Creating comprehensible regression models

• Mario Drobics
• Johannes Himmelbauer

In this paper we will present a novel approach to data-driven fuzzy modeling whichaims to create highly accurate but also easily comprehensible models. This is achieved by a threestage approach which separates the definition of the underlying fuzzy sets, the learning of theinitial fuzzy model, and finally an local or global optimization of the resulting model. The benefitof this approach is, that it allows to use a language comprising of comprehensible fuzzy predicatesand to incorporate expert knowledge by defining problem specific fuzzy predicates. Furthermore,we achieve highly accurate results by applying an regularized optimization technique.