Name
Abstract | ||
|---|---|---|
The EM algorithm is widely used to estimate the parameters of many applications. It is simple but the convergence speed is slow. There is another algorithm called the scoring method which is faster but complicated. We show these two methods can be connected by using the EM algorithm recur- sively. I. INTRODUCTION The EM (Expectation Maximization) algorithm(1) was originally proposed by Dempster et al.(2) for esti- mating the MLE (Maximum Likelihood Estimator) of stochastic models which have hidden random vari- ables. The algorithm is now used in many applications such as Boltzmann machine(3), Mixture of Expert networks(4)(5)(6) and also in HMM (Hidden Markov Model)(7). This algorithm gives us an iterative procedure and the practical form is usually very simple. However, the convergence speed is slow compared to the scor- ing method which is also used to estimate the MLE of these models. There are some works to accelerate the convergence speed of the EM algorithm(8)(9), but the procedure is usually not easy and need a lot of calculations. In this paper, we show that we can accelerate the EM algorithm by using it in a recursive way. The algorithm consists of two stages. In the first stage, we do one EM step with the given data set. In the second stage, we do another EM step not with the given data but with the data drawn from the model. Through these stages, we can have better estimator. We show the theoretical derivation of the algorithm in connection with the scoring method. We also show some results of computer simulations. They show the algorithm gives faster convergence speed. |
Year | DOI | Venue |
|---|---|---|
2000 | 10.1002/(SICI)1520-684X(200002)31:2<10::AID-SCJ2>3.0.CO;2-D | Systems and Computers in Japan |
Keywords | Field | DocType |
expectation maximization algorithm,hidden markov model,boltzmann machine,computer simulation,em algorithm,maximum likelihood estimate,stochastic model | Boltzmann machine,Forward algorithm,Markov model,Computer science,Expectation–maximization algorithm,Stochastic modelling,Artificial intelligence,Hidden Markov model,Perceptron,Machine learning,Mixture model | Journal |
Volume | Issue | Citations |
31 | 2 | 6 |
PageRank | References | Authors |
2.16 | 7 | 1 |
Name | Order | Citations | PageRank |
|---|---|---|---|
| Shiro Ikeda | 1 | 344 | 37.95 |