并行免疫克隆特征选择算法

Parallel immune clonal selection for feature selection

针对模式识别中传统的封装式特征选择算法,难以得到较好的特征子集和复杂度较高的分类器评价特征子集的耗时问题,提出了一种用于特征选择的并行免疫克隆算法,采用免疫克隆算法搜索特征,并利用并行算法评价特征子集,即将种群中个体的适应度计算并行在多个计算节点上同时进行.将该算法在Linux刀片集群上基于MPICH软件对UCI数据集进行特征子集选择算法仿真,特征子集采用最近邻分类并采用留一法验证评价.结果表明该算法选出的特征子集优于经典的顺序浮动前向搜索算法和标准遗传算法,与串行算法运行时间相比,在40个CPU时其加速比最高可达29.57.

Focusing on the time-consuming problem of wrapper feature selection when the feature subset is evaluated using high-complexity classifiers in pattern recognition, a novel parallel immune clonal selection for feature selection algorithm （PICFS） is proposed. The presented method uses an immune clonal selection for feature selection~ fitness of feature subset fitness is determined by evaluating the nearest neighbor classifier with leave-one-out cross-validation in multiple computing nodes at the same time. Experimental results on several standards UCI dataset sets show that the proposed algorithm outperforms the conventional genetic algorithm and classical sequential floating forward search algorithm in terms of classification accuracy and greatly reduce the running time based on MPICH using the Linux blade cluster, we have achieved a speed-up as high as 29.57 even when up to 40 processors are used.