Research Of Neural Network And The Ensemble Method Based On Granular Computing

Machine learning based on the data is an important content in the data mining technology, it mainly studies observed data to look for rules, and uses these rules to predict future data that can’t be observed. Granular computing(Gr C) is an effective approach to solve complex problems, by dividing the complex data, it transfers the complex problems to a number of relatively simple problems, so as to help us to analysis and solve complex problems. The basic idea of Gr C is to use simple easy enough satisfied approximate solution to alternate exact solution. Gr C has become the research hot spot in the field of artificial intelligence, soft computing and control science. All the theories and methods of grouping, classification and clustering used for analyzing and solving the problems are belong to the scope of Gr C.The artificial neural network( ANN) is another classic method of data mining,it is one kind of network system by simulating human brain information processing mechanism based on the development of modern biology research, which is also one of the soft computing techniques. As ANN has distributed storage of information, parallel processing, self-adapted and self-learning ability and other advantages, it has inestimable value in the fields of information processing, pattern recognition and intelligent control. However,neural network ensemble(NNE) can achieve the effect of improving network generalization ability by training multiple neural networks and combining their conclusions of multiple individual neural networks.Gr C theories and neural network have their own advantages in dealing with problems, and both of them have certain supplementary in some degree. The combination between them can solve the complex and high-dimensional problems well. The dissertation mainly studied Gr C based neural networks and neural network ensemble methods and proposed several network models and the corresponding algorithms.The main works of this dissertation included the following aspects:1. The dissertation studied the fusion of covering algorithm and Affinity Propagation(AP) algorithm, proposed a novel neural network classification model based on covering and AP clustering algorithm. AP algorithm, which does not need to define the class number, constantly searches suitable clustering center in the iterative process, and automatically identifies the center and its position from the data point. AP algorithm is a kind of deterministic clustering algorithm, and the clustering results are generally stable under multiple independent operations. Covering algorithm has strong comprehensibility, fast computing speed, high recognition rate, but it randomly selects training sample set from the original data set, the learning order in the covering algorithm directly affects the size of the covering areas and number, and greatly influences the effect of learning. Using AP clustering method as a pre-processor to cluster high dimensional massive data samples, and combining covering algorithm to automatically determine the center of the covering, and calculating its radius, and then converting these parameters into the weights and thresholds of the hidden layer of classification neural network. The number of hidden layer neurons is the number of covering. The output layer is the M-P neuron layer. By evaluating the input feature of hidden layer to complete the mapping from input to output. The introduction of AP clustering algorithm solves the problem of covering algorithm randomly selecting initial field center.2. The dissertation studied individual neural network generation and ensemble algorithm based on quotient space granularity clustering. Among all the parameters of AP clustering, the most important is the preference(P), which directly influences the number of clustering. Introducing the quotient space concept to the AP clustering analysis, which can find an optimal granularity from all possible granularities. Based on this, using different individual neural network to learn different categories of samples so that the diversity between networks can be improved. Making the number of classes and individual Neural networks is equal to determine the structure of NNE. Further, according to the degree of correlation between the input data and the sample category to adaptively adjust ensemble weights to improve the precision of NNE.3. The dissertation studied twice clustering based individual neural network generation method. The diversity between individual networks is essential to the generalization performance of NNE. In order to improve the precision and the diversity of individual networks to improve the performance of NNE, by varying the network training data that enables the samples to reflect the real data distribution, increasing the diversity between the training data to increase the diversity between the networks, thus improving the performance of NNE. Firstly, using k-fold cross validation method to divide the original dataset. k-fold cross validation, which learning samples from multiple directions, can effectively avoid falling into local minimum value. Because both the training and validation samples are as far as possible to participate in the learning, one can get satisfactory effect of learning. Secondly, choosing all the training samples to cluster for the first time to form once clustering subclasses, and then performing the twice clustering for each subclass to form the sample subsets of each subclass. Through AP clustering makes the otherness criterion of “similar in classes, different between classes” maximize, the samples in the class can reflect the real data distribution. Finally, according to the permutation and combination, selecting a subset from each twice clustering of each subclass to construct a training set. So the individual neural networks are generated with the bigger diversity, the smaller size of the training data and the training data can also reflect the real data distribution, and the ensemble of these individual neural networks can get better performance. Simulation experiments on nine datasets show that our proposed method here is effective.4. The dissertation studied the multi-side multi-granular neural network ensemble optimization method. According to the thought of divide and conquer that human perceiving complicated things from multi-side and multi-view and balancing the final decision. Combining feature selection, dividing attribute granularity of dataset from multi-side, and structuring multi-granular individual neural networks using different attribute granularity and the corresponding subsets. In this way, one can gain multi-granular individual neural networks with greater diversity, and get better performance of NNE.The dissertation studied several neural network and neural network ensemble models based on Gr C and their learing algorithms, and verified the effectiveness of these models by experiments.