Continuous Engineering for Trustworthy Learning-Enabled Autonomous Systems
摘要
Learning-enabled autonomous systems (LEAS) use machine learning (ML) components for essential functions of autonomous operation, such as perception and control. LEAS are often safety-critical. The development and integration of trustworthy ML components present new challenges that extend beyond the boundaries of system’s design to the system’s operation in its real environment. This paper introduces the methodology and tools developed within the frame of the FOCETA European project towards the continuous engineering of trustworthy LEAS. Continuous engineering includes iterations between two alternating phases, namely: (i) design and virtual testing, and (ii) deployment and operation. Phase (i) encompasses the design of trustworthy ML components and the system’s validation with respect to formal specifications of its requirements via modeling and simulation. An integral part of both the simulation-based testing and the operation of LEAS is the monitoring and enforcement of safety, security and performance properties and the acquisition of information for the system’s operation in its environment. Finally, we show how the FOCETA approach has been applied to realistic continuous engineering workflowsfor three different LEAS from automotive and medical application domains.