In this paper, we discuss the dual-problem of adjusting the mixture number and avoiding local optima in the estimation of a Gaussian mixture. This estimation is widely used in unsupervised-classification applications; however, its results are serially sensitive to the initial setting, which is difficult to optimize. It is also difficult to automatically designate the mixture number in advance. In much of the literature, these two issues are discussed separately, meaning that one is considered at the expense of the other. To overcome this problem, we present some strategies that automatically and simultaneously adjust the mixture number and escape from local optima. The evaluation results are very encouraging and show that the proposed strategies are effective.
Relation:
淡江理工學刊=Tamkang journal of science and engineering 9(2),頁155-166