Prev: contour color question
Next: Variable-Gain Altitude Control
From: PEDRO GARCIA on 9 Jul 2010 14:39
"PaulN" <paulnicholl(a)hotmail.com> wrote in message <1146737339.293065.111490(a)u72g2000cwu.googlegroups.com>...
> Thanks Maria.
>
> I'll play about with those values and see if it fixes things. I would
> have hoped though that, if I used stupid values, the HMMs would have
> given low recognition rates, rather than produce an error during
> training.
>
Hi PaulN,

Although it has been a long time since you experienced this error, I hope you can help me. Please. I am going mad with it.
I am experiencing the same issue but with dhmm_em function. I wonder if you founded a solution to this error of the toolbox.

Thanks in advance,
Pedro
From: Mero on 15 Jul 2010 17:13
On 7/9/2010 2:39 PM, PEDRO GARCIA wrote:
> "PaulN" <paulnicholl(a)hotmail.com> wrote in message
> <1146737339.293065.111490(a)u72g2000cwu.googlegroups.com>...
>> Thanks Maria.
>>
>> I'll play about with those values and see if it fixes things. I would
>> have hoped though that, if I used stupid values, the HMMs would have
>> given low recognition rates, rather than produce an error during
>> training.
>>
> Hi PaulN,
>
> Although it has been a long time since you experienced this error, I
> hope you can help me. Please. I am going mad with it. I am
> experiencing the same issue but with dhmm_em function. I wonder if you
> founded a solution to this error of the toolbox.
> Thanks in advance,
> Pedro

I believe the problem is that during the calculation, you are getting a
zero probability somewhere (ie: some symbol or vector could have been
generated by that model configuration at a time t). You could try
re-running the EM algorithm again with different initializations for the
HMM model parameter.

Another common adjustment is to use a Laplace estimator (ie: adding a
count of 1 to all possible transitions and emissions), similar to the
'Pseudotransitions' and 'Pseudoemissions' options of the _hmmtrain_
function from the Statistics Toolbox. They are used to "avoid zero
probability estimates for transitions/emissions with very low
probability that might not be represented in the sample sequences."

For the HMM with mixture of gaussians output(mhmm), a similar approach
could be done for the initial state and transition matrix. For the
output model, you might want to initialize the Mu/Cov from the data
itself (perhaps by clustering the data first into Q*M clusters and
initializing the mixture components from data in each of those clusters).

--Mero
 | 
Pages: 1
Prev: contour color question
Next: Variable-Gain Altitude Control