CAREER: Towards Safety-Critical Real-Time Systems with Learning Components

点击翻译按钮获取中文摘要

In the rapidly advancing world of artificial intelligence (AI) and its widespread applications, today's safety-critical systems, ranging from self-driving cars to surgical robots, increasingly rely on learning-enabled modules. Ensuring the temporal safety of these AI-driven systems is crucial, especially in high-stakes and time-critical settings. Alongside improving functionality, accuracy, and efficiency, these systems must be safe in meeting real-time constraints despite worst-case scenarios and extreme events. However, safety-critical systems with learning components differ significantly from traditional ones, exhibiting greater dynamism and complexity in various aspects. Designing performant real-time scheduling strategies and conducting rigorous yet appropriately pessimistic analyses for proving temporal safety becomes much more challenging. This project aims to address this challenge by creating a framework that seamlessly integrates real-time safety verification and assurance into the performance optimization process of AI-driven safety-critical systems. It will contribute to the advancement of critical technologies of modern autonomous systems with learning capabilities that need to respond to highly dynamic internal and external environments. Moreover, the project places a strong emphasis on integrating research into educational and outreach activities to promote diversity in STEM education and broaden participation in computing and engineering.This project aims to provide real-time safety guarantees while optimizing the average performance and efficiency of safety-critical systems with learning components. To achieve this, the project will devise innovative real-time scheduling strategies specifically tailored to handle the intricate and dynamic computational workloads of these systems for achieving temporal safety with minimum performance loss. Theoretical analyses will be derived to bound the worst-case timing behavior while simultaneously maximizing average performance and efficiency. Additionally, it will pioneer safe reinforcement learning mechanisms designed to efficiently learn and optimize system performance while ascertaining temporal safety upon deployment. The project lays the foundation for addressing the grand challenge of providing temporal safety guarantees for AI-driven safety-critical systems while simultaneously maximizing their overall performance, including enhancing the capability of achieving mission-level goals, optimizing energy-saving, and improving system efficiency.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.