A Quasi-Bayesian Approach to Estimating Parameters for Mixtures of Normal Distributions
提出一种易处理的准贝叶斯方法，通过添加伪数据来估计混合正态分布参数，避免最大似然估计的奇异性问题，并帮助选择局部最优解。蒙特卡洛模拟表明该方法能一致地降低均方误差，且能轻松分析最大似然法难以处理的数据集。
This article proposes a very tractable approach to estimating parameters for mixtures of normal distributions. The analyst proceeds as if, in addition to the data, he or she had observed some pseudo data points drawn from each distribution whose values reflect his or her priors. The approach eliminates the singularities associated with maximum likelihood estimation and offers guidance for choosing among alternative local maximum likelihood estimates. Monte Carlo analysis establishes its consistent potential to improve mean squared errors. Data sets on which maximum likelihood estimation has presented difficulties are shown to be readily analyzed with the quasi-Bayesian procedure.