Thread Subject:

Hidden Markov Model : a question ?

I'm using the HMM toolbox for a data. As you know, MATLAB has a
"hmmtrain" function that gives you estimates of the hmm parameters.
but, to use this code, you need to have a very good initial guess of
the parameters. my question is what if the only thing you have is the
data string itself(the observations) and nothing more?
 then how can
you come up with a good initial guess for the transmission and
emission matrices?

you can use kmeans to come up with state prototypes. usually the transition
matrix is either randomly initialized or uniformly initialized.


"Qaaddi" <bassqqad@yahoo.com> wrote in message
news:ef58fe2.-1@webcrossing.raydaftYaTP...
> I'm using the HMM toolbox for a data. As you know, MATLAB has a
> "hmmtrain" function that gives you estimates of the hmm parameters.
> but, to use this code, you need to have a very good initial guess of
> the parameters. my question is what if the only thing you have is the
> data string itself(the observations) and nothing more?
> then how can
> you come up with a good initial guess for the transmission and
> emission matrices?

Ganesan Ramachandran wrote:
>
>
> you can use kmeans to come up with state prototypes. usually the
> transition
> matrix is either randomly initialized or uniformly initialized.
>
>
> "Qaaddi" <bassqqad@yahoo.com> wrote in message
> news:ef58fe2.-1@webcrossing.raydaftYaTP...
>> I'm using the HMM toolbox for a data. As you know, MATLAB has a
>> "hmmtrain" function that gives you estimates of the hmm
> parameters.
>> but, to use this code, you need to have a very good initial
guess
> of
>> the parameters. my question is what if the only thing you have
is
> the
>> data string itself(the observations) and nothing more?
>> then how can
>> you come up with a good initial guess for the transmission and
>> emission matrices?
>
*****************************
Dear Mr.Ramachandran
Thanks for reply.
Actually i have a codebook of quantized vectors as my observations.I
want to recognize 15 patterns.
each pattern is a series of a number of codewrods.
I want to train 15 models to recognize 15 patterns and all what i
have is the sequence of codewords(observations).

The function [hmmestimate] needs the state sequence which i dont have
, and the function [hmmtrain ]needs a guess of initial Transitions
and Emissions matrices which also i dont have

Sorry, i didnt understand how k-means will help me?
plz, try to explain alittle .
Thanks alot

On Wednesday, June 6, 2007 2:19:31 AM UTC+9, Qaaddi wrote:
> Ganesan Ramachandran wrote:
> >
> >
> > you can use kmeans to come up with state prototypes. usually the
> > transition
> > matrix is either randomly initialized or uniformly initialized.
> >
> >
> > "Qaaddi" <bassqqad@yahoo.com> wrote in message
> > news:ef58fe2.-1@webcrossing.raydaftYaTP...
> >> I'm using the HMM toolbox for a data. As you know, MATLAB has a
> >> "hmmtrain" function that gives you estimates of the hmm
> > parameters.
> >> but, to use this code, you need to have a very good initial
> guess
> > of
> >> the parameters. my question is what if the only thing you have
> is
> > the
> >> data string itself(the observations) and nothing more?
> >> then how can
> >> you come up with a good initial guess for the transmission and
> >> emission matrices?
> >
> *****************************
> Dear Mr.Ramachandran
> Thanks for reply.
> Actually i have a codebook of quantized vectors as my observations.I
> want to recognize 15 patterns.
> each pattern is a series of a number of codewrods.
> I want to train 15 models to recognize 15 patterns and all what i
> have is the sequence of codewords(observations).
>
> The function [hmmestimate] needs the state sequence which i dont have
> , and the function [hmmtrain ]needs a guess of initial Transitions
> and Emissions matrices which also i dont have
>
> Sorry, i didnt understand how k-means will help me?
> plz, try to explain alittle .
> Thanks alot

Hi Qaaddi how do you quantized vector ?

I am working in character recognition. below is my properties of input image
type : binary image
width : 400
height : 100
then i divided each image into 10 horizontal blocks. After that i calculate the number of black pixels in each block.
Suppose this is my feature vector.
feature=[1 2 3 4 5 6 7 8 9 10];

how do i train using matlab hmmtrain command and how to define emission matrix.

Thanks