Consider the Bayesian belief network (Bbn) shown in figure 1. To learn
more about the domain of this Bbn refer to the section network
samples.
Figure 1: Bbn Representation of "Chest Clinic"
Experience
table or tables are necessary to use adaptation. To add experience table to all
discrete chance nodes in the domain click somewhere in the network pane, push
the right mouse button and then choose "Add Experience Table to All
Discrete Chance Nodes". It
also possible to add experience tables to a subset of discrete chance nodes in
the domain. To do the later select a specific discrete chance node then right
click and select "Add Experience Table".
In
our example it is justified to add experience tables to all the nodes in the
domain except "Tuberculosis or cancer" since this node is a logical
or and no experience can be gained on logical or nodes. To see the created
experience table select the desired node and right click then select "Show
Experience Table".
As mentioned
earlier, the experience table of a node represents the experience count of the
parent configurations. For example the experience table for node "Dyspnoea?"
which is shown in figure 2,
Has bronchitis |
yes |
no |
||
Tuberculosis or cancer |
yes |
no |
yes |
no |
Experience count(s) |
0 |
0 |
0 |
0 |
Figure
2: Experience Table for the "Dyspnoea?" Discrete Chance Node
Represents
the number of observations of different parent configuration. The value zero is
an invalid experience count thus the value must be greater than 0 to activate
adaptation. If our believe of the correctness of the present conditional
distribution probability is high then the experience count must have a high
value otherwise the value of the count should be low. In this case we assume
that our believe in the correctness of the current conditional distribution is
low thus we set the initial experience count to a low number, for instants
"10". Figure 3 show the initial experience count table for "Dyspnoea?"
discrete chance node.
Has bronchitis |
yes |
no |
||
Tuberculosis or cancer |
yes |
no |
yes |
no |
Experience count(s) |
10 |
10 |
10 |
10 |
Figure 3: The Updated Experience Table for the "Dyspnoea?" Discrete Chance Node
Note
that it is not necessary to activate experience count or enter the same
experience count for every parent configuration. For instants the initial
experience count values can be set to "10,0,100,0". Note that
adaptation requires at least a node with experience table otherwise it is not
possible to adapt the domain. Now add experience table to the every node in the
Bbn (except "Tuberculosis or cancer" node) and set the initial
experience counts to 10. The domain is now ready for adaptation.
An
adaptation step consists of entering evidence, propagating, and finally updating
(adapting) the conditional probability and experience tables.
Lets
concentrate on one of the nodes namely "Smoker?” The conditional
distribution probability of this node prior to any adoption is S (0.5,0.5). Now
enter the following observations:
then
propagate the evidence. Next push the adaptation button which is shown in figure
4.
Figure 4: The Adaptation Icon
Keep
clicking on the adaptation button a couple of times. Each time the adaptation
button is pushed the probability of this observation (i.e. P(All)) increases.
Now initialise the Bbn and observe the conditional distribution probability of
the "Smoker?” As you can see the conditional distribution
probability is no longer S(0.5,0.5). Actually no conditional distribution
probability is the same. This indicates that based on the new observations the
conditional distribution probabilities has been changed. I.e. if the experience
tables are now deleted or the values of the experience tables are set to zero
then the current distribution probabilities will be the new conditional
distribution probabilities of the nodes.
![]() |
HUGIN Expert A/S | , 2000 |