The Development of Learning Mechanisms and Their Applications
| Autor: | Chien-Hsing Chou, 周建興 |
|---|---|
| Rok vydání: | 2003 |
| Druh dokumentu: | 學位論文 ; thesis |
| Popis: | 91 Neural Networks are analytic systems that address problems whose solutions have not been explicitly formulated. The most appealing property of a neural network is the ability of the network to learn from its environment, and to improve its performance through learning. Learning mechanism is referred to as a procedure for modifying the weights and the corresponding parameters of a network to perform some tasks. Generally speaking, learning mechanisms can be roughly divided into three broad categories: (a) supervised learning, (b) unsupervised learning, and (c) reinforcement learning. For the aforementioned learning mechanisms, each learning mechanism has its own considerations, constraints, and limitations. In this dissertation, several new learning algorithms are developed for each learning mechanism and their applications are explored. The first issue of this dissertation is to explore the possibility of applying associative memories for locating frontal views of human faces in complex scenes. Compared with other face detection systems, one of the most appealing properties of this system is that the training task of associative memories can be easily accomplished. Clustering algorithms can be regarded as unsupervised learning algorithms. The second issue of this dissertation focuses on solving two most serious problems in cluster analysis. A modified version of Fuzzy C-means algorithm using the point symmetry distance is first proposed in this dissertation. The proposed modified version of Fuzzy C-means algorithm can deal with clusters with totally different geometrical properties on which many other clustering algorithms cannot perform well. Then a new cluster validity measure which adopts the same idea of "point symmetry” is also proposed in the dissertation. The new validity measure can be applied in estimating the number of clusters of different geometrical structures. In addition to the point-symmetry-based measure, another new validity measure is also proposed to deal with clusters with different densities and/or sizes. The new validity measure can be applied to reduce the edge degradation in vector quantization of image compression. Recently, several different approaches are proposed to apply self-organizing feature maps in cluster analysis. The success of these approaches fully depends on whether feature maps can be topologically ordered. Therefore, a novel measure is first proposed to check whether a feature map is topologically ordered. Then a healing mechanism is proposed to repair a feature map which is not topologically ordered. The last issue of this dissertation is on the topic of reinforcement learning. A new approach to fuzzify classifier systems is proposed and then applied to generate a neuro-fuzzy system. During the training procedure, the neuro-fuzzy system can incrementally construct its architecture and tune the system parameters without the need of desired outputs. |
| Databáze: | Networked Digital Library of Theses & Dissertations |
| Externí odkaz: |
|