Impact of Machine Learning on Safety Monitors

Abstract

Machine Learning components in safety-critical applications can perform some complex tasks that would be unfeasible otherwise. However, they are also a weak point concerning safety assurance. An aspect requiring study is how the interactions between machine-learning components and other non-ML components evolve with training of the former. It is theoretically possible that learning by Neural Networks may reduce the effectiveness of error checkers or safety monitors, creating a major complication for safety assurance. We present an initial exploration of this problem focused on automated driving, where machine learning is heavily used. We simulated operational testing of a standard vehicle architecture, where a machine learning-based Controller is responsible for driving the vehicle and a separate Safety Monitor is provided to detect hazardous situations and trigger emergency action to avoid accidents. Among the results, we observed that indeed improving the Controller could make the Safety Monitor less effective; it is even possible for a training increment to make the Controller's own behaviour safer but the vehicle's less safe. We discuss implications for practice and for research.

Bootstrapping confidence in future safety from past safe operation

“…However, this does not solve the problem of demonstrating that a type A AV will be as safe as required. Firstly, hazardous situations may be subtle to detect and hard to resolve; more importantly, it does not change the cost of statistical demonstration of safety, because the effectiveness of a monitor for a particular AV cannot be estimated independently of the specific AV [2], [17], [18].…”