How to Build an Influence Diagram

This tutorial shows you how to implement a small influence diagram in the Hugin Graphical User Interface. It requires that you have already constructed the Bayesian network from the How to Build a Bayesian Network Tutorial. The influence diagram you are about to implement is the one modeled in the Influence Diagrams Tutorial. It helps plantation owner Apple Jack to decide whether or not to give his apple tree, which is losing its leaves, some treatment. The qualitative representation of the influence diagram is shown in Figure 1.

Figure 1: The qualitative representation of the influence diagram used for decision making in Apple Jacks plantation.

Open the Network for Editing

First, you must open the network constructed in the How to Build BNs tutorial if it is not already open. Here is how to do it:

In Figure 2, the network has been opened and the Hugin Graphical User Interface is currently working in Edit Mode. We need to be in Edit Mode to edit the network, so if your network window shows the network in Run Mode, press the Edit Mode tool button. If you opened it in Edit Mode, you do not need to do anything.

Figure 2: The Network Window in Edit Mode with the network from the How to Build BNs tutorial.

Copying Nodes

In the influence diagram in Figure 1, there are three nodes very similar to those that we already have. In this case, the Hugin Graphical User Interface allows you to copy a group of nodes and paste them in another area of the Network Pane. Here is how to do it:

The Hugin Graphical User Interface generates new names and labels for the new nodes. You can keep the names and change the labels to Sick', Dry', and Loses' (you cannot use "Sick'" as the name because it contains the prime character which is illegal in names):

Perform the steps above for all three new nodes. Your network should then look as the one in Figure 3.

Figure 3: The network extended with Sick', Dry', and Loses'.

The next step is to add causal links from Sick to Sick' and from Dry to Dry':

Holding down the SHIFT key enables you to create more causal links sequentially without having to reactivate the Link Tool.

Adding a Utility Node

So far, the network we have constructed is still a Bayesian network. Now, we shall make the first change that makes it an influence diagram. This change is the addition of a utility node. The utility node we shall add is the Harv node (see Figure 1) representing the utility gained from the harvest. Here is how to add it:

The harvest depends on the state of Sick' and thus there is an link from Sick' to Harv. Add this link:

The utility of the harvest was specified to that found in Table 1.

Sick'="sick" Sick'="not"
3000 20000
Table 1: U(Harv).

You enter the values of Table 1 into the utility table of Harv as follows:

A Desicion Node and One More Utility Node

Now, you are about to add the decision node Treat (see Figure 1). This is done similar to the way you add chance nodes and utility nodes:

You add an action to a decision node in the same way as you add a state to a chance node:

The Treat decision node has an impact on the Sick' node so:

The new decision node represents the decision to give the tree some treatment or not. If the plantation owner (Apple Jack) chooses to give treatment this will cost him something which shall be modeled by the Cost utility node. The Cost node has the utility table shown in Table 2.

Treat="treat" Treat="not"
-8000 0
Table 2: U(Cost).

Now, add the Cost utility node to the influence diagram:

Filling in CPTs

When we copied the nodes Sick' and Dry', they inherited the CPTs of Sick and Dry. However, as both these nodes have become children of other nodes, their CPTs are no longer correct. Their new CPTs were specified to those found in Table 3 and Table 4.

  Treat="treat" Treat="not"
Sick="sick" Sick="not" Sick="sick" Sick="not"
Sick'="sick" 0.20 0.01 0.99 0.02
Sick'="not" 0.80 0.99 0.01 0.98

Table 3: P(Sick' | Sick, Treat).

  Dry="dry" Dry="not"
Dry'="dry" 0.60 0.05
Dry'="not" 0.40 0.95

Table 4: P(Dry' | Dry).

Now, your influence diagram is finished and it should look like the one in Figure 4. At this point it would be a good idea to save your influence diagram.

Figure 4: The complete influence diagram.

Compiling the Influence Diagram

You can now try out the influence diagram. First, compile the influence diagram:

In addition to the errors described in the How to Build BNs tutorial in the case of influence diagrams the compiler alse checks that there is a directed path through all of the decision nodes - if not it will return an error message. If the influence diagram does not compile, you have probably made some minor error. You should first check that the causal links are correct. Then, go through each of the CPTs/utility tables of the nodes.

What Should Apple Jack Do?

When the influence diagram has been compiled, first imagine that the only thing Jack knows about his tree is that it is losing leaves. Then, what will be the best thing for him to do? To find out this, follow these steps:

You should be reading something looking like that in Figure 5.

Figure 5: The influence diagram propagated with the evidence that Loses="yes".

You read 10234 as the expected utility of giving treatment and 11514 as the expected utility of doing nothing. This suggests that it will be best for Apple Jack not to treat the tree.

If you read other values than those specified above, you have probably mistyped something when filling in the CPTs. Then, check the CPTs/utility tables of all the nodes.

In a decision situation, your opinion about what to do will sometimes change when you learn more facts about your situation. Lets see what happens if Jack knows that it has been raining a lot lately and that the tree under no circumstances can be suffering from drought. Then the state of Dry can be set to "not":

You should read 9138.33 as the expected utility of giving treatment and 5918.33 as the expected utility of not doing anything. In this case, it will obviously be best for Jack to give the tree some treatment.

The reason for the difference between these two cases is that in the first case, it is likely that the tree is suffering from drought. Then, of course, the costs of treating the tree for a sickness will not pay off.

This finishes the tutorial. You should now be able use the Hugin Graphical User Interface to construct your own influence diagrams. However, if you want to create large and complex models, you should study the area more than just reading this tutorial. Also, you might be interested in learning more about the semantics of influence diagrams and the associated constraints imposed on their usage.


Back