Studies Of Computational Intelligence And Their Integrated Applications

Computational Intelligence (CI) is the set of methods. It has the capability of reasoning and learning from the infinite and inaccuracy environment. CI is the powerful computational tool for building more intelligent system. It is widely applied into the field of information processing, management decision, intelligence control, expert system, fault diagnosis, etc. There are three main methods in it: fuzzy system, artificial neural network, and genetic algorithms. During solving the practical computing problems, the methods can be the complements toward one another and work together.In computer-aided design (CAD) for cigarette formulating the main tasks are how to build a math model about chemistry index and internal quality of tobacco group, and how to optimize the restricted cigarette formulating prolem. As it is more complex and uncertain non-linear, so these three computational intelligence methods are integrated in it. At first the dissertation presents the improvements of neural network and genetic algorithm. Then they are integrated into cigarette formulating intelligent CAD system. The whole dissertation consists of eight chapters.Chapter one introduces the basic define of computational intelligence, and illuminates features and cooperation methods of these three techniques. Chapter two illuminates the principal and existing problems of the artificial neural network in brief.Chapter three studies back- propagation (BP) networks in special. According to the principal of BP algorithm, the cause of slowly convergence is analyzed when it is used in complex problem. The convergence performances of many algorithms are compared from three aspects: activation function, weight modification methods, and target function. The main factors are analyzed which influence the generalization capability of BP networks. Then an optimal method of network structure, and the method that initial weights with prior knowledge are presented. According to above, a mixed and improved BP algorithm is presented whose generalization capability is better. At last, with replacing the semi-linear function as the sigmoid activation function, a knowledge extracting from trained BP networks is brought forward. When it is applied into the neuro-fuzzy models of single-material tobacco, some useful rules can be extracted.Chapter four illuminates the principal of Kohonen networks. First an oddityvalue decompose method is presented to select structure of one-dimensional Kohonen network. By researching the relations between the initial weight and topological mapping, a new method of initial weight is presented. The algorithm of two-dimensional Kohonen network is improved from serval aspects such as neighborhood function, learning rate, etc. It is applied into tobacco clustering.Chapter five illuminates a new compound modularization network can be used to build the model for complex problems. The method can improve convergence speed and accuracy. Chapter six introduces the principal of genetic algorithm in brief. Through researching on the optimization problems with the basis of neural network model, a heuristic genetic algorithm is presented.In chapter seven, an integrated method is applied in the cigarette formulating system according to above conclusion. First it introduces the generic term and the traditional flow in the cigarette-formulating field. After comparing several expert systems, a new intelligent system is presented. According to the whole scheme, the integrated course is described step by step. An effective cigarette sensory-quality assessment and smoking measure model is built by using the compound modularization network. Then aiming at the cigarette-formulating maintenance system and new formulating system, different coding methods and mutation operator is presented. Through analyzing the output results of the test sample, it is proved that this algorithm can effectively find optimal answers.In the last chapter of the dissertation, we sum the main productions, and analyze the successful and deficient place of cigarette formulating system. A...