So we started out yesterday with an overview of what we would do. The general thing in

the first semester is that we're going to use symbolic methods, which is essentially

that we will have representations of the world, of the state of the world, which we can write

down as symbols. Symbols being things like that, anything we can give a name, which means

it's kind of an object we can do things with, which I can communicate to you or put into

my pocket, give a name, store somewhere, get it back. That's kind of a very simple conceptualization

of what goes on in the brain, which is basically anything out there in the world has a physical

symbol which we can actually connect to some kind of a structure in our brain being in

some kind of a state. The idea of early in AI was that we would manipulate physical symbols

in a clever way and reach AI that way. Those techniques we're going to look at in this

semester. In the next semester, we're going to do the same thing, only that the states

in the world the world is in might not be known because our sensors are buggy, our actions

don't work, there are things we can't observe, and many things are non-deterministic anyway.

But still, essentially, we're doing state-based stuff. Then we'll go use the techniques we

take next semester to go to what is often called sub-symbolic AI, where you give up

this idea that you have these physical symbols that you can manipulate, but otherwise we

have kind of the symbols smeared over neural structures. That can do other things well.

A different conception, it's more oriented to learning from brains. Turns out it can

do different things. There are some things that symbolic AI can do well and other things

that machine learning and neural structures can do well. So you get to learn both. We

do the easy stuff, or at least the stuff I consider easy, in this semester.

We're going to start out from giving ourselves a framework to work and to program in. Programming

will be in Prolog. The conceptual framework will be intelligent agents. Then we're going

to go through a series of algorithms that become increasingly more complex and that

actually get more complex when we add more structure to our world representations. That's

going to be the recurring theme. The more we know about the world, the more guided our

algorithms can be. Sometimes the price we have to pay for more complex systems actually

pays off because our algorithms become more guided. That's kind of what's been happening

in symbolic AI. More and more and more knowledge about the world.

This step, we know nothing about the world except of course world states have that form.

They have a name, essentially. We can recognize that we've been in a state already, sometimes.

And then we use that paradigm for game playing. Then we add a little bit of structure to the

world representations, which we call factor representations. That drives interesting new

systems. Then we add a lot of structure to our world representations and we get into

the realm of logics. Essentially, from doing computation at the level of world states,

we actually do computations at the level of descriptions. We've got one meta level up

and gain efficiency that way. In the last January or so, we're going to add time because

the real world in AI has time, so we better do something about it. Good. We add uncertainty

to it in the summer semester, which essentially means throw away logic, which really has these

states true or false, get real and say, well, we know this with 70% probability. Kind of

smear the truth values into a whole interval. Means we have to kind of look at everything

again and then we use the statistics techniques to do machine learning.