ABSTRACT
The Expectation Maximization (EM) algorithm is one of the most popular methods for obtaining the maximum likelihood estimate. One of its convenient properties is that it guarantees an increase in the likelihood function in every iteration (Dempster 1977). Moreover, since EM operates on the logscale, it is analytically simple and numerically stable for distributions that belong to the exponential family such as Gaussian. However, the conventional EM algorithm has some drawbacks. First, it may converge to a local optimum of the likelihood function. Second, the algorithm may suer from singularity. A great deal of eorts for EM algorithm to obtain the global optimal solution has been made, including the Deterministic Annealing Algorithm (Rose 1998) and the Minimum Message Length (Figueiredo 2002). Our primary goal is to achieve stable optimal solutions in an uncomplicated way when the conventional EM attains locally optimal solutions and other eorts to obtain the globally optimal solutions are rather complicated.
