Okay, I think five minutes we can do decision networks because they're extremely simple.

Remember what we did with Bayesian networks?

We have Bayesian networks for world models.

What we now do is we add action nodes, these little squares, and utility nodes, which are

these diamonds.

Essentially, where we have conditional probability tables here, we give ourselves utility tables

here with all the things that go into the utility computation.

Everything we did here was we added a utility table here.

Say we have, where's my example, V of something or the other, this.

That actually allows us to combine those three into a utility.

Then of course we have these action nodes, which is really where do we put the airport

in this example.

That actually is a second influence on the number of deaths and noise and costs.

The cost is essentially given by where you put the airport and then in the general tendency

of people to litigate the general cost of stuff.

For construction, it might be that you need to have a runway, which is, I don't know,

a couple of thousand cubic something or the other of concrete.

Then that has a cost in itself.

Depending on where you put the airport, you also have to see how far you have to truck

this thing in and so on.

We have these kind of Bezier networks where these round blobs are given by these conditional

probability tables.

Basically you have random variables here.

But here you have non-random variables.

You have variables where you can actually do stuff with.

You can decide.

You have decision variables rather than... You have decision variables that influence... There's

kind of an input node here, which influence this thing here.

Think about this like a spreadsheet.

When you have input variables and then you have stuff... You have formulae and then

you have outputs and then you kind of play around and move around the airport somewhere

and see what's the utility going to be like.

What you basically do is you basically cycle through the action node, through the values

of the action node.

You basically compute the expected value of the utility node here, given that action.

Given the action and the evidence you have so far, evidence about what the air traffic

will be and so on.

Then you return the maxim, the argmax essentially.

That action that maximizes utility.

Yes, that is my last slide here for today.

This is a typical medical decision network.

You have all kinds of things that can happen to a human.

You can have heart failure, typically probably the symptom.

You can have other conditions.

All of those influence the action, namely what is the treatment.

The treatment of course influences the intended results and the side conditions.

There are a couple of those.

The side conditions might actually also depend on other things.

You have a side condition which is para-something hypertension, which influences something or

the other syndrome, but also the sex of the patient, male or female, actually also influences