| Summary: | 碩士 === 國立中央大學 === 電機工程研究所 === 96 === In typical pattern recognition applications, there are usually only some vague and general knowledge about the situation. An optimal classifier will be definitely hard to develop if the decision function lacks sufficient knowledge. The aim of our experiments is to extract some features by some appropriate transformation of the training data set. In this thesis, we assume that the training samples are drawn from a Gaussian distribution. We also assume that if the data sets are in an imprecise situation, such as classes overlap. The overlap can be represented by fuzzy sets. Therefore, a fuzzy membership can be created according to the property of class overlap. For example, one can treat the closer training data of decision boundary as Support Vectors (SVs) in the center of classes overlap and let these points have higher degree of the fuzzy membership. That is because these points have higher contribution to the decision boundary. Relatively, one can treat the father training data of the decision boundary as SVs outside the margin and let these points have lower degree of fuzzy membership. In Support Vector Machines (SVMs), we define a fuzzy-penalizing parameter to balance both margin width and model complexity.
Finally, a powerful learning classifier is shown. It is the Fuzzy Support Vector Machines with the Uncertainty of Parameter C rule (FSVMs-UPC). In order to verify this classifier, the proposed method is compared with traditional SVM in experiment 1. Results show that the proposed FSVMs-UPC is superior to the traditional SVM in terms of both testing accuracy rate and stability. Experiment 2 shows our membership generation method concentrate on overlapping is a more feasible and better membership.
|
|---|
