Memory-limited model-based diagnosis

How should I compute my candidates? A taxonomy and classification of diagnosis computation algorithms

“…Only recently, a space-efficient algorithm exhibiting all desirable properties wrt. the features discussed here, along with a reasonable time performance, has been suggested [23].…”

“…Relevance: The best-first property is useful, e.g., if (one expects) there is a large number of diagnoses and the actual diagnosis is likely among the best diagnoses (e.g., when all components fail with an equal and small likelihood [13]), if focusing techniques are adopted where only the best subset of all diagnoses is further considered [13], if informative samples for sequential diagnosis should be computed [35], or if users intend to monitor the best diagnoses throughout the debugging process [23].…”

“…Impact: Forgoing the best-first requirement usually leads to a higher efficiency of diagnosis computation, as any-first algorithms can use more efficient (e.g., depth-first [26] instead of breadth-first [12] or uniformcost [19]) diagnosis search techniques. Also, generally best-first methods might be significantly more expensive than related focused best-first ones (cf., e.g., [23]).…”

“…Relevance: If conflict computation or theorem proving is very expensive in a diagnostic scenario (cf., e.g., [9]), then direct algorithms are preferable, or even the only feasible approach. If a systematic exploration of diagnoses (allowing, e.g., inferences that all diagnoses with a specific property, say minimum-cardinality, are already computed) is desired [13] or the general bestfirst property (see bullet 3) is relevant [23], or the used method should be optimized for certain sequential diagnosis problems [24], then a conflict-dependent approach might be the only viable choice. When adopting a conflict-dependent method, it is important to note that an adequate conflict generation approach-which is sound and complete wrt.…”

“…E.g., when DPIs expressible by means of propositional logic are the target use case of a diagnosis system, then a reasoner-dependent algorithm based on propositional logic might be preferable to a black-box one with otherwise equal features. If, on the other hand, different formalisms might be used to describe the faulty system [23], black-box techniques can be more expedient. They can always simply use the best reasoner for the particular problem at hand without needing to incorporate any modifications into the reasoner (or the diagnosis algorithm)unlike reasoner-dependent approaches, which require the incorporation of the necessary additional mechanisms into any adopted reasoner.…”