Data clustering with entropical scheduling

Data clustering is fundamental for training networks based on vector quantization and/or radial basis functions. This paper presents a new method of clustering data based on entropical scheduling. In the proposed entropical scheduling: 1) the assignment of a cluster to each data, for the update of the cluster center, is probabilistic, where the probabilities that each data belongs to individual clusters depend on the distances to the corresponding cluster centers; and 2) the entropy of assignment probabilities are gradually minimized in such a way that the nature of assignments evolves from more random to more nearest neighbor decisions. The proposed method aims at reducing the sensitivity of the K-means method to local minima through probabilistic assignments, while enhancing the efficiency of simulated annealing by incorporating the neighborhood constraint in random perturbation. Simulation results demonstrate the efficiency and robustness of the proposed method in comparison with conventional methods