Qualitative system identification: deriving structure from behavior

Qualitative reasoning programs (which perform simulation, comparative analysis, data interpretation, etc.) either take the model of the physical system to be considered as input, or compose it using a library of model fragments and input information about how to combine them. System identification is the task of creating models of systems, using data about their behaviors. We present the qualitative system identification algorithm QSI, which takes as input a set of qualitative behaviors of a physical system, and produces as output a constraint model of the system. QSI's output is guaranteed to produce its input when simulated. Furthermore, the QSI-made models usually contain meaningful “deep” parameters of the system which do not appear in the input behaviors. Various aspects of QSI and its applicability to diagnosis, as well as the model fragment formulation problem, are discussed.