Welcome back everyone.
Recall from the previous lectures
we are interested in rational, intelligent agents.
And this will be our approach to AI in this course.
In this lecture, let's examine different aspects
related to rational agents.
And,
specifically, agent environments in which the agent evolves.
And how this affects our design of
intelligent, rational agents.
So let's dive in.
Let's define first what's an agent.
An agent is anything that can be viewed
as perceiving its environment through sensors,
and act upon that environment through actuators.
An agents can be represented with this diagram in which we
have the environment and the agent.
The agent box actually has some sensors in it
that sense the environment, and gets percepts
from the environment.
And the agent can have actuators that actually
act upon that environment.
Those generating actions are deciding what to do
and generating actions.
So the first part here allows the agent actually
to sense the world-- sense the world in which it evolves.
And the second part actually is about acting upon that world.
So this is acting.
So in the middle, you have a product in which
the agent would do what?
Would do thinking and deliberating.
Here the agent, given what it's observed or sensed
from the world, it will be thinking, deliberating,
or deciding-- another term for deciding what to do,
based on what it senses.
So in general, you could see that as a cycle for the agent
in which it perceives the environment, decides
think what to do, act, and so on and so forth.
So we have a kind of loop or cycle
happening here of perception, thinking, and acting
upon that environment.
The last aspect to mention about agent,
is that they are actually made up of two main components.
And these are the architecture and the system, or the program.
The architecture is the physical makeup.
This is the hardware of the agent.
This is the box that you have here,
and to which the set of sensors and actuators belongs to.
So this is the physical object that is linked to the agent.
Or this is the agent hardware itself.
So the program is actually the mind of the agent.
This is where the deliberation, the thinking is happening.
And this is happening actually here,
where you have-- you get the percept from the sensors,
from the physical, but then that agent
has to think about what to do, what
would be the world if he does, or that, et cetera.
So we have this way two main components
that are not only complimentary, they have to be compatible.
So the program is limited by the architecture make up,
and they have to be working together in harmony.
So basically, you can't, for example,
ask your intelligent agent to dance if it's just
a simple box on your table.
It would have to be some kind of compatibility between the two
pieces.
Example of agents include human agent,
that have as sensors the eyes, the ears, and other organs.
The actuators are the hands, legs, mouth
and other body parts.
Or body agents has as sensors the camera, infrared,
range finders, and other kinds of sensor
that allows them to move safely in their environment.
And the actuators are the various motors
that makes the robot moves.
Actually agents are everywhere.
This includes a simple thermostat
that maintains the room temperature; to your cell
phone that includes a lot of functionality
such as e-mails, web pages, the weather; to the vacuum
cleaner that you could program to go around your house
to clean your house; to robot that is as we said.
The Alexa Echo which is one of the most recent additions
that you could add to your house,
that can order pizza for you, play your favorite music,
maintain your shopping list, switch your lights off, order
a Uber car for you, et cetera.
A self-driving car, which is one of the most advanced AI agents
that actually is a car driving itself in the street,
perceiving its environment in terms of pedestrians, traffic,
roads, other cars, and acting upon that by moving from one
point to another one in the city.
And finally, the human agent that
actually the most intelligent agent out there.
Let's take a very simple example of a vacuum
cleaner agent that actually evolves in a small environment
in which we have two rooms, Room A and Room B,
and potentially some dirt in the rooms to clean up.
So the percept of the vacuum cleaner in this environment
is the location and the content of the room.
For example, if the vacuum cleaner
has two sensors, a sensor for dirt and a sensor for location,
it could create a percept as a pair.
A, dirty, for example, is an example
to say that-- is a percept to say that the room A is dirty.
The action that the vacuum cleaner accomplish
are going left, going right, suck the dirt,
or do nothing if everything is clean.
And finally, we are defining actually
an agent function that actually is
a mapping between the set of percepts
and the set of actions.
And this we could write in a table
that actually has a mapping between percepts and actions.
So for example, if A is clean, then go right.
If A is dirty, then suck the dirt.
If B is clean, go left, then check whether there
is dirt afterwards in the left.
And if B is dirty, then suck it up.
So we're going to-- this is a simple actually precept action
table pairs in which we have simple precepts, not
a sequence of words but just simple precepts on the left,
and the corresponding actions.
So given this, the agent will just
need to look up in this table, and depending
on what it is perceiving from the environment
through its sensors, it will define what it's going to do.
So we could actually make this table longer
by not only keeping the precepts single percept by themselves,
but having combinations of percepts or sequence
of percepts.
For example, if A was dirty, and then A, it became clean,
what to do?
Right?
So going to extend this table to not only
maintain the percept, but also sequences of percepts.
So we are interested in well-behaved agents.
And these are rational that are defined according to Russell
and Norvig as follows, "For each possible percept sequence,
a rational agent should select an action that
is expected to maximize its performance measure,
given the evidence provided by the percept sequence
and whatever background knowledge or ability knowledge
the agent has."
So we want this agent to be rational in the sense
that it's going to do its best, given
what it is observing and given whatever background knowledge
the agent has.
More specifically, we define rationality
relative to a performance measure.
And we judge rationality based on the performance measure
that defines the criterion of success.
The agent prior knowledge of the environment,
the possible actions that the agent can perform,
and finally the agent precept sequence to date.
So when we define rational agents,
we group these properties under the acronym PEAS, which
stands for Performance, Environment, Actuators,
and Sensors.
Which is actually the problem specifications
for the task environment that the rational agents
is meant to solve.
So let's define PEAS for some intelligent agents.
And let's start with the self-driving car.
So for self-driving car, passenger
would care about the full performance, which
could be the safety of the passenger, its comfort,
and also the time to destination.
We care about the legal drive, we
don't want a self-driving car that
crosses red lights or that harms pedestrian.
The environment of a self-driving car
is all the roads, the other cars, pedestrians, road signs,
and so on and so forth.
The actuators are everything that
makes the car move or interact with the environment,
such as the steering, accelerator, brake, signal,
horn, et cetera.
And finally, the sensors of the self-driving car
is actually [INAUDIBLE] with a lot of different sensors.
These include a camera, a sonar, a GPS, a speedometer,
an odometer, and a keyboard, along with other kinds
of sensors that allow the car really
to perceive the environment.
As a second example, let's take the example vacuum cleaner.
You're probably familiar with the iRobot Roomba series that
started about 10 years ago, to clean--
that you could put in your room to clean the floor.
The performance of the vacuum cleaner
could be measured in terms of cleanness
of the room, the efficiency-- that is the distance traveled
to clean-- or the battery life.
You care also about the performance
in terms of security.
Putting a device in your room that actually has eyes,
it has cameras, and it is potentially connected
to the wi-fi or internet, and you
don't want this to be hacked.
So security is a very important aspect of this device
before you add it to your house.
The environment includes anything in the room,
such as tables, wood floor, carpets, and different objects.
Your kid's lego could be swallowed by the machines so,
the environment is anything around
in the room that actually constitutes the environment
of the vacuum cleaner.
Finally, the actuators are actually
the hardware of the vacuum.
That includes the wheels, the different brushes,
along with the vacuum extractors to clean the dirt.
And finally, the Roomba, or the vacuum cleaner,
includes several sensors.
These include the camera, a dirt detector, a cliff sensor-- they
didn't want the Roomba to go down the stairs,
fall down the stairs-- a bump sensor because you don't want
the Roomba either to ruin all of your furniture
so it has to really detect whenever there is something
that the Roomba would hit.
Finally, the sensors include a camera, a dirt detection
sensor, the cliff sensors so the vacuum cleaner doesn't go down
the stairs, the bump sensor that allows you
to-- that allows the vacuum not to bump
into all of your furniture.
And finally an infrared wall sensor
that allows to detect the size and the limits of the room.
So we have seen two example of agents
with two examples of environment,
the street and the room.
We could actually categorize environment
into different types.
And this includes whether the environment
is fully observable, in which an agent's sensors give it
access to the full picture.
That is, the complete state of the environment
at each point in time.
This is versus partially observable,
such as the car, the self-driving car,
that do not have access, for example, to what's
going on in three blocks away, or what's
going on behind the truck.
An environment is deterministic when
the next state of the environment
is completely determined by the current state,
and the action executed by the.
Agent if the environment is deterministic
except for the actions of the other agents,
then the environment is called strategic.
So basically, a car can't predict
the behavior of other cars, or the traffic conditions.
So the opposite of deterministic is non-deterministic,
or what we call stochastic, in which
uncertainty about outcomes is modeled or quantified
in terms of probabilities.
We have this possibility of making
this action for the agent.
And although it can also be episodic or sequential,
in which the agent experience
is divided into atomic episodes.
Each episode consists of the agent perceiving and then
performing a single action.
And the choice of action in each episode
depends only on the episode itself.
The environment is called static vs. dynamic
if the environment is unchanged why the agent is deliberating
or thinking.
The environment is called semi-dynamic
if the environment itself does not
change with the passage of time, but the agent's performance
score does.
So everything is the same, but the score is going down
as time goes down.
This is for example, the case of chess with a clock.
The environment is discrete versus continuous
when a limited number of distinct, clearly defined,
percepts and actions is there.
For example, checkers is an example
of discrete environment, while self-driving car evolves
a continuous one, in terms of both the actions and also
the time.
A single agent-- we talk also about single agent
versus multi-agent environments, because whenever the agent is
by itself or with other agents.
We also talk about simple versus multi-agents.
In a simple agent, an agent is operating by itself
in an environment, there are no other agents operating there.
And finally known versus unknown,
in which the designer of the agent
may or may not have knowledge about the environment make up.
If the environment is unknown, then the agent
will need to know how it works in order to decide.
So this is for example, a case of a Roomba
that you place in a room, so when you design one,
the designer of the Roomba does not know in which room
the Roomba would be put.
So the Roomba would be using different kinds of sensors,
including an infrared wall sensor,
to define with other limits and the distance interval.
So if we don't know the room, the designer
has to embed some intelligence in the agent in order
to actually get familiar with the environment in which it
will evolve.
For more infomation >> ఈరాశివారు భోగిరోజున ఇలా చేస్తే సంవత్సరమంతాఅద్భుతం - Horoscope For Pisces - Sankranti - Meena Rasi - Duration: 1:58. 
Không có nhận xét nào:
Đăng nhận xét