The next sections would be statistical learning which I think is important so I think we should

do that but there is not much there.

Then there is this entire section on knowledge and learning which I'm not too happy about

so I think I will try to condense that down as much as possible because a lot of that

is basically vague and wavy stuff that is difficult to understand what even the point

is that very much goes for Russell Norvig as well.

I'm not a fan of that chapter at all.

There is inductive logic programming in there which is kind of nice so I don't think I'm

going to skip that but the rest of the knowledge and learning chapter I might just condense

down to something that we can cover in one session.

Then we can go on to natural language processing which is also a section which I'm currently

not entirely happy about because it's somewhat outdated with respect to what it puts its

focus on.

In the year of our Lord 2024 I would guess that natural language processing should be

mostly about like training large language models, transformer architectures, attention

mechanism, transfer learning, these kinds of things.

So I might hopefully if I find the time which is always like a big question mark to just

redo that section entirely and focus on the stuff that actually matters nowadays.

Although like all of the kinds of word embeddings like glove and all of these things are still

like important so that still makes sense at least.

So ideally at some point we will still do natural language processing.

Oh, reinforcement learning, I forgot about that.

Yes, that will also be a thing.

Right I have to understand reinforcement learning.

Okay that I can do.

Yep.

Good.

So what did we do so far?

I'm going to skip all of the preface and administrativa stuff because basically that is entirely unimportant.

So the first thing we did is we talked about agents and environments.

So the takeaway message here is obviously the general idea of an agent architecture

which is this high level representation of what we are trying to do in AI as a field

in the first place.

So the assumption is that we want to create some kind of agent that interacts with some

kind of environment in various kinds of ways.

And the most important thing there, namely the thing that influences what our agent should

do or not do is obviously the notion of an environment.

So you should keep in mind what the PIS model is, i.e. performance measure environment actuator

sensors where we don't actually care about the actuators and we don't actually care about

the sensors.

We only care about what they do, i.e. we care about percepts and we care about actions that

our agent can do.

Do we have any questions about the PIS model as a general architecture?

Okay then the next thing that is important is how to classify the environment because

the nature of the environment influences what kinds of techniques we are going to use, i.e.

is the environment fully observable, is it partially observable, is it stochastic, is

it deterministic, these kinds of things.

Do we all know what that means, do we all know what the kinds of environments are, the

classification schemas and so on and so forth.

Any questions about that subject?