Researches On The Performance Of Genetic Algorithms And Their Applications In Clustering Analysis

Genetic algorithms (GAs) are random search techniques loosely based on the principles of nature evolution, which have become increasingly popular in recent years as a method for solving complex optimizing problems. The appeal of GAs comes from their simplicity and elegance as algorithms as well as from their power to find optimal solutions rapidly for difficult high-dimensional problems. What relation exists between GAs performance and fitness functions, namely, the solution for what kinds of functions is easy to find by using GAs, has aroused a great deal of interest among numerous researchers. Its solution is of great significance as people have tried to apply GAs to ever more diverse types of problems. In this dissertation, further study is conducted by constructing some fitness functions with expected order and definite length using Walsh polynomials. Detailed analyses of GAs performance and the landscapes of these functions result in a more reasonable explanation. On condition that population size is finite, random competition for survival will arise among different individuals with same or approximate fitness. The phenomenon is studied in detail and a solution is proposed to improve GAs performance by increasing population size slightly. A hybrid search method named BP network based genetic algorithm is also presented, which employs BP neural network to learn the corresponding relations between individuals. Then, the relations are used to direct the evolution of the individuals. In order to use relations among samples to speed up the search process, the relations are studied and an adaptive clustering algorithm for non-overlapped samples is introduced, which is insensitive to their distributions. Furthermore, to categorize samples on different curves, three algorithms are proposed which are parametric curve detection algorithm based on GAs, adaptive clustering algorithm for linear distributed samples, and connecting and fitting of curve segments based on GAs. A fundamental framework for classification system is studied in the last chapter, which introduces new ideas for the design of complex classification systems.