An architecture for diagnostic reasoning based on causal models

Sammanfattning: This thesis proposes an architecture for diagnostic reasoning employing causal models. The architecture embodies a knowledge representation at the problem-solving layer. That is, the way a problem is solved—in this case to arrive at a diagnosis—is an integral part of the knowledge representation. Diagnostic problem solving normally requires an explanation of what caused an observed behaviour of a system. Finding the cause for an observation or symptom is often crucial if the correct diagnosis for a malfunction in a device is to be found. Causal reasoning requires a deep modelling of the domain. The causal model relates observations to each other and provides explanations of how and why an observation was caused. Diagnosing a complex device, with diagnoses having symptoms that are not mutually exclusive, often requires model-based reasoning techniques, where causal reasoning is one possible choice. A causal model is based on a qualitative and quantitative interpretation of the observations done on the device that is to be diagnosed; each causal model is uniquely related to a specific problem scenario. A symptom to a diagnosis is justified by determining how it was caused. Causes to a symptom are described through a causal model called a scenario model. Each scenario model has constraints associated with it. A constraint can represent both assumptions and conditions needed to be satisfied when applying a scenario model. The whole scenario model, or parts of it, is not used when constraints are not satisfied. Diagnosing the device behaviour is done by testing to find out which scenario models in the knowledge base correspond closest to the actual observation. The knowledge base contains a number of scenario models, each scenario model is generated from a domain model when building the knowledge base. The domain model is a collection of causal relationships—relating events causally with each other. The proposed architecture was used for diagnosing performance problems in computer systems. A prototype implementation based on the proposed architecture was successfully diagnosing different performance problems. Further, the same knowledge base was successfully used to diagnose the same kind of performance problem at different computer systems. Some of the key problems efficiently solved by the architecture were: Handling multiple diagnoses in a non-single fault domain. How to represent knowledge genetically thereby making a knowledge base portable. Allowing for more realistic assessments of the total complexity of an implemented system.