Okay

welcome to the next AI lecture.

The quiz seems to have worked.

I must say I'm very happy with the way people are discussing and asking questions.

Even though with all the errors you find it's work

but that's fine.

We want to improve the learning materials.

So keep on doing that.

That's the best way of learning.

Some people seem to read the text very carefully and find all the things where I've been making slides deep at night or in a hurry or something like this.

It's very good that you do that.

So we're still at classifying agent architectures.

Remember, mathematically agents are extremely simple.

They're just a function from percept sequences to actions.

Very simple

except we're not mathematicians.

We're computer scientists

which means we want something we can actually implement.

Actually

we want to design agents that we could at least imagine giving a body and letting run around.

And just abstract functions you can't build and run around.

You have to do all the decisions and designing that we're doing.

Basically

I want you to

after this course

be in a position to design agents.

Kind of pick the right algorithm

pick the right architectures

and so on.

So we looked at a succession of ever more complex agent designs.

And the goal-based agents are those that we're going to basically do with mostly in this semester.

And leave the utility-based agents and learning agents to next semester.

But I want to show you kind of the progression and the kind of natural kind of endpoint here is that whatever performance element

whatever agent design we've done

we can act

we can add a learning layer.

And the learning layer is triggered by the outside performance measure

which drives a critic that basically says

you're not doing well enough, right?

Which in turn may drive a learning element

which is something that can change certain parameters in the performance element.

And for that

it is important that we actually know what the performance element actually looks from the inside

because we have to know what the parameters that can be changed and optimized actually are.

And then

of course

we might have a problem generator that actually would allow the agent to explore.

In the back of our minds

we still have this idea that we want to build rational agents.

And rational agents don't have to know the environment.