If you know that there's one thing,

one problem is assessing probabilities,

knowing what the probabilities of lots of different things are.

There's the second question is,

if I know those things,

what do I do?

Should I play this card or that card?

That's the question at hand.

How do I find out what the best thing is?

We humans have some kind of an intuition of this.

We're pretty good at surviving in our environment,

and we usually know jumping down there, bad idea.

Having a beer now, good idea, something like this.

But the question is,

how do we actually bring that into an artificial agent?

How do we describe it formally enough that we can actually

program it and have an agent that also has the feeling,

jumping down here, not a good idea,

and having more electricity, good idea.

Let's look at a couple of other examples that I had prepared.

I want to give a lecture somewhere,

because that's why, at 10.15.

There are two plans.

I get up at eight, leave at 8.40,

and arrive at nine. One plan.

Another thing would be to get up at 9.50, brush my teeth,

come to the, leave at 10.05, be there at 10.15.

There are more possibilities, but these two plans.

These plans, of course, have different merits.

The first one is safe,

and I can set up and do my mic and so on.

The second one gives me more sleep time,

which is also good. What do I do?

Both of the plans are correct,

meaning they both get me into

the right lecture hall at the correct time.

How do I judge?

Well, to actually know how to judge them,

we need more information.

We need more information about what can go wrong,

and what is the probability of these plans

succeeding in a partially observable,

non-deterministic, and whatever environment.

Of course, how important to me it is to be at the lecture in time.

If the lecture is just some lecture to three students,

we're going to be late anyway.

That might be different from me

giving a lecture for a job interview,

which I really want that job.

Depending on these things,