Abstract:
The main contribution of this research in the qualitative reasoning area of Artificial Intelligence is the development of the qualitative system identification algorithm QSI. QSI's input is a description of the qualitative behaviors of the system to be identified. Its output is a constraint model (possibly containing "deep" parameters absent in the input) of that system, in the format of Kuipers' qualitative simulation algorithm QSIM. The QSI approach to qualitative modeling makes no assumptions and requires no knowledge about the "meanings" of the system parameters. QSI is discussed in detail. Other contributions are a new method of eliminating a class of spurious QSIM predictions, and an algorithm for postdiction. Unlike other approaches to spurious behavior reduction, the method presented here does not require restricting assumptions about the input model. A particular kind of spurious behavior is shown to be caused by pure QSIM's insistence on assigning only point values to "corresponding value tuples" associated with model constraints. The solution put forward here preserves the overall complexity of the algorithm, while producing fewer incorrect predictions, as shown by the presented reports of the case runs and proofs. Postdiction is the task of finding out the possible pasts of the system under consideration, given the laws of change and the current state. For obtaining the algorithm, a different scheme of interpreting the tree built by simulation is imposed, as well as the handling of the "flow" of time. Issues of thig reasoning task, which is promising for diagnosis applications, are discussed.
