Unique Kernel Diagnosis by Anders Henriksson

The idea of using logic in computer programs to perform systematic diagnosis was introduced early in computation history. There are several systems using punch-cards and rulers described as early as the mid 1950’s. Within the area of applied artificial intelligence the problem of diagnosis made its definite appearance in the form of expert systems during the 1970’s. This research eventually introduced model based diagnosis in the field of artificial intelligence during the mid 1980’s. Two main approaches to model based diagnosis evolved: consistency based diagnosis and abductive diagnosis. Later kernel diagnosis complemented these two approaches. Unique kernel diagnosis is my contribution to model based diagnosis within artificial intelligence. Unique kernel diagnosis addresses the problem of ambiguous diagnoses, situations where several possible diagnoses exist with no possibility to determine which one describes the actual state of the device that is diagnosed. A unique kernel diagnosis can per definition never be ambiguous. A unique kernel diagnosis can be computed using the binary decision diagram (BDD) data structure by methods presented in this thesis. This computational method seems promising in many practical situations even if the BDD data structure is known to be exponential in size with respect to the number of state variabels in the worst case. Model based diagnosis in the form of consistency based-, abductive and kerneldiagnosis is known to be an NP-complete problem. A formal analysis of the computational complexity of the problem of finding a unique kernel diagnosis reveals that it is in PNP.