The Key Technology Research, Genetic Fuzzy Classification System Build In Rules To Obtain And Explanatory Optimization

Learning and optimizing fuzzy rule-based classification system by genetic algorithms is an important branch of pattern recognition, it posses important value of both in theoretical research and practical application because of its superb ability for tacking imprecise and uncertain information. Despite some successful designs and applications, there still remain difficult challenges that need to be addressed, including: 1. the preciseness problem caused by the search efficiency of genetic algorism when handing complex data; 2. the redundancy problem caused by the complexity of the set of classification rules, and the system is difficult to be understand by users with the regard to interpretability. Those two major challenges directly limit how this kind of system can be put into practices. The purpose of this thesis is to address two important challenges, and the major contributions are demonstating on the following three perspectives:(1) Linguistic modeling methods are adopted according to the concept of computing with words. The membership functions of fuzzy rules are consisted by five atomic words and four hedges that are transformed by similarity measure. Therefore, fuzzy rules can be expressed in an understandable way. On this basis, a method to include expert knowledge is integrated into the design of fuzzy system. For the reasons of expert knowledge are usually inadequate for the system, the genetic learning approaches are proposed for automatically generating accuracy fuzzy rules, with their coding, fitness function, genetic operators and special strategies are descried in details.(2) Because genetic algorithms are frequently difficult to perform global searching for high-dimensional and imbalance data, the niching technique, sharing and crowding are introduced in the fuzzy rules learning process for obtaining fine features or classification boundary. By analyzing the search ability of classical genetic machine learning approaches, the similarity level of one fuzzy rule from its neighbour's rules is defined, with the similarity level, fitness sharing and deterministic crowding are used for reducing the selection pressure of the individuals with low fitness values, and maintain the diversity of population. Therefore, the genetic learning processes are able to search all subspaces of these complex problems efficiently. Moreover, the similarity value of different fuzzy sets are calculated and cached before the learning processes for cutting down the computing load of similarity level. Experiments of different approaches are carried out on a suite of test problems and some well-known classification problems, and the results show that the sharing and crowding methods have higher search ability and they are able to obtained accurate fuzzy classification rules sets.(3) The existing measures of interpretability are reviewed, and the implementations of these measures are analyzed in different stages of system modeling. Because appropriate number of fuzzy rules and number of conditions should be taken into account for decreasing the complexity of linguistic systems, the multiobjective similarity-driven simplification method and NSGA-â…¡are introduced. The simplification method is designed for optimizing the existed fuzzy rules set, convert multi-objective to single-objective by using aggregating function; the NSGA-â…¡is based on Pareto theory and is able to search for a number of non-dominated fuzzy rules sets with different tradeoffs between accuracy and interpretability. The experiments investigate the major strength and weakness of the above approaches, C4.5 and GP-COACH.