Author Topic: structural learning  (Read 12150 times)

Offline ngh

  • Newbie
  • *
  • Posts: 6
    • View Profile
structural learning
« on: July 25, 2010, 13:33:58 »
Dear all
I have bunch of data that I like to inject to the Hugin software to get the possible BBN model.

I save my data as a .dat file and used structural learning to inject data into the Hugin software. Following is a sample file regarding to my data captured from a custom network.

As log as I import data to Hugin, I get messages such as to many data to learn, etc.


A, B, C, D
10, 100, 200, 100
30, 200, 100, 300
.
.
.
.



Please somebody help me to solve this problem.


Regards,
Ngh

Offline Anders L Madsen

  • HUGIN Expert
  • Hero Member
  • *****
  • Posts: 2278
    • View Profile
Re: structural learning
« Reply #1 on: July 26, 2010, 10:02:01 »
Hi,

To give you the most specific reply it would be useful if you please specify the exact wording of the error message.

If you are using our demonstration version HUGIN Lite, then this tool is limited to 50 states and 500 cases.

As your data is continuous by nature and since HUGIN does not support structure learning with continuous nodes, you have to discretise the nodes. This can be done using the "Analyse" dialog of the Learning Wizard (press "Analyse" in the Data Preprocessing step of the Learning Wizard).
HUGIN EXPERT A/S

Offline ngh

  • Newbie
  • *
  • Posts: 6
    • View Profile
Re: structural learning
« Reply #2 on: July 26, 2010, 11:29:32 »
Thanks for your email.

I am using Higing Lite user 7.3

Following is data I tried to import to Hugin using structural learning as well as error message I receive:

T,A,B,E
2,0,0,0
66 1008,576,504
314,2976,3072,1584
561,3924,3996,2124
809,4752,4464,2448
1057,5328,4536,2664
1305,5595,4690,2818
1553,5688,5076,2934
1801,5600,5200,3024
2049,5572,5414,3096
2296,5668,6178,3141
2544,5772,6276,3192
2792,5815,6347,3223
3040,5873,6449,3260
3288,6048,6412,3292
3536,6039,6381,3321
3783,6005,6310,3345
4031,6168,6288,3368
4279,6290,6389,3387


Error message: The limits of this software have been exceeded (too many states or too many cases)


Best regards,
Naghmeh

Offline Anders L Madsen

  • HUGIN Expert
  • Hero Member
  • *****
  • Posts: 2278
    • View Profile
Re: structural learning
« Reply #3 on: July 26, 2010, 12:19:22 »
If you enter the data as is without discretization of the nodes, then each node will have one state for each value in the data. Thus, each node will have 19 different states producing a total of 76 states (above the limit of 50 states).

If you, on the other hand, discretize each node to have two states (using Equal distribution in the Analysis dialog of the Learning Wizard), then you will obtain the attached network (where links have been directed randomly).

I also attach the data file constructed from your latest post (I have added an "," in the second line)

I hope this helps.
HUGIN EXPERT A/S

Offline ngh

  • Newbie
  • *
  • Posts: 6
    • View Profile
Re: structural learning
« Reply #4 on: July 26, 2010, 12:49:34 »
Thanks a lot it realy helped  :D

Offline ngh

  • Newbie
  • *
  • Posts: 6
    • View Profile
Re: structural learning
« Reply #5 on: July 26, 2010, 13:40:34 »
Thanks for your email.
I gone through the learning wizard and I captured the model using the .dat file you sent me in the previous post.
The connections between states are different in my captured model and the one you already sent me.
I just want to know why the results are different.
Many thanks
Ngh

Offline Anders L Madsen

  • HUGIN Expert
  • Hero Member
  • *****
  • Posts: 2278
    • View Profile
Re: structural learning
« Reply #6 on: July 26, 2010, 14:27:57 »
There can be a number of reasons why the results are different as a number of decisions have to be made in the learning process. Using your data the following decisions have to be made

1) A discretization of each node (did you use the same as me)

2) The Learning Wizard found one "ambiguous region" which has to be resolved. I selected to included one of the two links at random

3) Direction on the links. I selected the orientation of the links randomly as I do not know what the nodes represent

Each of these decisions may produce the difference-
HUGIN EXPERT A/S