Learning Qualitative Models From Example Behaviours

Qualitative models provide the domain knowledge for qualitative reasoning systems, and the automatic generation of such models is an important knowledge acquisition task . Qualitative Simulation is a key qualitativ e reasoning technique . It is proposed that simulation systems like Kuipers ' QSIM [Kuipers 86] contain a powerful description language, based as they are on qualitativ e mathematics, for the automation of model capture . A simulation system uses qualitative models and mathematical knowledge to generate behaviours . The Genmodel program described here uses the same knowledg e as a description language to proceed in the opposite direction, from behaviours to models . It is assumed that a set of dynamic qualitative syste m behaviours are available to form a training set of positive examples . These behaviours are either presented b y a domain expert or are extracted from numerical data b y curve fitting or other techniques. The program's task is to search for qualitative relationships between function s that will permit the example behaviour . These relationships together form a model of the system that produced the behaviours .