摘要
智能优化的嵌入式特征选择算法相比于传统嵌入式方法和过滤式方法,通常能得到规模更小和质量更高的特征子集,但智能优化中的群体寻优策略和嵌入式方法的评价机制导致其计算代价过高。为解决上述问题,提出了一种基于代理模型和人工免疫系统的特征选择算法。利用RBF神经网络构建代理模型,对部分新产生的个体进行预测和评价,避免了频繁调用评价准则而导致的时间损耗。在训练神经网络之前,利用相关系数概念将所有特征进行归类,以避免输入节点过多而影响模型的预测质量。此外,提出的人工免疫系统设计了3种不同的变异算子用以进一步提高算法的求解质量。对5组UCI数据的测试结果表明,该算法能够在保证求解质量的同时显著减少优化时间,其优化时间相对于无代理模型的同类算法最多减少了44.94%。
Wrapper-based method utilizing intelligence optimization worked better than traditional wrapper-based and filter-based methods since it could obtain smaller and more effective feature subset.However,it was very time-consuming for its populationbased search mechanism and learning procedure for estimating a solution.A novel feature selection algorithm based on surrogate model and artificial immune system was proposed to solve the problem.In this algorithm,RBF neural network was used as the surrogate model to save the time consumed on the learning procedure.A classification method based on the notion of correlation coefficient was designed to cluster the features so that the number of nodes in input layer was controlled.Besides,three mutation operators were designed to enhance the artificial immune system.Experiments on five UCI datasets showed that the proposed algorithm could obviously reduce the optimization time and obtain good results simultaneously.The time was saved at most by 44.94％ of it obtained by the algorithm without surrogate model.
出处
《计算机工程与设计》
CSCD
北大核心
2014年第6期2174-2178,共5页
Computer Engineering and Design
基金
国家青年基金项目(61301232)
河南省教育厅自然科学研究重点基金项目(12A520013)
关键词
特征选择
代理模型
人工免疫系统
RBF神经网络
群体寻优
变异算子
feature selection
surrogate model
artificial immune system
RBF neural network
group optimization
mutation operator
