Abstract

This thesis work focuses on one half of the knowledge acquisition problem for fuzzy systems, namely the acquisition of a fuzzy rule set from a set of input-output data. This acquisition is commonly done by the so called fuzzy rule extraction methods. An example of an existing method for fuzzy rule extraction is the method of Ishibuchi. This method suffers from a combinatorial explosion and relies heavily on the generalizing power of a trained artificial neural network.
A novel algorithm is presented, which is called the genetic programming (GP) rule extractor. It makes use of Darwinian evolution to find rule sets which are suited to problem domain data sets. To test its quality a generic applicable method called the research cycle has been devised which is able to test the quality of any rule extractor. The research cycle enables a user to create a reference framework for the expected answer of a knowledge state changer such as a rule extractor. The implementation and integration of software representing the research cycle resulted in a workbench. This workbench was able to fulfill the demands of the tasks of the research cycle. It produced results on which conclusions can be drawn about the characteristics of the genetic programming rule extractor and the Ishibuchi rule extractor.
Unlike previous work in the area of rule extraction, the introduced genetic programming extractor is general in its applicability and scales well to problems with high-dimensional input spaces.