Okay, so welcome back to AI.

We are still in the part of AI where we talk about learning agents, in particular, learning from examples.

So, we are still in the part of AI, and we are still in the part of AI.

Whatever that means, the training set.

We have looked at various ways of doing that.

We have looked at regression and classification.

We have looked at support vector machines and neural networks.

And we have basically started looking at applications of these.

And one of them that we are going to look at today is probably the best thought of as in the context of a learning agent.

Remember that a learning agent.

The performance element here, and various other learning related stuff behind here.

The idea was that the learning stuff could actually go in and change the stuff that's in the performance element.

And we think about the performance element in say a utility based agents.

And the way we talked about that a while ago.

And the method was a Bayesian network.

And the way we learned a learning agent would be able to learn models.

And then for instance the Bayesian networks we talked about the alarms.

And then we talked about a graph that's easy to do.

And then we talked about the variables.

And the thing that I did last week I was, Oh you as the agent designer actually make them.

And if you do well, then it's a good model.

The better answer is, well the agent learns those.

And they interact with the world.

They start to try and walk or listen to alarms.

And then they make experiences and call the police because of a burglar alarm, because John called or something like this.

And it turned out to be all wrong.

So what we want to do today is we want to look at how could we learn?

How could we learn these things?

How could we learn parameters or tables?

And if you remember those, right, there were lots of little numbers in there.

And those of them, we think of those as parameters of the Bayesian networks that we want to learn.

And that's something we want to look at today.

Very relevant, you think about learning agents, learning utility based agents.

It's a relevant, that's one of the relevant applications.

Yes.

I would say that's still up to the designer.

You can think of doing heavy-duty learning, always trying to adjust your probabilities.

Which is probably fine while you're still learning how to write a bike or something like this.

While you're explicitly in a learning phase for writing a bike that's, I don't know, an afternoon or something like that.

Then you do heavy-duty updating and doing experiments like making a sharp curve or something like this.

And then later, the remaining 60 years of your life.

You're basically not learning that much about biking.

You're just using it.

So it might depend on how often you learn.

How heavily you invest in this part, cost computation and energy and all of those kind of things.

And how often you just basically just utilize the model you have.

A good learning agent would actually reflect or at least do something with these things.

But again, not in AI too.

There's a couple of things I want to say about this.

One is, and not there are other learning tasks, just in Bayesian networks.