分裂-合并竞争学习的研究

针对竞争学习在给定的输出节点数目少于实际类数目时的学习结果会在几类数据之间振荡的问题,提出了MPTOC策略以及基于此策略的分裂一合并竞争学习算法．在假设数据集中的数据对其相应节点产生大小等于二者距离“吸引力”的基础上,算法通过计算网络中获胜节点在不同方向的“吸引力合力”分布,间接描述该节点附近数据的分布情况;采用高维空间模糊熵的方法确定该节点主要的“合力”方向,并将该节点在这几个方向上进行分裂一合并学习,从而实现MPTOC策略．通过对二维随机分布数据的实验结果验证了所提出算法的正确性和有效性．

Research on Splitting-merging Competitive Learning

An MPTOC strategy is presented,which guarantees each cluster has an output unit at least during learning.The concept of "attractive force" in mechanics is adopted to describe the relation between a unit and its corresponding data,which equal to their Euclid distance.Based on MPTOC splitting-merging competitive learning the distribution of data around their winning unit is estimatea indirectly through computing the unit attractive forces.And the unit splitting directions are assigned through method of fuzzy entropy in high dimension space.Then the unit is split and learned along these directions.To avoid over-segmentation of input dataset,the results of splitting-learning are merged with the help of their means and variances.Experiments in 2D space validate the proposed algorithm.