Merging Indiscriminable Diagnoses: An Approach Based on Automatic Domains Abstraction

This paper describes an algorithm which automatically abstracts domains of variables in qualitative models specified with propositional logic, by mapping sets of values at the detailed level into new qualitative values at the abstract level. The abstraction is driven by the subset of observations available and/or their granularity, and guarantees that given such restricted information the discrimination power of the model is unchanged (i.e. the aggregated sets of values are those that couldn't be further discriminated at the detailed level). By considering increasingly restricted classes of available observations (and/or granularity of observations), a set of abstract models can be derived that can be exploited with model selection at problem solution time. We further describe an implementation of the algorithm in the context of a diagnostic agent for a space robotic arm, discuss experimental results obtained using the abstracted models and compare them with corresponding results obtained using the detailed model.