The Safe and Effective Use of Low-Assurance Predictions in Safety-Critical Systems

The algorithm-design paradigm of algorithms using predictions is explored as a means of incorporating the computations of lower-assurance components (such as machine-learning based ones) into safety-critical systems that must have their correctness validated to very high levels of assurance. The paradigm is applied to two simple example applications that are relevant to the real-time systems community: energy-aware scheduling, and classification using ML-based classifiers in conjunction with more reliable but slower deterministic classifiers. It is shown how algorithms using predictions achieve much-improved performance when the low-assurance computations are correct, at a cost of no more than a slight performance degradation even when they turn out to be completely wrong.