Author Topic: Testing BBN model  (Read 12896 times)

Offline abul

  • Newbie
  • *
  • Posts: 18
    • View Profile
Testing BBN model
« on: April 24, 2009, 20:39:02 »

I would request answers to the following questions please.

1. I monitored a system for 5 variables and recorded 1000 cases (1000 time instances). Continous variables (from the system) were modelled as Interval nodes in Hugin with 3 states each. Then, I used the learning wizard to build the BBN using the NPC algorithm. Also I ran the EM learning which computed the CPTs for me.

2. Now I would like to test the BBN (with another the set of data) which I have created. What are the tools in Hugin to do this?

3. Also I have converted my BBN to an ID for decision making? I would like to check how my decisions on the test data are going to be? Basically I want to make sure the ID is making right decisions on test data, as well.

4. Can BBN make predictions with respect to time? That is (if we assume a case to correspond to a time instance), can it estimate the marginal probabilities, say 10 cases in future?

5. How can we feedback our decisions and marginal probabilities back to the system to make any corrective actions in the system settings? Is this possible in Hugin?

I would highly appreciate your response.



Offline Martin

  • HUGIN Expert
  • Hero Member
  • *****
  • Posts: 613
    • View Profile
Re: Testing BBN model
« Reply #1 on: April 27, 2009, 14:24:45 »
1. ok

2. The Analysis Wizard of the Hugin tool can perform various kinds of analysis on data accuracy.
In runmode, click Wizards -> Analysis Wizard. Also check this facility out in the Hugin GUI Manual.

3. You must add utility nodes to the network and specify utilities. In runmode one can then inspect the expected utility for decisions given evidence.

4. Such a system can be modelled as a time-sliced Bayesian network.  
An example for constructing a time sliced model using OOBNs can be found on our webpage:

5. This is not possible directly in the GUI, but using the Hugin API one can programatically transfer the configuration of the last slice to the first slice in the above example to approximate an infinite number of time slices.
« Last Edit: October 05, 2012, 12:04:23 by Anders L Madsen »
Hugin Expert A/S