【摘要】 聚类分析是把ｎ个个体分成ｋ个内在相近类群的一种多元统计分析方法。非系统（又称动态）聚类一般能得到比系统聚类更为合理的结果，但稳定性差的问题非常突出。以目标函数为最小迹［Ｍｉｎｔｒ（Ｗ）］为条件的动态聚类全局最优解的算法分以下三步。第一步：对个体依次易组试分，若这种试分能优化目标函数，则固化试分。进行一至多轮的试分改组，直至任一个体的改组均不能改善目标函数时为止，记录下目标函数值及相应的分类结果。第二步：设定一临界正值Ｃ＿ｉ，当试分改组增大目标函数，但又不超过Ｃ＿ｉ时仍实施改组，该轮试分过程一般会使目标函数劣化，但应对可能出现的目标函数最小化植和相应的聚类结果作出记录。对所有个体试分后，改变（降低）Ｃ＿ｉ值：Ｃ＿（ｉ＋１）＝αＣ＿ｉ（０＜α＜１）。以上两步交互运算多次。第三步：类群的重组过程，合并组中心欧氏距离最近的两类，并把平方乘积和阵迹tｒ（Ｗ＿ｉ）量大的类群一分为二以保持总组数ｋ不变。重复以上过程多次，在一定轮次（１０～１２）内目标函数未有改善时结束寻优过程。经多组模拟和实用数据运算，该算法对一般聚类分析问题都能达到全局最优解。更多还原

【Abstract】 CIuster analysis is to partition n objects into k non-empty coherent clusters,There are wide ranges ofapplications in genetics and other areas.Non-hlerarchical clustering usually produces local instead of global optimal re-sults in most cluster analyses.The poor robustness is the main problem of existing methods.A new algorithm wasproposed in this paper to reach the global optimization.The algorithm begins with obtaining a trial transfer of an ob-ject between initial（or intermediate） groups.The first step is to transfer those objects from one group into another,ifsuch transferring decreases the tr（W） value（or other criterla）,i.e.,Δ<(0(Δ=Q_new-Q_old),until no further transfer-ring can be made.The better solution ls recorded at the end. In the second step,worsening state is forced tO accept,provided that it is under proper criterion C_i,i.e.,0≤Δ≤C_i,This step usually reverses the objective function,but thepossible minimal solutions should be recorded in the process,The above two steps are repeated alternat ively withchanging criterion:C_i+1=αC_i.The third step is a group reforming process,which provides dramatic change of inter-mediate groups;The abeve three steps are repeated until there is no progress in 10～12 rounds.The algorithm is reli-able to get the global optimum in most cases.更多还原