|
I am getting observations in real-time, and my goal is twofold: 1) Given the probability distribution on the states before the observation was made, update this distribution based on the most recent observation.(Thus perform incremental, step-by-step updates of the probability distribution I am in). 2) Adjust the transition and observation probabilities(model parameters) on the go with each observation made. I have some sort of prototype model that I am starting off with, but I want to be able to fine-tune it as real-time conditions change. I read some tutorials online, but nothing seemed to specifically deal with this problem. Also, I would be curious to learn if there is a Java implementation of this type of approach. Thank you very much! |
|
As far as I know a completely online HMM as you mention, trained on a single sequence rather than a sequence of sequences, is still an open research question. One way of understanding why this is hard is by realizing that the main assumption behind standard learning of HMMs is that while symbols in a sequence are not IID, different sequences are, so you can maximize the likelihood of observing these sequences and get a reasonable estimator. I don't think there's any natural exchangeable way of looking at this setting you presented. I think, however, that this is an attempt you're making at modeling a specific real-world problem. If so, you might have better luck trying out a less restrictive model, or making stronger assumptions about the generative process in your data, in a way that would occasionally "leak" information about the hidden states allowing you to revise your estimates. It'd be nice if you could share your actual problem, and then see whether there has been any research done on a variant of it. Thanks Alexandre, this was really helpful.
(Mar 07 '12 at 00:19)
Viktor Simjanoski
|
