| Our goal in this dissertation is the designing of a novel hybrid system integrating fuzzy logic, neural networks, genetic algorithms and compensatory operations. In order to realize the above goal, we have developed various new techniques by improving conventional methods and discovering novel principles.; At first, we have found that (1) data granularity of conventionally used non-primary fuzzy sets is not appropriate to contain heuristic information for effective fuzzy reasoning; (2) commonly used fuzzy reasoning methods may result in unreasonable behaviors under some circumstances. To solve these problems, we developed a compensatory fuzzy reasoning methodology using new primary fuzzy sets with appropriate data granularity according to the philosophical principles of Taichi and properties of increasing and decreasing functions.; Secondly, A Fuzzy Neural Network with Knowledge Discovery (FNNKD) is designed to perform the compensatory fuzzy reasoning. The FNNKD is more effective than either Takagi-Sugeno's fuzzy system or Jang's ANFIS because: (1) all parameters in the FNNKD have physical meaning and therefore they can heuristically be initialized to speed up the training of the FNNKD based on sample data. (2) The FNNKD can learn commonly used fuzzy IF-THEN rules from given data.; Thirdly, a compensatory genetic fuzzy network using dedicated FNNKDs as basic building blocks is developed to process not only crisp input/output values but also fuzzy input/output sets. Compared with Wang's method, Jang's method and Sugeno-Lang's method, the compensatory genetic fuzzy neural network has impressive abilities of knowledge discovery.; Finally, extensive simulations on a highly nonlinear function approximation, a cart-pole balancing system, a chaotic times series prediction, a gas furnace model identification, compression of a fuzzy rule base and expansion of a sparse fuzzy rule base have strongly indicated that the compensatory genetic fuzzy neural network is an efficient and robust softcomputing system with the ability to discover fuzzy knowledge from both numerical data and fuzzy data and make heuristic fuzzy reasoning based on trained fuzzy rules. |
