Okay, sorry for being late. We were talking about agents. Remember agents are

these very abstract objects that live in an environment. They can perceive the

environment by collecting percepts and they can act on the environment using

actuators or effectors or something like that, however you want to call it. And

we've started looking at what kinds of agents there are, trying to give you an

idea, because this is the central metaphor we're going to use to embed the

theoretical stuff we're going to do in this course. So we're going to use this

kind of a diagram to concentrate on what goes on, wait, here, right? What is the

mechanism that drives the agent and we're assuming that the mechanism gets input from the

sensors, which are not part of the mechanism. That's a debatable choice, but that's what

the choice we're going to make. And the mechanism chooses actions which are then executed by

the actuators. And again, we're not going to look at the actuators themselves. In a

robotics class, which this is not, those would be very important topics and a tighter integration

of the deliberative mechanism and actions and the deliberative mechanism and the sensors

would be an important topic, but we're not going to do this. In AI2, we're going to

look at how uncertainty about your sensors, you know that your sensors are faulty 10% of the time

or something like this, really plays into your deliberations, the agent's deliberations, but

right now, we don't. We looked at agent programs, which is what we're going to develop, or at least

the theory for it, and we looked at rationality. Rationality being optimizing

the expected performance of an agent, given the agent configuration and the environment,

and most importantly, a performance measure. Those are the three big components which are fixed in

the optimization. And since rationality, that's important to realize since rationality only

optimizes the expected performance, that's kind of an escape hatch. The agent doesn't have to be

perfect, it doesn't have to know everything, it can still be rational.

It's basically optimizing within the boundaries of what the agent can perceive,

of what the agent can, what the actions are like and what the environment is.

We looked at the P's descriptions, performance measure, environment, agent, and

what's the S again? The sensors, yes.

And the last thing we looked into is classifying the environments. And by and large, we have these

always dichotomies, in one case a trichotomy, where we have the easy thing like static,

deterministic, fully observable, episodic and all of those kind of things on the left-hand side,

and not those things, non-static, non-observable and those kind of things.

On the right-hand side, we're going to look at the easy part

this semester, and we're going to building on that kind of one by one

go to the more difficult environments next semester.

I have a couple of examples here. If you have the game of solitaire as an environment, then as you

know, and I'm assuming you've played solitaire before, it is not fully observable. You have this

card stack on the left, and you do not know what's in the card stack. After the first round, you might,

but it's still not observable. There's a difference. You have a memory

in your agent mechanism of what it is, but it's not directly observable.

If you look at Bach-Gammon or something like this, Bach-Gammon is fully observable. You know everything,

but it's not deterministic because you roll a die.

So not all of your actions do. You may want to roll a six, but it turns out to be a three instead.

Solitaire is not episodic. Bach-Gammon also not. Internet

shopping is not fully observable. You can't see everything. Of course, I've ordered them so they

become more interesting and interesting. To the right, calling a taxi, just as one of the everyday

things we're doing, is an environment which has none of the good properties. It's not deterministic,

it's not episodic, it's not static. The environment changes even though you don't change it,

and possibly no other agent changes it. It might just start to rain or something like this.

That's a change in the environment. It's not discrete because there are uncountably many