Fuzzy Rule Mining And Application Based On Co-evolution

Nowadays, data mining technology has the very important economic and theoretic value, not only in industrial production but also in commercial decision field. However, there are still many unresolved problems, and we are still in the period of creating new concepts and researching new methods. In 1980's, the development of traditional artificial intelligence became slower, while vast advances have been made on computational intelligence (CI), which includes primarily artificial neural network, evolution computing (including genetic algorithm, evolutionary strategy, evolutionary programming, and genetic programming), and fuzzy system. Nowadays, CI has been the mainstream of the intelligence research, and been successfully applied to many fields such as machine learning, process control, economic prediction, engineering optimization. Specially, more and more researchers focus their interest on the combination method on CI's three branches. This thesis mainly focuses on the methods combining evolution computing with fuzzy system. Co-evolution is a new method of evolution computing, which can be used to improve the evolution capability. It parallel evolves two or more species that have the coupled fitness. The individuals of different species could fit others well, and combine to be one complete answer of the problem. The main two parts of the thesis are as follows: 1. Fuzzy system modeling with Co-evolution. Fuzzy reasoning system simulates the way people express and manage information. Genetic fuzzy system, which integrates fuzzy logic and evolutionary computation, has been widely applied in classification and control. Fuzzy modeling process can be clearly divided into two separate and coupled search processes, searching the membership function of fuzzy variables, and searching the classic fuzzy rules. In this thesis, two species are generated for these two processes, a genetic algorithm species and a genetic programming process, and they are optimized simultaneously by co-evolutionary method. For the purpose of improving the evolving efficiency and the results'quality, the significance of the attributes and the tree pruning operator are involved. In most applications, the sample databases are huge, and have many redundant attributes. This thesis preprocesses the sample databases with rough set theory method before the learning process. We evaluate the attributes'significances, and reduce the redundant attributes. The significances are also used to accelerate the evolving process. Code bloat in genetic programming greatly affects the performance of the system. So this thesis introduces a tree pruning operator into GP to limit the code bloat. This operator also can speed up the evolution and improve the intelligibility of the GP result. The system is tested on a medical database. The result of the performance parameters validates the advantage of our method. 2. Co-evolution for fuzzy rule mining in time series. Temporal sequence data are ubiquitous in many fields, and today there has been an increased interest for methods that can extract useful information from large sequence databases. One of the specific problems is rule mining: extracting interesting and unexpected regularities, or rules, from the database. Based on fuzzy theory and co-evolution method, in this thesis a method for fuzzy rule mining based on fuzzy sets and co-evolution is proposed. This thesis proposes fuzzy set concept into the traditional rule language, and defines L a ??T →Lc,CF, a fuzzy rule language with the reliability, CF. It means that if La happens, then Lc happens within the following T time units with the reliability of CF. The symbol set is composed by the fuzzy subsets. The form of the rule language is flexible, and a very wide range of rule formats is supported. Some parameters of the fuzzy rules'performance are proposed, such as rule R's τ-happen, τ-Support(R), Confidence(R), and Bf(R). Then the concept of reliability CF is given as...