In the first step, hmmgenerate function is used to find initial sequence of emissions and states which takes state transition and emission probabilities as an input. What are the different ways to find these probabilities. Can we simply guess these probability matrices with the help of medical experts.
Hi,
If no experts are available to broadly specify a good set of initial parameters, come unsupervised tool can still be used like some clustering tool, based on which, initial parameters can be inferred. For instance, K-means. Blind stochastic methods also are used for such tasks.
Regards,
Hello. As HMM is a Bayesian estimation problem, prior knowledge is your friend. Use as much information as you can about the likely model structure, noise levels, number of states, unlikely transitions, etc. The suggestion about unsupervised clustering is a good one. Check out Rabiner's Proc IEEE survey paper if you haven't already.
A word of caution: the EM algorithm gives local convergence only. If you have too many parameters or start a "long way" from a good solution, your HMM will probably not be much use.
Hello Aman,
How did you find emission and transition probabilities Aman Singh ?
- Badges
- Science topic
- Similar topics
- Statistics
- Time Series