Comparing Methods for Knowledge-Driven and Data-Driven Fuzzy Modeling: A Case Study in Textile Industry

The aim of this study is to compare different approaches to fuzzy systems design from different perspectives: knowledge-driven versus data-driven and rule-based (flat) versus tree-based (hierarchical). More specifically, our comparison is focused on two of the arguably most important criteria in fuzzy systems design, namely accuracy and interpretability. We compare two approaches to data-driven fuzzy modeling, namely fuzzy rule-based inference using the well-established Takagi-Sugeno approach, and so-called fuzzy pattern trees, an alternative approach that has been proposed only recently. In contrast to the flat structure of fuzzy rule systems, pattern trees are hierarchical models. These methods are compared in the context of a concrete case study, namely the modeling of color yield in polyester high temperature dyeing as a function of disperse dyes concentration, temperature and time. As a baseline, we include Mamdani models designed in a knowledge-driven way. Our results show that, at least in this particular application, Takagi-Sugeno systems offer the best predictive accuracy, whereas Mamdani models are preferable in terms of interpretability. Fuzzy pattern trees seems to offer a good trade-off between both criteria.