CPS: Small: Neuro-Symbolic Learning and Control with High-Level Knowledge Inference

人工智能在网络物理系统中的使用受到数据可用性、任务环境复杂性以及对可表达和可解释的高级知识表示的需求等挑战的限制。为了应对这些挑战，该项目旨在通过整合机器学习、控制理论和形式化方法来开发一套神经符号学习和控制工具。这一成果有望在机器人系统、自主系统和联网的网络物理系统等网络物理系统中得到应用。研究任务主要集中于提高网络物理系统中神经符号方法的数据效率、可解释性、可扩展性和弹性，特别关注网络物理系统中广泛的在线交互可能不安全或不可行的网络物理系统，但高级知识在具有可证明正确性保证的复杂和对抗性环境中对于完成任务是有用的。具体地说，神经符号学习和控制算法提供了以下三个关键特征。首先，它们将来自不确定数据的高级知识(例如，时态逻辑公式)的无模板推理融入到神经符号学习中，以实现数据效率高的强化学习，包括离线训练和在线微调。其次，该方法能够在多智能体对抗性环境中实现可伸缩和弹性的强化学习，在这种环境中可以推断高级知识来加速智能体的学习过程。最后，该算法使用基于神经符号学习的未知动态自适应控制，为复杂任务规范提供了可证明的保证。这项研究的成果将通过一系列教育活动分享，包括本科生和研究生课程以及研究讲习班。此外，将实施外展计划，以吸引K-12学生和科学和工程社区的不同群体，并扩大他们的参与。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。

The use of artificial intelligence in cyber-physical systems is limited by challenges such as data availability, task environment complexity, and the need for expressive and interpretable high-level knowledge representations. To address these challenges, this project aims to develop a set of neuro-symbolic learning and control tools by integrating machine learning, control theory, and formal methods. The results are expected to find application across cyber-physical systems such as robotic systems, autonomous systems, and networked cyber-physical systems. Validation in a testbed environments should facilitate safe deployments in real-world physical environments with provable guarantees and robustness against potential adversaries.The research tasks mainly focus on improving the data efficiency, interpretability, scalability, and resiliency of the neuro-symbolic approaches in cyber-physical systems, with special focus on cyber-physical systems where extensive online interactions may not be safe or feasible, but high-level knowledge can be useful in completing tasks in complex and adversarial environments with provable correctness guarantees. Specifically, the neuro-symbolic learning and control algorithms offer the following three key features. Firstly, they incorporate template-free inference of high-level knowledge (e.g., temporal logic formulas) from uncertain data into neuro-symbolic learning to enable data-efficient reinforcement learning with both offline training and online fine-tuning. Secondly, the approaches are capable of achieving scalable and resilient reinforcement learning in multi-agent adversarial environments where high-level knowledge can be inferred for expediting the agents' learning processes. Lastly, the algorithms provide provable guarantees on complex task specifications using neuro-symbolic learning-based adaptive control with unknown dynamics. The results of this research will be shared through a range of educational activities, including undergraduate and graduate courses, as well as research workshops. Additionally, outreach programs will be implemented to engage K-12 students and diverse groups in the scientific and engineering communities and broaden their participation.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.