Author Topic: A list of questions (sorry about it)  (Read 20429 times)

Offline grgcln

  • Newbie
  • *
  • Posts: 1
    • View Profile
A list of questions (sorry about it)
« on: December 04, 2007, 01:37:03 »
Hello everyone,
I am a HUGIN newbie. As I am learning the new software, I am trying to figure out some things (which I could not find on the help file). I do apologize if the list of questions sounds too long.

1) Continuous nodes : is there a chance to have probability distributions other than Gaussian? If not, is there an automated way in Hugin to compute the coefficients of a linear combination of Gaussian distributions to approximate a non-Gaussian probability density function? Or anything at allo along those lines? 


2)  How can I perform parameter learning of a network parameters (with continuous node) from a data file of samples? What is the structure of the data file supposed to be? I think I was only able to find an example for discrite nodes.


3) Any implementation of Monte-Carlo-Markov-Chains in the software?


4a) Where can I read more about the batch capabilities of Hugin?
4b) What if I want to investigate different scenarios of evidence, is it possible to automate the process?
4c) How are results saved on file?

5) When I create the network and execute (-lightning icon) I only see number up to the 2nd digit. How can I increase the accuracy (or display of digits)?


MANY MANY Thanks everyone!
 


Offline Martin

  • HUGIN Expert
  • Hero Member
  • *****
  • Posts: 613
    • View Profile
Re: A list of questions (sorry about it)
« Reply #1 on: December 28, 2007, 14:43:10 »
Hello.

(I know you have allready received an answer in the support email correspondance - so this is just to share it with the other users of the forum.)


Quote
1) Continuous nodes : is there a chance to have probability distributions other than Gaussian? If not, is there an automated way in Hugin to compute the coefficients of a linear combination of Gaussian distributions to approximate a non-Gaussian probability density function? Or anything at allo along those lines?

Exact reasoning with nodes having continuous distributions is
difficult.  It is possible for CG nodes because of the nice theory
developed for CG distributions.  See, for example, Cowell et al
(1999).

Other continuous distributions can be approximated using CG
distributions, but there is no automated tool in the Hugin software to
do that.

On the other hand, there is a table generator in Hugin that can
discretize continuous distributions based on mathematical expressions
for the distributions.  The node type for discretized continuous
distributions is the "interval" node type.


Quote
2)  How can I perform parameter learning of a network parameters (with continuous node) from a data file of samples? What is the structure of the data file supposed to be? I think I was only able to find an example for discrite nodes.

Data for continuous nodes are simply numbers.

Parameter learning for CG nodes is not supported, but they can be
present and will affect the learning process for the discrete nodes in
the network.  Interval nodes are discrete and can be learned.


Quote
3) Any implementation of Monte-Carlo-Markov-Chains in the software?

No


Quote
4a) Where can I read more about the batch capabilities of Hugin?

In the HTML help pages.


Quote
4b) What if I want to investigate different scenarios of evidence, is it possible to automate the process?

What do you mean by "investigate"?  There are a number of "analysis"
functions in the Hugin GUI.  Also, it is possible to create case files
to hold scenarios of evidence.


Quote
4c) How are results saved on file?
5) When I create the network and execute (-lightning icon) I only see number up to the 2nd digit. How can I increase the accuracy (or display of digits)?

Under View -> Belief precision.
Hugin Expert A/S

Offline Martin

  • HUGIN Expert
  • Hero Member
  • *****
  • Posts: 613
    • View Profile
Re: A list of questions (sorry about it)
« Reply #2 on: April 28, 2011, 14:32:34 »
Just a small update to anyone skimming through this old post

Quote
Parameter learning for CG nodes is not supported

Parameter learning for CG nodes IS supported from Hugin 7.1 onwards.
Hugin Expert A/S