報告書番号

ER99001

タイトル（和文）

Trends in Boosting Research and Applications

タイトル（英文）

Trends in Boosting Research and Applications

概要 （図表や脚注は「報告書全文」に掲載しております）

観測されたデータ中の代表データに注目し分類する従来方法とは異なり、分類の難しいデータに注目するブースティングアルゴリズムについて調査し、その学習特性を分類問題について解明したし、現在のブースティング改良研究の流れを報告した。また、この方法が適用されている応用例を調査し、ブースティングの潜在能力の高さを明らかにした。今後ブースティングの研究動向の把握、適用研究情報の収集が重要であることを指摘し、ブースティング研究に関する代表的なウェッブサイトの開設、運営が情報収集、適用促進に重要となることを報告した。

概要 （英文）

This report analyzes the research trend of boosting methods from a practical point of view and presents some applications based on these methods. We investigate in the following three features as results of the analysis. 1. The original AdaBoost, which was proposed by Freund and Schapire, is a hard margin classifier like a Support Vector Machine. AdaBoost is very useful and can achieve good generalization performance in the low noise regime (few outliers, low input noise). 2. In the noisy case (lots of outliers, high input noise), the generalization performance of AdaBoost becomes less comparative to other algorithms: AdaBoost concentrates too much to some outliers while generating a hard margin. 3. Now, some improved boosting methods are researched to avoid overfitting by using a soft margin concept. In this report, AdaBoost$_{reg}$ and $\nu$-Arc are introduced as examples of improved boosting methods that can deal with noisy data and avoid overfitting. AdaBoost$_{reg}$ is briefly introduced as the first improved boosting method. Moreover, $\nu$-Arc is presented which has connections to the soft margin approach of SVMs. This report presents the following applications based on Boosting: ・ Empirical evaluations of boosting methods on benchmark data sets・ Application A: Handwritten digit recognition・ Application B: A non-intrusive monitoring system for household electric appliancesThe boosting methods are one of the new learning strategies. There are only a few applications based on these methods yet. However, boosting methods have high potential for many applicationfields and it is undoubted that the number of applications based on the boosting methods will be increasing as in the case of Support Vector Machines. Support Vector Machine community made a great success for collecting and spreading research information by building Support Vector Machine homepage. Therefore, in order to collect and spread information about new strategy of boosting, we will arrange and operate a representative web-site about boosting.

報告者

キーワード