CAREER: Switching and Logic in Control

[00:93762 . hespanha2]在本提案中，我们在自动控制领域划定了一个五年的研究和教育计划。该计划的总体目标是为致力于科学研究的职业生涯奠定基础，并将其融入南加州大学的教育计划。我们的研究和教育工作将集中在混合控制系统领域，即结合连续动态与离散逻辑的系统。我们对利用离散逻辑进行在线学习和适应的监督控制算法特别感兴趣。在这种情况下使用逻辑是在超过15年前的控制文献中提出的。早期的大部分工作都是理论性的，其动机是希望构建能够稳定非常大的系统类的通用控制器。这些控制器的通用性是以非常差的性能为代价实现的，并且所得算法不能用于大多数实际应用。从那时起，在这一领域取得了很大的进展，现在监督控制实际上被提出作为一种技术，以获得非常高的性能和鲁棒性，以应对过程或控制目标的突然变化。我们研究的科学目标是开发一个框架，以加快利用切换和逻辑的学习算法的分析和设计。为此，我们将分离监督控制算法设计中出现的基本问题，并以系统的方式对其进行探索。我们在本提案中表明，尽管文献中提出的许多算法起源于根本不同的方法，但它们具有关键结构并表现出几个共同的特性。一个通用的框架将有助于更好地理解它们的主要特性和局限性。我们在这里分离出两个我们认为对监督控制的广泛建立至关重要的具体主题：1。将这些算法的应用范围扩展到复杂非线性系统的控制。特别是，当只有输出反馈可用和/或当有显著的测量噪声和未建模的动态。开发理论工具来优化这些系统的性能。在这里，性能应该从瞬时和渐近的意义上理解。

0093762HespanhaIn this proposal we delineate a five year-research and education plan in the area of Automatic Control. The overall objective of this plan is to build the foundation for a career devoted to scientific research and its integration in the educational program at the University of Southern California.Our research and education efforts will be focused in the area of hybrid control systems, i.e., systems that combine continuous dynamics with discrete logic. We are particularly interested in supervisory control algorithms that utilize discrete logic for the purpose of online learning and adaptation. The use of logic in this context was proposed in the control literature more than fifteen years ago. Most of the early work was of a theoretical nature, motivated by the desire to construct universal controllers capable of stabilizing very large classes of systems. The universality of these controllers was achieved at the expense of very poor performance and the resulting algorithms could not be used in most practical applications. Since then a great deal of progress has been made in this area and now supervisory control is actually proposed as a technique to attain very high performance and robustness to sudden changes in the process or the control objectives.The scientific goal of our research is to develop a framework to expedite the analysis and design of learning algorithms that utilize switching and logic. Towards this end we will isolate the fundamental issues that arise in the design of supervisory control algorithms and explore them in a systematic fashion. We show in this proposal that although many of the algorithms proposed in the literature originate from fundamentally different approaches, they share key structures and exhibit several common properties. A common framework will provide a better understanding of their key features and limitations. We isolate here two specific topics that we judge crucial for the wide establishment of supervisory control:1. Extend the domain of application of these algorithms to the control of complex nonlinear systems. In particular, when only output-feedback is available and/or when there is significant measurement noise and unmodeled dynamics.2. Develop theoretical tools to optimize the performance of these systems. Here, performance should be understood both in a transient and in an asymptotic sense.