Hi,
What are the techniques to combine together multiple HMMs related to the same process?
Mental example:
I train 3 different HMMs by observing the behavior of 3 animals. All these HMMs have the same number of parameters and emissions. Then, I discover that I was actually observing the same animal, in three disjointed time intervals.
Moreover, for each HMM I have the same number of observations.
I want to combine the 3 HMMs to a single HMM that better models the animal behavior.
Of course this is just a mental experiment, but it fits very well my problem.
Edit:
Up to now, my reasoning is the following:
- I have 3 HMMs trained from different observation sequences from the same animal.
- Not every model has the same "wheigth": models with higher P(observations | A, B, pi) should wheight more when combining (is this correct?)
- One idea is to, for each probability, do a wheighted average for each model
Please go through these links:
https://cs.stackexchange.com/questions/29682/combining-multiple-hmm-models
and,
https://www.google.com/amp/s/www.researchgate.net/post/How_we_can_combine_several_HMM_together_for_gaining_better_results/amp
Best
I apologize, I just realized that my "mental experiment" was badly stated. I updated the description and the goal in bold.
Anyway, I had already seen those links, but I was wondering if I could do better than that. For example, in the first ink, the two HMMs refer to two different "objects": accelerator and video stream. In my case, each state of the HMMs refer to the exact same thing. The only difference between my HMMs is that they've been trained from observations taken in different time intervals.
Is there a way, for example, to combine the three HMMs in a single HMM without increasing the number of states?
Dear Federico Taschin
You can initially train the 9 observations in one model instead of using 3 models.
Yes, that would be undoubtedly better.
Unfortunately, my assumption is that at the beginning I don't know that the observations are taken from the same animal.
The best solution is the use of a pseudo 2D-HMM. The following links are for face recognition:
Conference Paper Pseudo 2D Hidden Markov Models for Face Recognition Using Ne...
https://www.degruyter.com/downloadpdf/j/ipc.2014.19.issue-1/ipc-2015-0001/ipc-2015-0001.pdf
Article Hidden Markov Models For Two-Dimensional Data
Article A face recognition system based on Pseudo 2D HMM applied to ...
- Badges
- Science topic