Diagnosing Tree Structured Systems

This paper introduces the algorithm TREE DIAG for computing minimal diagnoses for tree structured systems. Diagnoses are computed by descending into the tree, enumerating the input combinations that might be reponsible for a given incorrect observation, and combining the diagnoses for the subtrees generating these inputs into diagnoses for the whole system. We prove soundness and correctness of the algorithm and show experimental results that indicate that it compares favorably to Reiter's hitting-set-based algorithm and El Fattah and Dechter's SAB. Extensions of the algorithm related to general acyclic systems, use of fault modes and the practical application to the software diagnosis domain are discussed.