关键词: 退火算法, 聚类算法, 期望最大化, 有限混合模型, 自由能量方程, 模型选择

Abstract: Drawbacks of finite-mixture-model-based clustering algorithms are as follows：the dependence on the choices of initial parameters；the convergence to a local optimum；the incapability of cluster number determination. To tackle these problems, a new model is developed by introducing a penalty function into the negative free energy function. The learning of this model presents a similar annealing process as that of the simulated-annealing algorithm, which may increase the chance of global optimization. A generalized model selection method is introduced for cluster-number determination. To avoid either a sudden temperature decay in an annealing process or an over-or under-estimation in the number of clusters, the weight of penalty function is adaptively controlled by a novel method based on the overlapping degree between clusters. Experimental results show that the proposed approach outperforms the conventional algorithms.

Key words: annealing algorithm, clustering, expectation maximization, finite mixture model, free energy function, model selection