Model-Based Diagnosis or Reasoning from First Principles

model that lets us automate the task of locating a detected misbehavior’s cause.1 Although several ways to locate causes exist, they all involve searching for the parts of the system that might have caused the misbehavior. In recent years model-based diagnosis has proven successful in numerous application domains, and the underlying methods are now powerful enough to handle many real-world problems in a commercial setting. Two approaches to model-based diagnosis exist. The consistency-based approach defines a diagnosis as a set of assumptions about a system component’s abnormal behavior such that observations of one component’s misbehavior are consistent with the assumptions that all the other components are acting correctly.2,3 This approach requires only the model of the component’s expected correct behavior. Many publications in the context of AI use the term “model-based diagnosis” to refer to consistency-based diagnosis; we follow that convention in this article. Unlike consistency-based diagnosis, which uses the natural way of reasoning from causes to effects, the abductive approach reasons from effects to causes.4 Also unlike consistency-based diagnosis, abductive diagnosis deals with fault models (models of the system’s faulty behavior). This approach defines a diagnosis as a set of abnormality assumptions that covers (or, in terms of logic, implies) the observations. (Luca Console, Daniele Dupré, and Pietro Torasso analyzed the logical definitions of model-based diagnosis in the literature and proposed a unified framework that describes the relationship between consistency-based and abductive reasoning.5) Without denying the importance of diagnostic reasoning systems based on other techniques, such as causal reasoning and Bayesian approaches, we consider model-based diagnosis as one of AI’s tremendous achievements in recent years.