Okay, so the quiz is over.

Welcome to the last week of AI1.

What you've just done is not the last quiz.

So the quiz, to remind you, I'm trying to do two things with that.

One is kind of to keep you working continuously.

And the other one is to give you quiz material for preparing for the exam, which is still a ways off, but it's coming for sure.

And that means that the quiz has to cover all material that could be in the exam, including this week.

So we'll just formally have another quiz. You can just have a go at it and try to increase your bonus points.

But of course that's as you please. All is optional.

And even that is not the last one, because we're going to kind of hijack the quiz mechanism to give you a little incentive for filling out the Alia evaluation.

So that will count as a quiz. I'm going to show you details about that tomorrow.

So you can fill out the survey. At the end you get an individual token.

And if you fill out that token into the only quiz 16 question, then you'll get 100% of that.

So that's certainly a good way of increasing your bonus points.

Okay, good.

So last week we looked at planning algorithms.

So you can think of them as either search with structured world representations.

That's one way of thinking about planning. Or you can think of them as logic-based agents with a special consideration of changing environments.

When you change something by your own actions, things aren't what they used to be, which is exactly what we want to do by the actions.

Okay.

And we looked at two kinds of algorithms. One very classical one, namely partial order planning.

The idea there is that instead of looking at strings of actions or sequences of actions that kind of build on each other, we generalize that to partial orders of actions.

Where we don't commit on a full linearization in the beginning, we only do that essentially when we execute.

So this kind of algorithm was even more classical, was a very normal search algorithm.

But with an interesting heuristic.

Because we have structured world representations now.

Remember in strips we have the states being sets of facts which are conjunctions of literals, atoms actually, in the facts.

And the actions have these preconditions add and delete lists and so on.

Because we can look into the states, because we can look into the actions, we can actually build heuristics.

And the heuristics was just counting the optimal or any solution length in a relaxed problem and that gives us guidance.

And that was what the delete heuristic actually gave us.

A relatively simple general heuristic to speed up search.

And again, that's what you get by having structured world representations.

That is not something you can do with atomic world representations because you have nothing to look at.

And those things are relatively efficient.

But remember in the beginning when we talked about rational agents, we had this classification of environments.

The environments could be fully observable or only partially observable.

They could be static or dynamic.

The actions could be deterministic or not.

And we had a whole list of those.

And there was always kind of the easy way, fully observable for instance.

And the hardware, partially observable.

And we very strictly only did the easy stuff.

So at the end of the semester we're going to kind of explore a little bit of what would happen if we relax one or the other for that.

And I'm going to kind of tell the story of what would happen in terms of planning.

What would you do? What would the agents do?

And we're going to do this much better next semester.

But we can kind of take first steps that hopefully give you a little bit of an intuition of what needs to change mostly.

For when you maybe come back next semester.

So that's our plan for now.