Original subtitles

- [Instructor] As we get ready

to build agent AI-based products and solutions,

first, let us get an understanding

of the many different types of agents

that we can build and design.

Can you imagine why we need different kinds of agent AI?

Let me give you a clue.

It is related to the type of problems

that you want to solve using agents

and how you want to design your agent to behave.

Let us look at the factors that contribute to the need

for different types of agent AI in the first place.

Do you want the agent to be reactive or proactive?

Is the agent in a fixed environment with no changes

or is it in dynamic environment?

Then it has to capture more factors from the environment

and perceived responses.

Are you building a single agent

or are you building a multi-agent system

with multiple agent AI debating and collaborating?

How exciting.

Let us look at different types of agent AI one at a time

with an example.

A simple action agent is designed to take action

if a condition is met in the environment.

It's simple.

For example, a simple action agent will turn on

the sprinkler if the heat goes

beyond a certain threshold in a factory.

A model-based agent AI has a model that provides knowledge

but perceives the state of the environment to take action.

An example of this is a vacuum cleaner

that picks up spilled waste in a mine.

This agent knows it has to look for spills

and keeps going,

and when it meets the condition,

it actuates the change to pick up the spill.

A goal-oriented agent aims to reduce the distance

between the action and the goal

so that the best possible way can be chosen

from multiple possibilities.

For example, a chatbot that schedules appointments

for patients efficiently can be set up

as a goal-oriented agent.

You can set the goal as reward points

for the agent AI upon successful feedback from customers

who confirm their appointments.

In this case, the agent has to take input

from the customer on several considerations

to pick the right time.

And if they say that time does not work,

it has to reflect upon its own decision and adapt

and change its approach till it reaches the goal

of the customer accepting the proposed appointment.

Did you see?

The agent AI is automating tasks,

but it is also human-centered in its design

to delight the customer.

A learning agent, as the name implies,

is capable of learning.

All AI is capable of learning,

but when you design an agent AI to be a learning agent,

you have to design it with learning capabilities

and the ability to self-evaluate and critique itself,

and an ability to improve.

Let's stop and think for a moment

how this is so different from AI

where we have trained the AI with new data

to provide learning to the AI.

In contrast to that,

a learning agent AI can learn on its own

as long as we design it

with the right capabilities to learn.

In the future lessons, we will learn

how agent AI can do reflection and improve itself

and how we can design to achieve this.

A utility agent uses a utility function to reach a goal,

but also optimizes other variables

such as cost or speed.

For example, an autonomous food delivery sidewalk robot,

like Kiwibot, can use a utility-based agent

to reach its goal to reach customers for delivery,

but can optimize its path

for speed, efficiency, and cost.