And then I gave you kind of a little history excursion.

Most of the techniques we will see from now on are actually motivated by kind of posing

AI problems.

One of the wonderful things about AI is that it's really too hard for the current technology.

And that's been kind of an invariant.

The AI has always been too hard for the current technology because all the things we're solving,

we've talked about this, become computer science.

So AI is too hard for computer science, which is why it's a good source for computer science,

because we try new stuff.

And sometimes we succeed.

At the time in the 70s, looking at line graphs, which was kind of a simplified problem of

doing vision, something like that, intelligent people and cats and mice and so on can do

very, very well, but machines at the time couldn't.

So they said, we want to do vision.

What can we do reasonably with current technology?

And remember, in the 70s, computers were things that filled a room like this.

And you didn't have a terminal in front of you.

You had punch cards, which you could program, and then you would actually give to the operator,

which was the operating system, essentially was a human that took your punch cards and

put them into the machine and gave you a printout the next morning.

So that's where this comes from.

And the problem is too hard for that day's technology.

Why?

Because it leads to a big configuration problem.

So the analysis was that to understand these things, you have to show which of the edges

are convex and concave and where the boundary to open space lay.

And you can analyze this, and you basically have the configuration problem that you have

to kind of put marks on all the line segments and have them be consistent, meaning if I

have an edge here with a labeling, the same label or the fitting label should be on the

other all the connected.

So still, no problem.

Waltz discovered we have a finite number of these edges, and then you just have the configuration

problem of making them all fit.

Obviously you use a search technique.

At least at that point in time, nothing else existed essentially.

So what can I do?

Well, maybe iterative deepening search and maybe this and that.

And all of that was nowhere near a solution.

So the innovation step was, OK, let's invent constraint satisfaction problems, which he

did and then had a technology that actually worked in actually having this here, which

is essentially you have a couple of three-way intersections.

Here is one, there is one, there is one, and two-way intersections, each of which you can

actually think of as a variable, which can have one of, so you had two variables with

two intersection, two junction, which had six value, and you had the three intersections

which had those, what is it, five, nine, twelve values, and then you just had to basically

assign a value to every slot.

The constraints are, if you have these points, the line in a way gives you the constraint.

They have to match up.

If you program this and once you understand the idea of a constraint satisfaction problem,

you can solve configurations like this.