ブースティング法に発想を得た確率モデル学習アルゴリズム

概要

論文の詳細を見る
The methods which use ensembles of classifiers to improve accuracy in supervised learning have received much attention. Especially, much empirical and theoretical work has been made on AdaBoost. Given a base learner and a set of training examples, AdaBoost maintains of training examples and constructs and ensemble of classifiers by applying the base learner to the weighted training set iteratively while updating the weights. The empirical studies have reported that AdaBoost improves accuracy dramatically in many domains. Boosting was proposed and had been analyzed in the context of concept learning (i.e. Boolean-valued function estimation). However, many real classification problems will be intrinsically stochastic and they will not fit into the framework in which AdaBoost has been analyzed. Nevertheless, little attention has been paid for Boosting in the context of probabilistic model learning so far. In this paper, we propose an algorithm which is inspired by the Boosting to learn probabilistic of AdaBoost and our algorithm. We also investigate the relation between our algorithm and AdaBoost. It shows that our algorithm closely related to AdaBoost and they coincide under some restricted settings. These results shed light on an extension of the Boosting to the probabilistic model learning.
社団法人人工知能学会の論文
2000-01-01