Identifying CpG Islands: Sliding Window and Hidden Markov Model Approaches - Mathematical Concepts and Methods in Modern Biology - page 294

FIGURE 9.9
Output from the Baum-Welch algorithm for simulated data for a sequence of length 500
(Panel A), 1000 (Panel B), and 100,000 (Panel C). The HMM model parameters used
for the simulation are presented above Panel A.
1000, and length 10,000, respectively. The results are presented in Figure 9.9 .The
example illustrates what should be intuitively clear: the longer training sequences
yield more accurate estimates for the HMM parameters. Due to the stochastic nature
of the process, shorter sequences are more likely to not contain the necessary number
of transitions and emissions to accurately estimate the HMM parameters. Notice
that since we use a single training sequence in all cases, the estimates for the initial
distribution cannot be accurate. Using a set of training sequences vs. a single very
long training sequence would be advantageous if obtaining estimates for the initial
distribution is important.
Exercise 9.16. Experiment with the Baum-Welch algorithm for the Dishonest
Casino application in the CpG Educate suite. Simulate sequences of various lengths
using HMM models with different sets of parameters. How well in your opinion is