I apologize for being late.

The side effects of having the quizzes where it's kind of not so important that I'm here

from minute one.

And so I sometimes kind of forget which day of the week it is.

So I may get a little bit careless at the beginning times.

Okay so, we have completed the reasoning and building agents that can deal with uncertainty

chapter.

And in a way, we've kind of slowly but surely made our way towards the kind of POMDPs, which

is the most general models that our agents can actually take on.

And now we basically still have two topics left over.

One is machine learning.

And we're going to kind of go relatively bottom up way into that top look at the theory and

practice of these.

And then we're going to use statistical learning techniques and so on to deal with natural

language.

Natural language being one of the big prerequisites we looked at all in the beginning as one of

the components of intelligence.

But that's still in the future.

So the first kind of learning, the simplest kind of learning that we're going to tackle

is supervised learning.

Learning by being told.

So really the performance measure of that is actually matching some kind of an input-output

function.

So as always, we try to understand the problem of what is learning as a mathematical problem.

A mathematical problem is something that has problem instances.

Try to learn this.

Try to learn the stock market development of Mercedes-Benz stocks or something like

this.

That's an instance of the problem.

And then we have solutions and kind of everything that's every function that we learn, every

time series that behaves like the stock of Mercedes-Benz over time is a solution, a possible

solution.

And then there might be better solutions and worse solutions.

So here the kind of idea is that we want to learn a function, a function of inputs

to outputs based on examples.

Examples are just input-out, specific set of input-output pairs.

But that's not all.

That's kind of the general thing.

We also, and that's maybe the thing to actually remember, is we're also saying we're giving

our set a hypothesis set, the set from which our functions may actually be picked.

Okay?

So that's an important parameter that kind of if you go out formalizing what an inductive

learning problem might be that you might actually easily forget.

So that's this funny set H. From which set do we pick our solutions?

What are the allowed solutions?

And so we basically, the function that we should learn, the target function, remember

the development of Mercedes-Benz stocks, that's the target, and you get some kind of examples

like on Friday it was 206 euros and on Saturday it was zero or something like this.

It's going to be there on Monday or so.

Okay?