A Short Tutorial on Model-Based Diagnosis∗

The model-based approach to diagnosis started to be investigated by A.I. researchers in the late seventies, as a possible alternative to the expert-system approach [9]. At that time diagnosis was one of the largest categories of expert systems in use; nonetheless the use of expert knowledge in diagnosis started to show some major limitation [18], especially when applied to diagnosis of artifacts rather than medical diagnosis. The acquiring and maintenance of the needed expert knowledge became a bottleneck in the deployment of diagnostic expert systems. Thus researchers started investigating ([11], [3], [25], [10], [15]) how to exploit, rather than expert knowledge, the then called deep knowledge. The idea was to exploit objective information about the system behaviour (whereas expert knowledge could be regarded as the expert’s subjective view of the system), that could be available for other purposes than diagnosis. In the case of engineered artifacts, this corresponded to exploiting models of the diagnosed system that had been created while designing it. Design models usually adopt a component-oriented approach, that is, they describe a system in terms of its components and the way these components interact. Moreover, since such models are not specifically designed for diagnosis, they do not include knowledge about how the system behaves in the presence of faults. Therefore many of the early works cited above started from the assumption that (i) the model contained information about the sole correct behaviour of the system, and (ii) that the task of diagnosis was to identify which components of the system were broken or mis-functioning. This approach, systematized by Reiter [24], is the one known as consistency-based diagnosis. ∗This is an evolving document, and by no means a complete account of the literature on model-based diagnosis. If you have any comments and/or corrections please contact me at [email protected]