ADS-1-10 Acceleration of the Adaboost with Exact Plane Search

概要

論文の詳細を見る
The Adaboost is so far known as one of the most successful boosting algorithms that can produce a highly accurate classification rule by combining multiple relatively low accurate rules. The algorithm is interpreted as a certain coordinate descent method for iterative minimization of an exponential-type function. Unfortunately due to the simple line search strategy of the algorithm, its speed of minimization is not necessarily efficient. In this preliminary report, as a candidate of more efficient boosting algorithms, we propose a novel iterative minimization algorithm which employs an exact plane search in place of the line search in the Adaboost. The exact plane search is realized simply by an explicit formula. A numerical example demonstrates that the proposed algorithm successfully accelerates the speed of minimization while its computational complexity comparable with that for the Adaboost.
社団法人電子情報通信学会の論文
2009-03-04