Control in the presence of limited information - systematic design methods with performance guarantees for distributed systems.

从卫星星座和飞机到电网、计算机网络和制造过程，我们的现代社会严重依赖于复杂系统的运行，其中许多系统可以建模为动力系统网络。控制行动是必不可少的，以确保这些系统的正常和安全运行，并达到规定的性能目标。然而，在这种分布式系统中，有限的通信和有限的处理资源对控制设计提出了一个主要挑战：在缺乏“完美信息”的情况下，设计具有稳定性和性能保证的控制器。在这个项目中，我们将探索信息在线性和非线性网络系统控制中的作用，目的是开发系统控制设计方法。在这种情况下，任何控制设计都应该完全基于有限的局部信息，但它应该为全局系统提供保证。这需要分布式控制方法。为此，我们将研究如何使用测量数据来解释分布式系统背景下缺失的系统信息。许多现代流程生成并存储大量数据。因此，开发利用测量数据进行控制器设计的方法是一个令人兴奋的、积极研究的课题。机器学习和控制的合并是解决这一挑战的一种流行方法。波士顿动力公司（Boston Dynamics）的Atlas机器人表演后空翻等引人入胜的演示凸显了人工智能的潜力。然而，一个悬而未决的问题仍然存在：如何保证在嘈杂的现实条件下安全运行？复杂算法容易出现bug，性能对初始条件很敏感。这些不确定性在安全关键系统（如航空航天系统）的控制中没有地位。在这个项目的初始阶段，我们将建立一个低复杂性的学习框架，允许直接从数据中计算具有稳定性、性能和鲁棒性保证的未知线性系统的控制策略。我们将探索如何将这个框架引入动态博弈和时变系统领域。前者是一种强大的工具，可以模拟分布式系统中战略代理之间的互动，而后者既存在于特定类型的游戏环境中，也存在于网络系统的环境中，其中代理之间的互动可能会随着时间的推移而进化和改变。然后，我们将把注意力转向非线性系统，考虑数据驱动控制以及在分布式控制背景下克服信息缺乏的替代方法。例如，我们将探索如何利用系统属性（如无源性）来获得控制器保证，尽管可用信息有限，以及如何有效地使用可用的部分信息，而无需用数据替代所有系统依赖关系。该项目的最终目标是开发系统的方法，为线性和非线性分布式系统设计稳定，鲁棒和高效的控制器。这些方法的发展可以改变当前和未来各种工程学科的一系列技术，包括但不限于电力系统、工业过程、空中交通管制、农业、机器人和太空探索。

From satellite constellations and aeroplanes to power grids, computer networks and manufacturing processes, our modern society heavily depends on the operation of complex systems, many of which can be modelled as networks of dynamical systems. Control action is indispensable to ensure proper and safe operation of such systems, and to meet specified performance objectives. However, restricted communication and limited processing resources in such distributed systems present a major challenge for control design: designing controllers with stability and performance guarantees despite a lack of "perfect information".In this project, we will explore the role of information in the control of linear and nonlinear networked systems, with the aim of developing systematic control design methods. In this context, any control design should be based solely on limited local information, yet it should provide guarantees for the global system. This calls for distributed control methods. Towards this end, we will study how measured data can be used to account for missing system information in the context of distributed systems. Many modern processes generate and store large amounts of data. Consequently, developing methods to utilise measured data for controller design is an exciting, actively researched topic. The merger of machine learning and control is a popular approach addressing this challenge. Fascinating demonstrations, such as Boston Dynamics' Atlas robot performing backflips, highlight its potential. However, a major open question remains: how to guarantee safe operation under noisy real-world conditions? Complex algorithms are susceptible to bugs and performance is sensitive to initial conditions. These are uncertainties which have no place in the control of safety critical systems, such as aerospace systems. In the initial phase of this project, we will build upon a low-complexity learning framework allowing to compute control policies with stability, performance and robustness guarantees for unknown linear systems directly from data. We will explore how this framework can be introduced to the areas of dynamic games and time-varying systems. The former represents a powerful tool to model the interactions between strategic agents in a distributed system, whereas the latter is encountered both in the context of certain classes of games and in the context of networked systems, where interactions between agents may evolve and change over time.We will then turn our focus to nonlinear systems, considering data-driven control as well as alternative methods to overcome lack of information in the context of distributed control. We will, for instance, explore how system properties, such as passivity, can be utilised to derive controller guarantees, despite limited available information, and how available partial information can be used efficiently, without substituting all system dependencies with data.The ultimate goal of the project is to develop systematic methods to design stable, robust and efficient controllers for linear and nonlinear distributed systems. The development of such methods can be a game-changer for a range of current and future technologies across various engineering disciplines including, but not limited to, power systems, industrial processes, air traffic control, agriculture, robotics, and space exploration.