摘要
竞争学习在聚类分析中是一种重要的学习方式,次胜者惩罚竞争学习(RPCL)算法虽能自动选择合理的类别数,但其性能对学习率和惩罚率的取值较敏感,其变种惩罚控制竞争学习(RPCCL)算法将所有的竞争单元当成冗余单元进行惩罚也不合理.文中提出一种可区分惩罚控制竞争学习算法(DRPCCL).算法中获胜单元的学习率会在迭代过程中自适应调整.同时该算法使用一种可区分惩罚控制机制来区分竞争单元中的冗余单元和正确单元,给予冗余单元较重惩罚,正确单元轻微惩罚,使得算法能自动确定正确类别数和中心点位置.最后通过实验对比分析证明DRPCCL算法的聚类效果比RPCL算法和RPCCL算法更准确.
Competitive learning is an important approach for clustering analysis. The rival penalized competitive learning ( RPCL) algorithm has the ability of selecting the correct number of clusters automatically, but its performance is sensitive to the selection of learning rate and de-learning rate. In fact, it is unreasonable that all the rival units are treated as redundant units to be penalized in the variant algorithm called rival penalization controlled competitive learning ( RPCCL) . In this paper, a discriminative rival penalization controlled competitive learning ( DRPCCL) is presented. The learning rate of winning units adaptively adjusts during iteration in the proposed method. Meanwhile, a discriminative penalization controlled mechanism is used to discriminate the redundant units and the correct units in the rival units. The correct units and redundant units are given a slight penalization and a heavier penalization respectively, which makes this algorithm get exact number of clusters and reasonable centre of clusters. The experimental result demonstrates that compared with RPCL and RPCCL, DRPCCL achieves more accurate performance.
出处
《模式识别与人工智能》
EI
CSCD
北大核心
2014年第5期426-434,共9页
Pattern Recognition and Artificial Intelligence
基金
国家自然科学基金项目(No.61175026)
国家"十二五"科技支撑计划项目(No.2012BAF12B11)
科技部国际科技合作专项项目(No.2013DFG12810)
浙江省自然科学基金重大项目(No.D1080807)
浙江省国际科技合作专项项目(No.2013C24027)资助
关键词
聚类分析
竞争
胜者惩罚竞争学习(RPCL)
可区分的惩罚控制机制
Cluster Analysis, Competitive Learning, Rival Penalized Competitive Learning ( RPCL),Discriminative Penalization Controlled Mechanism