一种基于生物地理学优化算法的集成选择方法  

Ensemble Selection Method Based on Biogeography-based Optimization Algorithm

作　　者：丁智国[1,2] 费敏锐[1] 马海平[1] 

机构地区：[1]上海大学机电工程与自动化学院上海市电站自动化技术重点实验室,上海200072 [2]浙江师范大学数理与信息工程学院,金华321004

出　　处：《系统仿真学报》2014年第5期996-999,共4页Journal of System Simulation

基　　金：国家自然科学基金(61074032);上海市科委重点基础项目(10JC1405000)

摘　　要：集成选择发生在集成结论合成之前,所有的个体学习器已经生成后对某些不满足精度要求或多样性要求的个体进行剔除,从而用较少的个体集成获得同原始集成相同,甚至更好的泛化性能,同时降低系统的存储和计算开销。基于一种新的优化算法——生物地理学优化算法(Biogeography-based optimization,BBO),对获得的初始集成个体依据个体的预测精度和个体之间的多样性准则设置目标函数,进行优化选择,提出了一种新的基于BBO的集成选择算法(BBO based Ensemble Selection Algorithm,BBOESA)。基于UCI数据的仿真结果表明,该算法在不降低,甚至提高了常规的Bagging集成学习的泛化性能的同时,使集成个体数量有了大幅度的减少。Ensemble selection is defined for the setting where the base learners have already been generated, and no more base learners will be generated from the training data. In the ensemble process, only some base learners who meet the standard of accuracy and diversity in some extent, are employed in order to achieve the better, at least equal generalization performance than including all of them, as well as to reduce the memory demand and computation cost. Based on the biogeography optimization algorithm, a new ensemble selection algorithm, BBOESA, was proposed to select the optimized base learners for the final ensemble. Experiment results on the UCI benchmark datasets demonstrate that the proposed algorithm has a better generalization performance than Bagging＇s and the number of base learner has decreased dramatically.

关 键 词：集成学习 集成选择 生物地理学优化算法 泛化性能 

分 类 号：TP273[自动化与计算机技术—检测技术与自动化装置]