rapid integration of machine learning components, threatens to outpace our understanding of their robustness properties and lead to severe failure and safety risks. This project will build a rigorous framework for analysis and design of robust, risk-constrained data-driven control algorithms and architectures for dynamical networks. The proposed innovations will have broad societal impact by enabling enhancements to the safety, efficiency and robustness of various emerging complex networks that are vital to future society. The application focus will be on autonomous energy grids and distributed multi-robot teams, but the fundamental knowledge advancements can also impact other emerging critical infrastructure networks. Furthermore, the integrated research and education plan features an extensive array of activities that will train scientists and engineers and promote public understanding of data-driven control in dynamical networks, especially around issues of robustness, risk, and safety.This project will significantly advance knowledge via a transformative robust and risk-aware integration of model-based control and data-based learning approaches that explicitly incorporate uncertainty from finite-data estimates. The overall approach will combine performance and robust-ness guarantees from model-based stochastic control, dynamic game theory, and risk-based distributionally robust optimization with modern non-asymptotic martingale concentration bounds and bootstrap techniques from statistics. Robustness to uncertainties in finite-data model estimates will be explicitly incorporated in two innovative ways: (1) a rich stochastic dynamic game framework that combines adversarial inputs and multiplicative noise to promote robustness to both highly-structured parametric and non-parametric uncertainties; (2) the use of axiomatic risk theory and distributionally robust optimization to guarantee meaningful risk-based safety constraints. On this basis, the project will develop techniques for fully data-driven dynamic output feedback control, active exploration, and locally optimal, robust, and provable convergent algorithms for nonlinear systems using iterative stochastic dynamic games with locally learned and refined models. The proposed research will also advance knowledge by developing innovative self-tuning architectures and regularized policy optimization algorithms for network control architecture design, and illuminating how underlying network structure and size impose fundamental limits on data-driven control and estimation. The approach will be illustrated, refined, and validated via numerical simulations and hardware testbed experiments for autonomous energy grids and distributed heterogeneous multi-robot teams. A complementary education plan will: (1) create interactive exhibitions via extensive educational outreach programs at the Perot Museum of Nature & Science and UT Dallas; (2) enhance curriculum for dynamical networks with new courses, interactive web technologies, and project-based learning; (3) mentor senior capstone design projects and undergraduate researchers that directly support the proposed research.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.