A Study of Application Ant Colony Optimization in the Support Vector Machine

• Main Authors: Chen-Yu Xiao, 蕭辰宇
• Other Authors: Yum-Shiow Chen
• Format: Others
• Language: zh-TW
• Published: 2009
• Online Access: http://ndltd.ncl.edu.tw/handle/68696348714882801660

碩士 === 元智大學 === 工業工程與管理學系 === 97 === Recently, support vector machine (SVM), one of the novel techniques for pattern classification, has been widely applied in various fields, such as bioinformatics, text categorization, and so on. However, enormous different in datasets features may increase the difficulty of classification. Furthermore, the subset of features will impact on executive time and accuracy. Thus, a feature selection is an important step in pattern classification problems. A set of selected features is followed by the classification procedure. The purpose of this thesis concerns how to select one of the best subset of features to reduce the error of classification. In this study, we propose a feature selection algorithm based on the Ant Colony Optimization (ACO). The ACO which is a simulator on the behavior of ants in their searching shortest paths to food sources is a metaheuristic algorithm. The ants will leave chemistry called Pheromone on their track. The higher pheromones is aggregated in feature, the more probability the feature will be selected. Following the selection of features, the selected feature-subset is classified and evaluated the error by SVM. This hybrid method is named as ACO-SVM. We apply two real-world datasets which are from the domain of credit risk to verify the proposed hybrid model. The result shows that the proposed method can improve the error efficiently. For studying the influence of classification, we discuss several different rates of training sample and the different sequence of separating data in this study as well. The result displays that the accuracy is positively related to the rate of the training sample, and the accuracy of the post-separated data is better than pre-separated ones. Moreover, adding some suitable rules of local search can auxiliarily diminish the error of classification.