Automated Determination of Qualitative Distinctions: Theoretical Foundations and Practical Results

Automating the generation of qualitative models at a level that is tailored to support a particular task is crucial to the deployment of model-based systems technologies in practical applications, because reusability of models is of vital importance. If this task cannot be solved, models in a library will either be ineffective, because they are too coarse for solving a particular problem, or inefficient, because they are too fine-grained. The key question to be answered is, "what are the distinctions in the domains of the system variables that are both necessary and sufficient to achieve a particular goal in a certain context and under given conditions?". In our approach, the goal is defined by a set of target partitions of the domains of selected variables (e.g. output variables), the context is given by the structure of the modeled system, and the conditions are represented by a set of initial variables and their possible distinctions (e.g. possible observations). The task includes problems such as determining the appropriate qualitative values of variables in order to enable prediction at the desired level or discrimination for diagnosis, deciding whether or not changes can be modeled as discontinuous ones, and determining when a deviation of a parameter can be considered significant. We have analyzed and formalized the task for relational behavior models, implemented an (incomplete) algorithm, and carried out first experiments. The paper first elaborates on previous theoretical foundations, defining the goal and the specification of algorithmic solutions. We then outline the implemented algorithm and present and discuss some experimental results of applying this prototype. We conclude with some open problems and discuss alternative approaches.