Real-Time Scheduling of Multistage IDK-Cascades

An IDK classifier is a software component that categorizes each input provided to it into one of a fixed set of “classes,” or outputs an “I don't know” (IDK) to indicate that it is unable to classify this input. An IDK-cascade is a linear arrangement of different IDK classifiers for the same classification problem, which are executed in sequence on a given input until one outputs an actual class (rather than IDK). Given a multistage computation that must be completed within a specified hard end-to-end deadline and a choice of classifiers, one deterministic and one an IDK-cascade, for each stage, the problem of determining which IDK-cascades to schedule in order to minimize the expected end-to-end response time while guaranteeing to meet the specified deadline is considered. Different variants of this problem are defined, and optimal algorithms for solving them are derived.