Generating fuzzy models from deep knowledge: Robustness and interpretability issues

The most problematic and challenging issues in fuzzy modeling of nonlinear system dynamics deal with robustness and interpretability. Traditional data-driven approaches, especially when the data set is not adequate, may lead to a model that results to be either unable to reproduce the system dynamics or numerically unstable or unintelligible. This paper demonstrates that Qualitative Reasoning plays a crucial role to significantly improve both robustness and interpretability. In the modeling framework we propose both fuzzy partition of input-output variables and the fuzzy rule base are built on the available deep knowledge represented through qualitative models. This leads to a clear and neat model structure that does describe the system dynamics, and the parameters of which have a physically significant meaning. Moreover, it allows us to properly constrain the parameter optimization problem, with a consequent gain in numerical stability. The obtained substantial improvement of model robustness and interpretability in "actual" physical terms lays the groundwork for new application perspectives of fuzzy models.