Efficient algorithms for training the parameters of hidden Markov models using stochastic expectation maximization (EM) training and Viterbi training

Table 2 CPU time use for different models

CPU time (sec) per iteration	dishonest	extended dishonest	CpG island
	Casino	Casino	Model
Baum-Welch training	8.85	5.94	22.22
stochastic EM training K = 1	5.12	3.42	5.42
stochastic EM training K = 3	6.02	4.42	10.30
stochastic EM training K = 5	7.06	5.38	14.84
Viterbi training	4.42	2.84	5.00

Overview of the CPU time usage in seconds per iteration for Viterbi training, Baum-Welch training and stochastic EM training for the three different models. For each model, we implemented each of the three training methods using the linear-memory algorithms for Baum-Welch training, Viterbi training and stochastic EM training. The number of state paths that are sampled for each iteration and each training sequence in stochastic EM training is denoted K.

ISSN: 1748-7188