Fuzzy genetic programming - offline acquisition of fuzzy rules using variable-length GAs

G. Steinöcker. Fuzzy genetic programming - offline acquisition of fuzzy rules using variable-length GAs. 2002.

  • Gerrit Steinöcker
Abstract In order to be able to extract functional relationships from data one can try to develop anexact mathematical model for the problem, e.g. in physics, or create an approximation, e.g. using statistics or numerics. In the fields of medicine, economics and industries, however, there are still many tasks that cannot be solved satisfying with any of these methods. Therefore, the search for alternative approaches is an important goal. One possible approach is to use machine learning. In particular this can be done by imitating the human learning style. Genetic algorithms are one implementation form of the simulation of human learning behavior. Still, not all problems can be solved with these algorithms since they cannot be described in an appropriate way. Some problemscan only be described or solved properly with fuzzy systems. Therefore, a synergy between fuzzy systems and genetic algorithms, so-called fuzzy genetic programming, is needed.Firstly, we described the requirements and organized them as follows: we start with giving a brief overview on fuzzy systems. Then an introduction to genetic algorithms is presented, which allows a deeper insight of the working of genetic algorithms. Using genetic algorithms we motivate genetic programming and combine it with fuzzy systems to fuzzy genetic programming.Secondly, we developed a testing environment for FGP which was used for applying the theory. This environment is coded in C++ and is based on the libraries of Helmut Hörner and LibFuzzy++ which are described in detail. This is followed by a description of the additional work done, to be able to deal with fuzzy systems and fuzzy rules in special.Finally, we show that fuzzy genetic programming is not only capable to deal with easy problems like the minimum or maximum problem but also with distorted data and rather complicated cases like the UCI wine data set.