chapter01

imminentpoppedΤεχνίτη Νοημοσύνη και Ρομποτική

23 Φεβ 2014 (πριν από 3 χρόνια και 6 μήνες)

94 εμφανίσεις



1

Artificial
Intelligence

CS 404


Berrin Yanikoglu

To Know


Basic history of AI


Know some of the important events or at least what
happend in different eras


Difficulty of defining intelligence and some of
the attempts


Fleeting nature of the definition


Difference of humanly/rational thinking/acting


Turing test


Rational agents






2

Course Info


Webpage:



http://people.sabanciuniv.edu/~berrin/cs404
/



for info, expectations, lecture notes, assignments, ...



Linked from SuCourse


3

On May 12th, 1997, the

best chess player in the

world, Gary

Kasparov,

lost a six
-
game chess

match to a computer

named “Deep Blue 2”


What was so significant

about this event?

Being able to program a computer to defeat a Grand Master level

chess player had been
a long
-
standing goal of the science of

artificial intelligence

-

and now it has been achieved

What is Artificial Intelligence?


Intelligence is difficult to define and understand
,
even for philosophers and psychologists who spend
their lives studying it. But this elusive quality is, to
many people, the
characteristic that sets humans
apart from other species


“What is intelligence, anyway?

It is only a word that people use
to name those unknown processes with which our brains solve
problems we call hard.

But whenever you learn a skill yourself,
you are less impressed or mystified when other people do the
same.


This is why the meaning of “intelligence” seems so elusive:

It
describes not some definite thing but only the momentary
horizon of our ignorance about how minds might work.”



-

Marvin Minsky, AI researcher

What is Artificial Intelligence?

Smart programs?



Not really. Studying what is possible and underlying theories
are very important.



How does a slow, tiny brain (biological or electrical) perceives,
understands, and manipulates a complex world?















6



Studying AI


Started out in 1950s


The Dartmouth meeting in 1956



Turned out much more difficult than anyone had
imagined



Currently encompasses a large variety of subfields,


from general areas such as perception and logical reasoning
to


specific tasks such as playing chess, writing poetry



bringing together
philosophy
,
logic
,
computer science
,
cognitive science and cognitive neuroscience




7

What is AI?

So it’s not surprising that defining artificial intelligence (AI) is hard. In general,
artificial intelligence is the field of science devoted to making

computers
perceive, reason, and act in ways that have, until now, been reserved for human
beings
.


4 types of historical definitions:

acting or thinking humanly







acting or thinking rationally

8

Thinking Humanly:
Cognitive Science


How do humans think?


Requires scientific theories of internal activities of the brain


What level of abstraction? ``Knowledge'' or ``circuits''?


How to validate?



Bring together computational models from AI and
experimental techniques from psycho
-
physics to model the
human mind



1) Predicting and testing behavior of human subjects (
Cognitive
Science
;
top
-
down)


2) Direct identification from neurological data (
Cognitive
Neuroscience
;
bottom
-
up)


Both approaches are now distinct from AI



Most of the machinery of the human mind is not available to
our

conscious experience



9

Thinking Logically


Let’s give up imitating the human thinking process, but try to
achieve logical thinking



Laws of formal logic to formalize the thinking process


Socrates is a man; All men are mortal => Socrates is mortal



Difficulties:


How to take informal knowledge and state in formal terms especially
when knowledge is less than 100% certain



Even problems with just a few dozen facts can b
e prohibitive


10

Acting Humanly
-

Turing Test


11

Acting Humanly
-

Turing Test


UCSD Turing Page:


http://cogsci.ucsd.edu/~asaygin/tt/ttest.html



Captchas


Tests to identify
humans from bots on the Internet,
to deny services to webcrawlers or
spammers


12



Acting Humanly
-

Turing Test



Chatterbots:


One of the most complex (and entertaining!)
chatter bots are at:
http://www.simonlaven.com/



choose complex chatterbots, choose Barry, go to
downloads and then choose CGI talk (
this year they
dont have the online CGI talk, and the downloaded
program didnt work for me)



Or this one works with no downloads: The
ALICE

Foundation


...


Human:
You sound pretty smart


ALICE:

In what way do I sound like that?

Human: never mind, hey, how is the weather?

ALICE:

Oh well thanks for trying to explain it to me anyway.
You've got my full attention. Sunny.

Human: where are you?

ALICE:

I am in Dr. Richard S. Wallace's computer in Oakland,
California. Where are you?

Human:
i

am in
istanbul
, do you know where it is?

ALICE:

Tell me a little about your life in
istanbul

do

me know
where it is.



...

13



Acting
Rationa
lly



Rationality
: ideal concept of intelligence



Rational behavior:

doing the right thing


that which is expected to maximize goal achievement,

given the
available information


Reflexes do not necessarily involve

thinking
---
e.g., blinking reflex
, but
they are rational


T
hinking

should be in the service of rational action




Caveat: computational limitations make perfect rationality
unachievable


design best program for given machine resources



We will emphasize rational agents in this course.




14

Why do we want artificial intelligence?



T
o
relieve our mental labour
, just as machines
relieved our physical

labour last century



I
t should
make the machines themselves easier
to use



I
t might give some insight into the workings of
our own minds









15

History of AI


1943 McCulloch
and
Pitts:
Artificial Neuron Model


1950 Turing's ``Computing Machinery and Intelligence'‘



1950s Early AI programs, including Samuel's checkers program,

Newell
&
Simon's
Logic Theorist

(proving theorems)
, Gelernter's

Geometry Engin
e, Shannon
and Turing writing chess programs


Shortage of computer times => Development of time sharing (=> DEC)


Creation of LISP (McCarthy)



1956
Dartmouth meeting:

``Artificial Intelligence''
coined


1965 Robinson's complete algorithm for logical reasoning

resolution method



1960s

Early development of knowledge
-
based systems
;
Minsky’s microworlds

(blocks
as home to various projects: vision, planning, nat. Lang. Understanding, ...)


ANALOGY

program (what is this figure most similar to?)


Algebra STUDENT program (one egg costs ... How much does twenty eggs cost?)




16

History of AI


1966
--
74
Dose of Reality



Very little domain knowledge:


Swithing from one domain to another, the programs failed miserably



AI discovers computational complexity


Early programs worked by representing the basic facts and trying out a
series of steps to solve the problem which was only tractable within
micro worlds; NP
-
completeness showed that scaling up to larger
problems was not always viable





Neural network research almost disappears



17

History of AI



1980
--
88 Expert systems industry booms


After all, they work, even if in limited domains



An
expert system

is a software designed to replicate the decision
-
making
process of a human expert, within a narrow topic. At the heart of every
expert system is a
knowledge base

representing ideas from the specific
field of expertise



A
knowledge
-
based system

derives knowledge from experts as well as
other sources like government regulations, statistical databases, company
guidelines, etc.



In practice,
the terms
expert system

and
knowledge
-
based system

are
often used interchangeably



While a database contains only facts, a knowledge base also contains a
system of if
-
then
rules

for determining and changing the relationships
between those facts



18

Digression:
Expert System
s


Expert systems are widely used in many different areas:



American Express uses one to automate checking for fraud and

misuses of its
no
-
limit credit card. This has to be done in 90

secs while the customer waits,
and the cost of an error can be high



DENDRAL, an expert system that examines the spectroscopic

analysis of an
unknown chemical compound and predicts its molecular structure



DEC’s XCON configures complex computer systems. It

reportedly does the
work of > 300 human experts, with fewer mistakes



PIERS, an expert system used to diagnose blood samples in

St Vincent Hospital,
Sydney



...



Current success is in reasonably narrow topics, eg mineral prospecting, medical
diagnoses, air traffic control, etc. But the real goal is to build something that has
a broad understanding of the world
-

which requires
common sense



History of AI



1988
--
93
Expert systems industry
start losing its power


Successful only in very narrow domains


Building a successful expert system is much more than simply
buying a reasoning system and filling it with rules



1985
--
95 Neural networks return to popularity



1988
--

With strengthened foundations, AI becomes hot
again
-

r
esurgence

of probabilistic and decision
-
theoretic
methods
, genetic algorithm
s
,
belief networks
,...


20

Current State


Which of the following can be done at present
!
?



Play a decent game of table tenni
s


Drive along a curving mountain road


Drive in the center of
Istanbul


Play a decent game of bridge


Discover and prove a new mathematical theorem


Write an intentionally funny story


Give competent legal advice in a specialized area of law


Translate spoken English into spoken Swedish in real time




21

Current State


Which of the following can be done at present
!
?



Play a decent game of table tenni
s


Drive along a curving mountain road


Drive in the center of
Istanbul


Play a decent game of bridge


Discover and prove a new mathematical theorem


Write an intentionally funny story


Give competent legal advice in a specialized area of law


Translate spoken English into spoken Swedish in real time




22

Current State



Limited domain speech/natural language understanding
programs



Chess playing programs (machines)



Medical expert systems challenging doctors



...




23

Artificial Intelligence and the Human
s



What does the advent of the intelligent machine mean for human
beings?



Are artificial intelligences just extensions of human intelligence?



When AARON creates a drawing, who is the artist, Cohen or AARON?




When expert systems make decisions, who is responsible? the user, the
programmer, the software company, or somebody else?



S
hould

we think of intelligent machines as some new sort of life, one
with which we must now share the world?



Could AIs be our evolutionary successors?



How will AI affect our own sense of self?



AI is beginning to force us to confront these hard
philosphical

questions…




Syllabus


http://people.sabanciuniv.edu/~berrin/cs404/syllabus.htm



In short (AIMA 3rd ed.)


Introduction:


Chapters 1
-
2


Problem Solving


Chapters 3
-
6


Knowledge and Reasoning


Chapters 7
-
9


Planning:


Skipped with just a brief overview


Uncertain Knowledge and Reasoning


Chapters 13,14,16; skip 15 and 17


Learning


overview + one classification method (decision trees)


Communicating, Perceiving, Acting


overview + one problem in computer vision


Conclusions


25

Seeing, Hearing and Understanding

An intelligent computer must be able to recogni
z
e its surrounding
environment and adapt to changes in it. To do this it must be able
to “see” and “hear” what’s going on


Computer vision

is the capability of a computer to mimic the ways
that human brains process and interpret light waves to produce a
model of reality. Though it’s very easy for people to do that, it’s
very difficult for computers to do build and update their models


Hearing, Seeing and Understanding

The ability of a computer to recogni
z
e the speech of a user and
take action based on the words spoken is called
speech recognition

or
voice recognition.
The computer matches spoken words against
stored speech patterns to determine what was said

Natural language processing


is the ability of a computer to build
knowledge representations corresponding to the
meaning

in
sentences made up of recogni
z
ed words. This is very difficult,
because human language is full of ambiguities, vagueness and
depends on a lot of commonsense knowledge of the world

Machine Learning


We’ve seen how difficult
collecting and maintaining knowledge

is. If there
was a lot, it could be impossible to do by hand



It would help if the machine could
build up its own knowledge from
experiences in the world
, like a child learning how to walk. The ability of the
machine to discover knowledge from observations of the world is called
machine learning





For example, some of the best game
-
playing programs learn from
past
experiences
. If a move pays off, a learning program is more likely to use that
(or similar moves) in future games. If a move results in a loss, the program will
remember to avoid similar moves

Robots
-

AI Embodied



Japanese companies such as Honda,


Fujitsu and Sony are racing to


develop humanoids





The Honda ASIMO (right) is a good


example





Improved walking stability over


earlier models





Smaller size is about marketing
-



and Robocup eligibility





Intelligence quite limited
-

some


commands sent by remote control





Simple voice recognition functions


trigger pre
-
programmed actions






Will cost about the same a luxury


car