一种改进的AdaBoost快速训练算法  

An Improved Ada Boost Training Algorithm

作　　者：张苗燕 王登飞 魏宗寿[1,2] 

机构地区：[1]兰州交通大学自动控制研究所,甘肃兰州730070 [2]甘肃省高原交通信息工程及控制重点实验室,甘肃兰州730070

出　　处：《西北工业大学学报》2017年第6期1119-1124,共6页Journal of Northwestern Polytechnical University

基　　金：甘肃省基础研究创新群体计划(1606RJIA327);陇原青年创新人才扶持计划(2016-38);甘肃省科技支撑计划(1604GKCA009);兰州交通大学校青年基金(2016024)资助

摘　　要：针对AdaBoost算法误检率及收敛速度问题,结合改进的细菌觅食优化算法的思想,提出一种基于改进细菌觅食的AdaBoost弱分类器优化权重算法。采用改进的随机化佳点集方法构造初始种群,改进的趋化策略、变次数游动策略及变概率迁徙策略来全局优化搜索最佳弱分类器。对最佳弱分类器的加权系数作以改进,其加权系数不仅与错误率有关,也应与对正样本的识别能力及弱分类器的可靠性有关。选取4种UCI数据集进行实验验证,基于Matlab的仿真结果表明,改进方法获得了较好的检测性能。Aimed at the problem of mis-decetion rate and the convergence speed, and combined with improved Bacterial foraging optimization algorithm, this paper presented an improved AdaBoost algorithm named optimal weighting algorithm of weak classifiers based improved BF-based AdaBoost. In this paper, adopted an reformative good point set based randomization method to construct the initial population, and used some strategies such as improved chemotaxis direction policies, variable frequency winding tactics and changing probability of migration operations to soulord the weak classifiers. In order to modify the weight coefficients of optimal weak classifiers, the weighting coefficient was not only related to the error rates, but also the recognition of positive samples and the reliability of classifiers. Experiment results of simulation by selecting four UCI data sets based on MATLAB indicate the improved method obtains better detection performance than traditional AdaBoost algorithm.

关 键 词：ADABOOST 细菌觅食优化算法 随机化佳点集 弱分类器 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]