Lernen hierarchischer Fuzzy-Regelmodelle

In our modern society and due to the ever-growing flood of information data mining methods are increasingly applied in many areas. However, often the data is unstructured and not manageable which makes it difficult to extract interesting and relevant information. Data mining methods help find coherences in large data sets and make them useful for humans. In order to interpret new knowledge, it is important to generate understandable and simple rules captured in easy models. This research work deals with methods from the field of intelligent data analysis that automatically find patterns as rules in data. Rule-based systems are particularly applied when dependencies in data need to be modeled in a manner that is comprehensible to humans. Due to their simple structure, the rules closely reflect how humans act and think and can therefore be understood and interpreted directly by the user. In this research study, a rule-learning approach that automatically generates fuzzy rule bases has been extended and the influence of different parameters on the accuracy of the trained fuzzy model is examined. Even though these models are generally understandable, they suffer from the fact that usually a huge amount of rules has to be generated to explain all the details of the underlying data. The focus of this thesis is on so-called hierarchical, rule-based models. This type of model hierarchy consists of different layers of simple rule models that describe the concept of the origin of the data in each layer according to a certain degree of detail. In the upper layers only a few rules exist that roughly approximate the concept, while rules further down in the hierarchy concentrate not only on details, but also on artifacts and outliers. A hierarchy of models is usually induced by a local learner, which completely explains the data throughout the hierarchy levels. By using a simple learning algorithm with an understandable hypothesis language, the hierarchy remains interpretable even in the event of complex concepts in the data. These hierarchical models help explore large amounts of data and may also be used