AN OBJECTIVE BASED CLUSTERING APPROACH TO DESIGN OF MANUFACTURING CELLS

Cluster-seeking algorithms such as K-means and Isodata have been used in many fields for clustering patterns. Problems often arise from inte rpretation and evaluation of clustering results because of lack of eva luation criteria. This paper presents a new concept called objective-b ased clustering to improve the cluster-seeking algorithms by defining relationships between parameters of the algorithms and evaluation crit erion of solutions. Two phases, an initial phase and a final phase, ha ve been developed, which are used before and after the clustering-seek ing processes. The initial phase aims st the optimal selection of init ial cluster centres based on the objective; the find phase, from an ap plication point of view, corrects any misclustering caused by absolute similarity as only criterion for clustering. Both phases are domain d ependent. To further improve the performance of the Isodata algorithm, an optimization model has been developed to link to the Isodata algor ithm. The objective and constraint functions are expressed as function s of the parameter variables required by the Isodata algorithm. The op timization program selects a group of parameters for the modified Isod ata (with the initial and final phases) routine, which in turn returns an objective function value in each iteration. In such a way an optim al solution can be found. This approach has been used to form manufact uring cells, and the results show that improvement on the solutions ha s been achieved. A case study is provided to illustrate the approach.