一种基于累积适应度遗传算法的SVM多分类决策树  

SVM decision-tree multi-classification strategy based on genetic algorithm with cumulative fitness

作　　者：朱庆生[1] 程柯[1] 

机构地区：[1]重庆大学计算机学院,重庆400030

出　　处：《计算机应用研究》2016年第1期64-67,74,共5页Application Research of Computers

基　　金：国家自然科学基金资助项目(61272194)

摘　　要：针对基于遗传算法(genetic algorithm,GA)的支持向量机(support vector machine,SVM)多分类决策树算法(GA-SVM)中全局优化缺陷的问题,通过重新定义遗传适应度函数(fitness),提出一种累积适应度(cumulative fitness),进而衍生出新算法CFGA-SVM。该算法从根节点开始逐层构造二叉树,对根节点基因实值编码,通过基因分裂操作产生子代种群,然后利用累积适应度筛选出新的种群,筛选出的种群并不一定是当代局部最优,但一定是所得二叉树中全局最优,从而提高分类精度,最后以此循环直至算法结束。通过在UCI的artificial characters数据集上的实验结果表明,CFGA-SVM较之DT-SVM与GA-SVM算法在全局优化能力、分类精度上有明显提高,进而验证了该算法的可行性与有效性,可在大规模样本的分类应用中推广。Aiming at the defect in global optimization of the SVM decision tree based on genetic algorithm （ GA-SVM）, this paper redefined the fitness function, which was the key component in genetic algorithm, and proposed an optimized cumulative fitness and a new SVM decision tree based on cumulative fitness genetic algorithm （CFGA-SVM）. This algorithm took advan- tage of global optimization by the cumulative fitness in the gene selection phase of GA, thus brought a more appropriate popula- tion into the following inheritance operation. Experimental results on the UCI artificial characters dataset verify that CFGA- SVM is better than the traditional DT-SVM and GA-SVM in the aspect of classification accuracy and global optimization meas- urement. It has wide application prospects especially with huge training sample.

关 键 词：多分类 支持向量机 遗传算法 累积适应度函数 全局优化 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]