Realizing the Promise of Artificial Intelligence for Unmanned Aircraft Systems through Behavior Bounded Assurance

摘要

A key value proposition for incorporation of Artificial Intelligence (AI) and Machine Learning (ML) methods into aviation is that they offer means of understanding data in ways that allow hitherto unprecedented insights for decision making, whether by a human or a machine. When these techniques are applied to cyber-physical systems, such as unmanned aircraft systems (UAS), they can result in positive societal impacts (e.g., search and rescue). However, the advantages of such techniques must be balanced against appropriate safety and security requirements so that taken together the system can ensure an acceptable level of confidence and assurance in both civilian and military applications. To this end, there is a need for the capability to suitably characterize such techniques and assess how they can be integrated into a viable assurance framework that can maximize safety and security benefits while bounding the inherent risk of non-determinism arising from such these approaches. This paper focuses on assurance and behavior bounds for decision making systems from a) algorithmic functional performance; b) schedulability analysis and candidate scheduling paradigms; and c) processor architectures (including multi-core) to support minimized interference in general. We will place particular emphasis on machine learning approaches for control, navigation and guidance applications for unmanned systems. This paper will review available and emerging approaches (e.g., formal methods, modeling and simulation, real-time monitors/agents among others) to ensuring behavior assurance for unmanned systems engaged in missions of moderate-to-high complexity. The intent is to examine behavior assurance for advanced autonomous operations within a holistic life-cycle process