SIDIA - Extending prediction based diagnosis to dynamic models

The paper summarizes results of a study about how the 'general diagnostic engine' (GDE, de Kleer/Williams) may be used to diagnose complex, dynamically modelled systems. To deal with complexity we extended GDE to hierarchical model and integrated filters for conflicts and candidates. Diagnosis at different levels in the hierarchy rendered our application tractable. Conflict and candidate filtering assumptions, retractable when necessary, were a good means for efficiently pruning the candidate space in typical situations, without loosing GDE's principal ability to diagnose unexpected faults. Concerning the dynamics in our application, a special extension of episode propagation was developed. An informationtheoretic probe selection procedure for candidates' discrimination was adapted. All extensions to GDE proved easily integratable without touching the basic mechanisms of prediction-based diagnosis. In particular, GDE's clear separation between diagnosis and behavior prediction allowed the straightforward intergration of the new predictive engine. However, our work also showed that the application of GDE to complex and dynamic models requires further elaboration of some of its basic features. In particular, GDE's information theoretic probe selection procedure and the predictive engine should be supported stronger heuristic knowledge.