forum.hugin.com
User Forums => HUGIN GUI Discussion => Topic started by: coldfire on March 13, 2009, 18:05:36

Hello All,
I am a newbie in HUGIN GUI. I am studying uncertain reasoning in order to give suggestions or persuasions to the user. Lets consider I want to simulate the scenario similar to one given in examples as well.
e.g. given properties of being "Fit" (healthy) as,
"Eat Vegetables" , "Do exercise", "Sleep Well" ,etc.... these are the nodes on which I have to reason that out of 10 possible properties of being "Fit", the person only follows 4, so the percentage of being fit will be 40% but
 How should I make its BN?
 Are the CPT values same as probabilistic weight assigned explicitly to some node? Because I am not sure.
Apart from that, for uncertain reasoning, I should not have to give the CPT values for the expected value .e.g in the above case of being "Fit". IF that is the case than if there are 20 properties of being fit, than I had to enter more than hundred values in the CPT of node "Fit"?
 How to predict without entering the CPT values?
One more question regarding the same above scenario is:
If I have 2 expected values like "Fit" and "Fit2" from the given 10 properties, whereas, Fit precentage will be if 6/10 properties are matched than we will say its 60% Fit and if the rest of them will be assigned to "Fit2" i.e. 40%. How to achieve this?
I am quite novice in it. Sorry for that. Still learning.
thanks alot

A Bayesian network consists of an acyclic, directed graph (DAG) and a set of probability distributions. In a fully specified Bayesian network each node has a (conditional) probability distribution given its parents. This implies that in order to do probabilistic inference (belief update) using a Bayesian network you have to specify a prior distribution for each node with no parents and a conditional probability distribution (CPD) for each node with parents. The CPD of a node X with parents A and B will (in the discrete case) specify one distribution over the states of X for each configuation of A and B.
The number of entries in a CPD grows exponentially with the number of parents of the node. This implies that you may have to specify many numbers in a CPD. However, a CPD can be specified by hand, using the Table Generator, which makes it possible to specify the CPD using mathematical expressions relating configurations of the parents to states of the child or it can be estimated from data (with very large CPDs you need very much data).
In your example you may consider using the Table Generator to define an appropriate expression for computing the distribution over the child node given a configuration of the parents. For instance, you may be able to define an expression which computes the number of properties being present and set the probability of Fit being true as the number of properties being present divided by the number of properties.
Hope this helps

Using Table Generator is not an easy task....i am stuck in building expressions and completing it successfully. I tried few scenarios given as printscreen in the hugin help docs as:
(http://img17.imageshack.us/img17/8509/ps2u.jpg)
but i could not made this scenario...... may be a couple of sample networks with expressions should be given with helping material to make things easier....
at one time, i tried to run the expression "Distribution(C1,C2)" over C3 (C1 and C2 are the parents of C3, all are interval nodes), it gave me some CPT values which appeared ok. But, when I tried again, the same thing, it doesnt work.

The example from the tutorial is attached. Notice that the functionality of the Table Generator depends on the subtype of the node.
With no background training the use of Bayesian networks can be a challenge. Therefore we offer a training course on HUGIN and our technology: http://www.hugin.com/Products_Services/Services/Training/ (http://www.hugin.com/Products_Services/Services/Training/)