We're still doing logic-based agents.

The idea there is that we have these world description languages and those basically

determine what we can do.

We've looked at procedures for automated theorem proving, which are the inference systems

that such an agent needs.

Remember, the agent has percepts, the agent has background knowledge, and for every percept,

like there's a breeze in cell 3-1, we want to get at the actual information that the

agent is interested in, namely not where there is a breeze, but where are the pits, or where

is the stench, where is the... and so on.

So we really need to have the entailment relation from this world description that knowledge

follows that that can be automated.

And the good thing about propositional logic was we had a really good inference procedures,

they were decision procedures, they come back after a finite time, they can be scaled

pretty well with DPLL, but the language for describing the world is extremely cumbersome.

Things can get big.

Even though we have a decision procedure, the descriptions get very big.

So we looked at first-order logic as a description language, where the descriptions are typically

relatively small.

Unfortunately, we do not have a decision procedure.

First-order logic is only semi-decidable, which means even though we have procedures

that if our conjecture is a theorem, or if our entailment is an actual entailment, then

they will come back after finite time, but if it's not, which is in the agent context

very often, we ask is there a pit here, and there actually isn't.

So that might actually take forever.

So what we're going to do next as the final chapter on logic-based agents is that we're

kind of looking at an intermediate.

The picture I want you to keep in mind, we kind of have a continuum almost of logical

systems, formal systems, and you can scale it by expressivity.

You essentially have first-order logic here, propositional logic down here.

There's stuff above, which we haven't talked about.

There's stuff below propositional logic, but nobody's interested in that.

And we have the decidability line here.

We have decidable logic below that line and undecidable ones above that line.

So it seems that, at least in this picture, there's a lot of space here.

Wouldn't it be nice if we could kind of go up and kind of make our logics as expressive

as possible and stop just before we hit undecidability?

Wouldn't it be very good to explore this space of logic, where we can do things but

not lose undecidability?

And that's really this space here.

The field is called knowledge representation and reasoning.

The big technology around this is what's called semantic technology.

You may have heard about the semantic web and all of those things, or the Google Knowledge

Graph if you want.

Bing has something similar and it's heavily used in industry.

That's what we're going to look at.

And if you think of it from a logic perspective, the logics here are called description logics.

Why?

Because as we will see, it's all about set descriptions.

And if you Google for knowledge, knowledge representation and reasoning, you'll find

things like this graphics, sometimes kind of tilted and said that is the knowledge stairs.