All language subtitles for 14-Lecture 1 Segment 14 Designing Rational Agent

What do we do in this course: we design rational agents. What's that?

An agent is some entity that perceives

and acts.

What does that mean, well here's an agent,

an agent that's trying to get at some kind of apple. We'll come back to

this agent again later.

We can think about this agent with the following diagram--we'll draw these

agent diagrams from time to time.

There's an environment and there's an agent.

What's the difference? The environment is what you don't control,

the agent is what you do.

So, in particular, where you draw that boundary--you might think when

you're driving the agent is like the car, it can turn left and it can turn right. But the agent

is also your arms. You have to move the steering wheel.

If the car is not working right maybe you want to think of the car as the environment, or

maybe your arms not working right--you threw out your elbow driving so hard.

Now the line between the agent and the environment is in your head. You can send

the signal to your arm but who knows what your trick arm is going to do.

And so where we draw this abstraction, just like everything else in computer science,

depends on the problem we're solving

and what abstraction is safe.

Okay, so what can an agent do?

It can perceive,

we have percepts coming in.

We have actions going to the environment. What does the environment do?

It does whatever the environment does. The world does its thing.

And then the agent

takes what comes in and it formulates an action

using the question mark.

What the heck is the question mark? That's the all important agent function.

That's the thing that maps inputs (sensation)

to outputs (actuation).

So a rational agent chooses actions that maximizes its utilities. Since it doesn't

know what's gonna happen, like this agent's not really sure whether it's going to succeed,

we have to talk about expectations. The characteristics of the

percepts, the environment, the action space tell us how to solve the problem.

If an agent can see everything, that's one thing. If the agent can't, that's another thing.

So, for example, if you think about a video game. Sometimes you can see the whole board,

sometimes there's the fog-of-war, and in general, you don't always have full information.

The techniques you use are different.

This class is largely about general AI techniques

that apply to a variety of problems types,

and about learning to recognize when and how a problem

can be solved by one of those techniques, by thinking about what kinds of

percepts do I have, what kinds of environment do I have. Is it adversarial,

is it fully observed,

and so on. And as we go through this course you'll get more and more of these tools

to deal with more and more complicated environments.