So we start a new part here, which is essentially learning from observations, which is machine

learning.

And I've already alluded to this, is that all our agents either need to have things

like transition models and sensor models and so on, baked in, means compiled in essentially,

or they have to learn them.

And of course, learning them is quite useful for a variety of reasons.

First of all, if you can learn things like transitions models, or learn about the world,

learn about your sensors and so on, then you as the agent designer don't have to actually

program it in.

If you don't have to program it in, then you don't have to know how convenient.

Very nice.

Of course, there's a downside to it.

You have to understand machine learning.

So machine learning is a huge topic by now, and I'm going to skim this much of it.

There's a huge research, very successfully, going on at the moment.

So successfully at actually that it kind of deteriorates research, because instead of

trying to understand what's actually going on, people now just basically say, oh, but

there's a deep learning framework from Google.

Let's just feed it some data, and hopefully something comes out.

And nicely, it often does.

For research, of course, I think that's a bad thing, because understanding is really

what should be at the end of research.

We only want to breed successful agents without understanding why.

We can do that by biological means.

So actually a lot of fun, too.

We're not going to really cover deep learning here.

Rather than that, I would like to give you the basics.

If you want to understand deep learning, take what you're learning here and add lots of

matrices to it and two ideas.

So I'm going to probably tell you about those two ideas if we still have time.

So we're going to look at, there's quite a lot of techniques here.

They're often embarrassingly simple, but very often they actually work.

And so I'd like to give you some notion of why they work.