摘要
为了提高传统AdaBoost(adaptive boosting)算法的收敛性能,提出一种基于多步校正的AdaBoost改进算法。在该算法中,训练样本的分布更新不仅与当前分类器有关,而且也需要考虑到前面的若干分类器;进一步地,新的算法在每一个分类器集成进来后会对前面产生的某些分类器权重进行修正。在UCI数据集Diabets,Heart-statlog和Breastcancer Wisconsin上的实验表明,该算法获得了更好的训练误差和测试误差的优化性能。这说明,利用多步校正策略不但可以提高成员分类器的搜索效率,而且可以进一步地改进集成分类器的整体性能。
The convergence oi the traditional AdaBoost (Adaptive Boosting) algorithm is improved by an AdaBoost algorithm with multi-step correction. In the algorithm, the update of the distribution of the training samples is related not only to the current classifier, but to previous classifiers as well. The algorithm modifies the weights of previously generated classifiers when a new classifier is aggregated. The experiments on the UCI "Diabetes", "Heart statlog", and "Breast cancer Wisconsin" datasets indicate that the modified algorithm achieves better performance in both training and test errors than AdaBoost. The multi-step correction not only enhances the search efficiency for new member classifiers, but further improves the overall performance of the classifier ensemble as well.
出处
《清华大学学报(自然科学版)》
EI
CAS
CSCD
北大核心
2008年第10期1613-1616,共4页
Journal of Tsinghua University(Science and Technology)
基金
国家“八六三”高技术项目(2006AA01Z115)