Okay.

We looked into planning as a framework and what the applications might be and what we're

actually doing in the modeling.

The idea is in the kind of strep-spaced planning that we have a declarative language for describing

the state of the world and the actions.

And we saw that if we have actions, something changes.

So that actually means two things.

One is that we need a variable for time.

Somehow we need to account for time.

So we introduced fluence.

And we have to account for change.

And whenever we change something, we are in a logic-based setting but also in many other

settings, we have the problem that we have run into the frame problem.

Right?

It's relatively easy to write down effect axioms and sensor axioms that basically describe

if I see if I have a percept at time t, what does that mean for the world?

It's also relatively easy to say if I have an action, what are the new things that are

now true?

What are the things that might not be true anymore in the new time step?

But the big thing is that we somehow have to account for what is called the frame of

reference, the stuff that's not changed by my actions.

And you can imagine that the stuff that's not changed by my actions is kind of the majority

of things.

Think about what you're doing when you're performing an action in this room.

Right?

There are 1.4 billion Chinese people who are completely unaffected by whatever I do in

this room.

Right?

Do I want to say, oh, yes, and I have 1.4 billion axioms by saying, oh, Mr. Wu in Shanghai

is not affected by call has a teaching AI in Alang.

You don't want that.

So how do we deal with this?

The trick in planning are these delete and add lists.

Right?

If you explicitly say the certain things are no longer true in the delete lists, then you

don't have to say what stays true.

That's the main thing.

The main effect of strips like planning, having add and delete lists.

And together with the declarative formatting where we use facts to describe states.

That's the essence of planning.

And now we want to look at planning algorithms.

We already saw one.

We saw partial order planning.

Which is kind of one way of doing it.

And we're going to see another currently more efficient way of doing planning.

That is what I want to show you now.

And surprisingly, the best planning algorithms nowadays are actually using heuristic search.

And there's mostly one heuristic that we're going to look at.

A reminder how the search works.

We build trees.