Approximate Model-Based Diagnosis Using Preference-Based Compilation

This article introduces a technique for improving the efficiency of diagnosis through approximate compilation. We extend the approach of compiling a diagnostic model, as is done by, for example, an ATMS, to compiling an approximate model. Approximate compilation overcomes the problem of space required for the compilation being worst-case exponential in particular model parameters, such as the path-width of a model represented as a Constraint Satisfaction Problem. To address this problem, we compile the subset of most “preferred” (or most likely) diagnoses. For appropriate compilations, we show that significant reductions in space (and hence on-line inference speed) can be achieved, while retaining the ability to solve the majority of most preferred diagnostic queries. We experimentally demonstrate that such results can be obtained in real-world problems.