CPS: Medium: Collaborative Research: Learning and Verifying Conformant Data-Driven Models for Cyber-Physical Systems

这个项目研究了建立数学模型的基本技术，这些模型可以安全地用于做出可信的预测和控制决策。数学模型是现代信息物理系统（CPS）的基础。例如，预测刹车时汽车如何移动的车辆模型，或预测1型糖尿病患者在注射胰岛素时血糖水平如何变化的生理模型。机器学习工具的成功产生了数据驱动的模型，如神经网络。然而，根据收集数据和学习模型的方式，有可能获得违反基本物理、化学或生理事实的模型，这些事实可能会威胁到生命和财产。该项目的方法是通过高级分析来暴露这些模型缺陷。该项目旨在通过指导活动，鼓励大学生女性和代表性不足的少数群体成员考虑从事研究工作，从而扩大对计算机的参与。该研究将揭露失败的证伪方法与严格证明一致性的验证方法相结合。此外，从数据中学习动力系统模型和施加核心网络物理领域知识的方法正在研究中。该项目将这些具有一致性保证的数据驱动模型应用于自动驾驶汽车的安全控制器设计、人类胰岛素葡萄糖调节模型和机器人群。通过建立以CPS为中心的远程教育框架，努力推进CPS教育。研究人员正在开发一系列低成本的硬件测试平台和自定进度的学习任务，这些任务将使学生接触到构建高度可靠和安全关键CPS的过程。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

This project investigates fundamental techniques for building mathematical models that can be safely used to make trustworthy predictions and control decisions. Mathematical models form the foundation for modern Cyber-Physical Systems (CPS). Examples include vehicle models that predict how a car will move when brakes are applied, or physiological models that predict how the blood glucose levels change in a patient with type-1 diabetes when insulin is administered. The success of machine learning tools has yielded data-driven models such as neural networks. However, depending on how data is collected and the models are learned, it is possible to obtain models that violate fundamental physical, chemical, or physiological facts that can potentially threaten life and property. The approach of the project is to expose these model flaws through advanced analysis. The project seeks to broaden participation in computing through mentoring activities that will encourage undergraduate women and members of underrepresented minority groups to consider a career in research.The research combines falsification methods for exposing failure to conform with verification approaches for rigorously proving conformance. Furthermore, approaches for learning models of dynamical systems from data and imposing core cyber-physical domain knowledge are under investigation. The project is applying these data-driven models with conformance guarantees to the design of safe controllers for autonomous vehicles, models of human insulin glucose regulation and robotic swarms. The effort is advancing CPS education by creating a framework for distance education focused on CPS. The researchers are developing a series of low cost hardware testbeds and self-paced learning tasks that will expose students to the process of building highly reliable and safety critical CPS.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.