2001 Volume 13 Issue 4 Pages 387-396
This paper discusses a Fuzzy c-Means regularized by fuzzy entropy, and proposes a new algorithm that estimates cluster numbers first and then improves clustering accuracy. In the first stage, this method estimates the cluster numbers by dividing and combining the clusters repeatedly. In the dividing process, the fuzzy entropy measures the fuzziness of the distribution of clusters, and in the combining process, new fuzzy relative entropy measures the resemblance of the shapes of clusters. In the second stage, we use Fermi-Dirac type distribution functions to represent membership functions for the clusters and adjust the shapes of Fermi-Dirac functions by applying a heuristic search method, Simulated Annealing, so as to minimize the fuzzy relative entropy. Numerical experiments show that the proposed algorithm can estimatethe cluster numbers ; the heuristic search method with the fuzzy relative entropy increases clustering accuracy for the date distributions overlapped with each other.