Career: Correct-by-Learning Methods for Reliable Autonomy

让人们在身体上高度自主的计算系统需要提供严格的安全保障。可以在这样的系统上使用形式化方法来提供数学证明，以确保正确的行为。然而，机器学习和数据驱动的方法现在是自主系统设计中不可或缺的一部分，它们对高度非线性连续函数和概率推理的依赖在很大程度上与形式方法中的逻辑和符号分析框架不一致。因此，缺乏正式的保证已成为阻碍自主系统更广泛部署和采用的关键瓶颈。该项目针对这一开放挑战，为自主系统的基于学习和数据驱动的控制和规划方法开发了正式的综合和验证技术。该项目为提高自动驾驶汽车和无人驾驶飞行器等现实自主系统的基本可靠性开发了理论基础以及实用技术和工具。这项工作建立在研究人员先前在连续域和混合域上的形式方法方面的工作基础上，以期在实际工程中统一符号方法和数值方法。这种逐次学习的方法可以确保一般的基于学习的决策算法在人工智能应用的各个领域都具有安全性和可信性。教育和外联活动是该项目的重要组成部分，因为所制定的方法只有在广泛的工程领域中被自主系统的新一代开发者采用时才有用。教育工作直接有助于提高实际工程中正式方法的自动化和可用性。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。

Computing systems that engage people physically with high degrees of autonomy need to provide rigorous guarantees of safety. Formal methods can been used on such systems to provide mathematical proofs to ensure correct behavior. However, machine learning and data-driven approaches are now an indispensable part of autonomous-systems design, and their reliance on highly nonlinear continuous functions and probabilistic reasoning has largely been at odds with the logical and symbolic-analysis frameworks in formal methods. As a result, the lack of formal assurance has become the key bottleneck that impedes the wider deployment and adoption of autonomous systems. This project targets this open challenge by developing formal synthesis and verification techniques for learning-based and data-driven control and planning methods for autonomous systems. The project develops the theoretical foundations as well as practical techniques and tools for improving the fundamental reliability of realistic autonomous systems such as autonomous cars and unmanned aerial vehicles. The work builds on the investigator's prior work on formal methods over continuous and hybrid domains towards the unification of symbolic and numerical methods in practical engineering. The correct-by-learning methods can ensure formal properties of general learning-based decision-making algorithms towards safety and trustworthiness in all areas of AI applications. The education and outreach activities are a crucial part of the project, as the methods developed are only useful if they are adopted by the new generations of developers of autonomous systems in a broad range of engineering domains. The education efforts contribute directly to improving automation and usability of formal methods in practical engineering.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.