Intelligent Agents - PKU · 2 Intelligence Agents 2.1 Agents 2.2 Agent programs 2.3 Rationality 2.4...

Intelligent Agents

2

AI Slides (5e) c© Lin Zuoquan@PKU 2003-2019 2 1

2 Intelligence Agents

2.1 Agents

2.2 Agent programs

2.3 Rationality

2.4 Environments

2.5 Agent structures


Agents

An agent is an entity that perceives and acts in an environment

Agents include– animal agents

– human agents

– robotic agents (robots)– software agents (softbots)– – internet agents– – – crawler– – – webbot– – – email agent– – – search agent, etc.– – chatbots– – – Cortana/Siri/GAssistant/Waston/Alexa/FMessenger/· · ·

Single agent or usually multi-agents (so-called distributed AI)


Agents and environments

?

agent

percepts

sensors

actions

environment

actuators

An agent is anything that can be viewed as perceiving its environmentthrough sensors and acting upon that environment through actuators


Sensors and actuators

A sensor measures some aspect of the environment in a formthat can be used as input by an agent

– vision, hearing, touch– radio, infrared, GPS, wireless signals– active sensing: send out a signal (such as radar or ultrasound)

and sense the reflection of this signal off of the environment⇒ IoT (Internet of Things)

Perception provides agents with information about the world theyinhabit by interpreting the response of sensors

Actuators– hands, legs, vocal tract etc.– automated taxi: those available to a human drivere.g., accelerator, steering, braking and so on


Agent functions

An agent is completely specified by the agent function : maps frompercept histories to actions

f : P∗→ A

For any given class of environments and tasks, we seek theagent (or class of agents) with the best performance

Find a way to implement the rational agent function concisely


Agent programs

The agent program runs on the physical architecture to produce theagent function

agent = architecture+ program

program = algorithm + data

The agent program takes a single percept as input, keeps internalstate

function Skeleton-Agent( percept) returns action

persistent: memory, the agent’s memory of the world

memory←Update-Memory(memory, percept)

action←Choose-Best-Action(memory)

memory←Update-Memory(memory, action)

return action

The algorithm is described by the pseudocode


Another form of algorithm

A procedure for the skeleton-agent that takes a single percept asinput, keeps internal state

Procedure of Skeleton-Agent

Input: percept

Output: action

memory the agent’s memory of the world

1. memory←Update-Memory(memory, percept)

2. action←Choose-Best-Action(memory)

3. memory←Update-Memory(memory, action)

The algorithm is also described by the pseudocode


Algorithm

Recall: an algorithm is an explicit effective set of instructions for acomputing procedure

– may be used to find the answers to any of a given class of ques-tions– can be precisely defined by computational models, e.g., Turing ma-chine, etc.

Analysis of algorithms, independently of the particular implementationand input

– time complexity: speed in secondse.g., the worst Tworst(n) or the average Tavg(n)– space complexity: memory consumption in bytes

Complexity analyzes problems rather than algorithms


Algorithm

O-notation: asymptotic analysis

T (n) is O(f(n)) if T (n) ≤ kf(n) for some k, for all n > 0

O(nk) for some k: polynomial time/space, called P or “easiness”Otherwise, exponential time/space, e.g., O(2n), called NP , hardness

– class of nondeterministic (Turing machine) polynomial problemsguess a solution and then verify whether the guess is correct in

polynomial time– NP -complete problems (the hardest subclass of NP )– – co-NP (-complete) is the complement of NP (-complete)“yes” and“no” answers reversed

Theorem: either all the NP -complete problems are in P or none ofthem is


Algorithm

NP = P ??if you could try all the guesses at once, or you were very lucky

and always guess right the first time

#P : counting problems corresponding to decision problems in NP

decision problems have a yes-or-no answer, counting problemshave an integer answer: how many solutions

at least as hard as any NP problems

PSPACE: require a polynomial amount of space, even on a non-deterministic machine

it is believed that PSPACE-hard problems are worse than NP -complete problems although it could turn out thatNP = PSPACE,just as P = NP


The pseudocode

The pseudocode is a simple language to describe algorithms– similar to programming language like Java, C, Python or Lisp– informal to use mathematical formulas or ordinary English to

describe parts

• Persistent variables (global and local variables): a variable isgiven an initial value the first time a function is called and re-tains that value on all subsequent calls to the function. The agentprograms use persistent variables for memory. Variables have low-ercase italic names.

• Function as values: The value of a variable is allowed to be afunction. Functions and procedures have capitalized names

• Indentation is significant: the scope of a loop or conditional


The pseudocode

• Assignment: ”x← value” means that the right-hand side eval-uate to the left-hand side variable

– Destructuring assignment: ”x, y ← pair” means that theright-hand side evaluate to a two-element tuple, and the firstelement is assigned to x and the second to y. Or ”for eachx, y in pair do”

• if-then (-else): ”if c then · · · else · · ·” means that if the con-dition c is hold then doing something; otherwise doing somethingelse

• for each: ”for each x in c do” means that the loop is executedwith the variable x bound to successive elements of the collectionc. Or ”while c do” or even ”loop · · · ”

• Generators and yield: ”generator G(x) yields number” de-fines G as a generator function


The pseudocode

• Lists: [x, y, z], [first|rest]

• Sets: {x, y, z}

• Arrays start at 1: as in usual mathematical notation, not 0 asin Java and C

• /* the explanations can be given as remarks */


The code

The algorithms in the pseudocode can be implemented in Java, C/C++,Python, Lisp and Prolog etc.

The (Lisp) code for each topic is divided into four directories:– agents: code defining agent types and programs– algorithms: code for the methods used by the agent programs– environments: code defining environment types, simulations– domains: problem types and instances for input to algorithms

(Often run algorithms on domains rather than agents in environ-ments)


A coding (Lisp)

(setq joe (make-agent :name ’joe :body (make-agent-body)

:program (make-dumb-agent-program)))

(defun make-dumb-agent-program ()

(let ((memory nil))

#’(lambda (percept)

(push percept memory)

’no-op)))


Rationality

A rational agent is one that does right thing– to achieve the best performance“goals” specifiable by performance measuredefining a numerical value for any environment history

Rational action: whichever action maximizes the expected value ofthe performance measure given the percept sequence to date

Limited rationality: computational limitations make perfect ra-tionality unachievable⇒ design best program for given machine resources

Rational 6= omniscient– percepts may not supply all relevant information

Rational 6= clairvoyant– action outcomes may not be as expected

⇒ rational 6= successful


Vacuum-cleaner world

Example: Robot cleaner, say, iRobot Roombat

A B

Percepts: location and contents, e.g., [A,Dirty]

Actions: Left, Right, Suck, NoOp


A vacuum-cleaner agent

Percept sequence Action[A,Clean] Right

[A,Dirty] Suck

[B,Clean] Left

[B,Dirty] Suck

[A,Clean], [A,Clean] Right

[A,Clean], [A,Dirty] Suck... ...

function Reflex-Vacuum-Agent( [location,status]) returns an action

if status = Dirty then return Suck

else if location = A then return Right

else if location = B then return Left

What is the right function? Can it be implemented in a program?


Vacuum-cleaner rational agent

Fixed performance measure evaluates the environment sequence– one point per square cleaned up in time T ?– one point per clean square per time step, minus one per move?– penalize for > k dirty squares?


PEAS (Performance/Environment/Actuators/Sensors)

To design a rational agent, we must specify the task environment

E.g., an automated taxi (intelligent vehicle):

Performance measure??

Environment??

Actuators??

Sensors??


Example: automated taxi agent

To design a rational agent, we must specify the task environment

E.g., intelligent vehicle (an automated taxi):

Performance measure?? safety, destination, profits, legality, . . .

Environment?? streets, traffic, pedestrians, weather, . . .

Actuators?? steering, accelerator, brake, horn, speaker/display, . . .

Sensors?? video, accelerometers, gauges, engine sensors, GPS, . . .


Example: Internet shopping agent

Performance measure??

Environment??

Actuators??

Sensors??


Internet shopping agent

Performance measure?? price, quality, appropriateness, efficiency, . . .

Environment?? web sites, vendors, shippers, . . .

Actuators?? display to user, follow URL, fill in form, . . .

Sensors?? pages (text, graphics, scripts), . . .


Environments

Solitaire Backgammon Internet shopping TaxiObservable??Deterministic??Episodic??Static??Discrete??Single-agent??


Environments

Solitaire Backgammon Internet shopping TaxiObservable?? Yes Yes No NoDeterministic??Episodic??Static??Discrete??Single-agent??


Environments

Solitaire Backgammon Internet shopping TaxiObservable?? Yes Yes No NoDeterministic?? Yes No Partly NoEpisodic??Static??Discrete??Single-agent??


Environments

Solitaire Backgammon Internet shopping TaxiObservable?? Yes Yes No NoDeterministic?? Yes No Partly NoEpisodic?? No No No NoStatic??Discrete??Single-agent??


Environments

Solitaire Backgammon Internet shopping TaxiObservable?? Yes Yes No NoDeterministic?? Yes No Partly NoEpisodic?? No No No NoStatic?? Yes Semi Semi NoDiscrete??Single-agent??


Environments

Solitaire Backgammon Internet shopping TaxiObservable?? Yes Yes No NoDeterministic?? Yes No Partly NoEpisodic?? No No No NoStatic?? Yes Semi Semi NoDiscrete?? Yes Yes Yes NoSingle-agent??


Environments

Solitaire Backgammon Internet shopping TaxiObservable?? Yes Yes No NoDeterministic?? Yes No Partly NoEpisodic?? No No No NoStatic?? Yes Semi Semi NoDiscrete?? Yes Yes Yes NoSingle-agent?? Yes No Yes (except auctions) No

The environment type largely determines the agent design

The real world is (of course) partially observable, stochastic, sequen-tial, dynamic, continuous, multi-agent


Agent structures

Agents interact with environments through sensors and actuators

Agent Sensors

Actuators

En

viro

nm

ent

Percepts

Actions

?


Agent structures

Four basic types in order of increasing generality:– simple reflex agents– reflex agents with state– goal-based agents– utility-based agents

All these can be turned into– learning agents


Simple reflex agents

Agent

En

viro

nm

ent

Sensors

What action Ishould do nowCondition-action rules

Actuators

What the worldis like now


Simple reflex agents

function Simple-Reflex-Agent( percept) returns an action

persistent: rules, a set of condition-action rules

state← Interpret-Input(percept)

rule←Rule-Match(state, rules)

action← rule.action

return action


Example


if status = Dirty then return Suck

else if location = A then return Right

else if location = B then return Left

(setq joe (make-agent :name ’joe :body (make-agent-body)

:program (make-reflex-vacuum-agent-program)))

(defun make-reflex-vacuum-agent-program ()


(let ((location (first percept)) (status (second percept)))

(cond ((eq status ’dirty) ’Suck)

((eq location ’A) ’Right)

((eq location ’B) ’Left)))))


Model-based reflex agents

Agent

En

viro

nm

ent

Sensors

State

How the world evolves

What my actions do

Condition-action rules

Actuators


What action Ishould do now


Example


persistent: last A, last B, numbers, initially ∞

if status = Dirty then . . .

(defun make-reflex-vacuum-agent-with-state-program ()

(let ((last-A infinity) (last-B infinity))


(let ((location (first percept)) (status (second percept)))

(incf last-A) (incf last-B)

(cond

((eq status ’dirty)

(if (eq location ’A) (setq last-A 0) (setq last-B 0))

’Suck)

((eq location ’A) (if (> last-B 3) ’Right ’NoOp))

((eq location ’B) (if (> last-A 3) ’Left ’NoOp)))))))


Goal-based agents

Agent

En

viro

nm

ent

Sensors


State


What my actions do

Actuators


What it will be like if I do action A

Goals


Utility-based agents

Agent

Enviro

nm

ent

Sensors

How happy I will bein such a state

State


What my actions do

Utility

Actuators


What it will be likeif I do action A



Learning agents

Performance standard

Agent

En

viro

nm

ent

Sensors

Performanceelement

changes

knowledge

learning goals

Problemgenerator

feedback

Learning element

Critic

Actuators


The agent projects

Design and implement a simple but useful agent– Just right now on the progress with the new knowledge charpter-

by-charpter selectively– by a programming language which you are familiar with– in a software environment or platform– single agent or multi-agents (group)

Options:

• Internet agent (say, smart shopping, chatbot)

• Intelligent robot

• Intelligent motor

• Intelligent drone

• Coding algorithms and finding applications


Intelligent Agents - PKU · 2 Intelligence Agents 2.1 Agents 2.2 Agent programs 2.3 Rationality 2.4...

Documents

Transcript of Intelligent Agents - PKU · 2 Intelligence Agents 2.1 Agents 2.2 Agent programs 2.3 Rationality 2.4...