Knowledge Representation - Computing

wonderfuldistinctΤεχνίτη Νοημοσύνη και Ρομποτική

16 Οκτ 2013 (πριν από 4 χρόνια και 24 μέρες)

71 εμφανίσεις

Knowledge Representation



The Edwin Smith papyrus


Title:


Instructions for treating a fracture of the cheekbone.


Symptoms:


If you examine a man with a fracture of the cheekbone, you will
find a salient and red fluxion, bordering the wound.


Diagnosis and prognosis:


Then you will tell your patient: "A fracture of the cheekbone. It is
an injury that I will cure."


Treatment:


You shall tend him with fresh meat the first day. The treatment
shall last until the fluxion resorbs.


Next you shall treat him with raspberry, honey, and bandages to
be renewed each day, until he is cured.


Searle’s Chinese Room


http://www.mind.ilstu.edu/curriculum/searle_chinese_room/searle_chinese_room.p
hp


Monolingual English speaker locked in a room,


Given


a large batch of Chinese writing


a second batch of Chinese script


a set of rules in English for correlating the second batch with the first batch.


A third batch of Chinese symbols and more instructions in English enable you "to
correlate elements of this third batch with elements of the first two batches" and instruct
you, thereby, "to give back certain sorts of Chinese symbols with certain sorts of shapes in
response."


Those giving you the symbols call


the first batch 'a script' [a data structure with natural language processing applications],


the second batch 'a story',


the third batch 'questions';


the symbols you give back the ‘answers to the questions’


the set of rules in English ‘the program‘


Can you be considered to understand Chinese ?


Approaches to Artificial Intelligence


Cognitive Scientists


think AI is the only serious way of finding out

how humans work


Engineers


want computers to do very smart things, quite independently of how
humans work


Strong AI


Want to build machines with human
-
like intelligence


Weak AI


Want to build machines that exhibit intelligent like behaviour but
believe machines will always be intellectually inferior to humans


Computer as a metaphor for the mind has been the dominant
approach for the last 60 years


Weak Vs. Strong AI
-

Philosopher John
Searle


WEAK AI


like Cognitive Science above

(I.e. about people)


uses machine representations and hypotheses to mimic human
mental function on a computer , but never ascribes those
mental properties to the machine.


STRONG AI


claim that machines programmed with the appropriate
behaviour, are having the same mental states as people would
who have the same behaviour


i.e. that machines can have MENTAL STATES.

What is Artificial Intelligence ?


Make machines behave as they do in the movies!


About the emulation of human behaviour


Make machines do things that would require intelligence if
done by humans

Boden
, M.A. (1977). Artificial Intelligence and Natural Man. Basic
Books, New York
.



Concerned with programming computers to perform tasks
presently done better by humans because they involve higher
mental processes such as perceptual learning, memory
organisation and judgemental reasoning

Minsky
, M.L. and Papert, S.A. (1969). Perceptrons. MIT Press,
Cambridge, MA.

What is Artificial Intelligence ?


Agreement that it is concerned with two things


Studying human thought processes


Representing these processes via machines


Computers


Robots


Artificial Intelligence is behaviour by a machine which if
performed by a human would be considered intelligent


“Artificial Intelligence is the study of how to make
computers do things at which, at the moment, people are
better”

Elaine
Rich,

Artificial Intelligence
, McGraw
-
Hill, 1983, p. 1

What are humans better at ?


Playing Games


Solving Puzzles


Common Sense Reasoning


Expert Reasoning


Understanding Language


Learning

AI

AI

Philosophy

Psychology

Anthropology

Neuro
-
Science

Linguistics

Computer Science

Intelligence


How does a human mind work ?


Can non
-
humans have minds ?


In terms of computing philosophy


Accept the idea that machines can do anything


Oppose this


machines incapable of sophisticated behaviour e.g.
love, creativity


OK in philosophy


How about in engineering/science terms ?

Intelligence


What does intelligence mean ?


Dictionary definition

1.
Someone’s ability to understand and learn things

2.
Intelligence is the ability to think and understand instead of
doing things by instinct or automatically



(Collins English Dictionary)

Someone’s

1
st

=> possessed by humans

2
nd

=> some flexibility, does not specify someone


Intelligence


what is thinking ?


Thinking


Activity of using your brain to consider a problem or create an
idea (Collins)


=> have to have a brain


Organ that allows learning and understanding


Is it possible for machines to achieve this ?


Can machines think ?

Why build intelligent machines?


Cheaper to build and maintain


Offer new possibilities


Better solutions to problems


Software relatively cheap to develop


Software can be changed easily

Why is AI relevant to us ?


Ai is concerned with how


knowledge is acquired and used,


information is communicated,


collaboration is achieved,


how problems are solved,


languages are developed, etc.


History of AI (Classical Period or Dark Ages)

mid 1940’s


mid 1950’s


Game Playing & Theorem Proving


State Space Searching


Alan Turing


McCulloch & Pitts


Von Neumann

Turing


Proposed
the concept of a universal
machine


Mathematical Tool equivalent to Digital Computer


Takes input and computes output via a Finite State
Machine







Must construct a different machine for each computation

Turing


Enigma Machine


Wrote the first program capable of playing a complete
chess game;


Reflections on intelligence:


Is
there
thought without
experience?


Is
there mind without communication?


Is
there
language without
living?


Is
there intelligence without
life?


i.e. can machines think?

Turing


Invented a game ‘Turning Imitation Game’


Can machines pass a behaviour test for intelligence


D
efined
the intelligent behaviour of a computer as
the
ability
to achieve the human
-
level performance in
cognitive
tasks


Predicted that by 2000
a
computer could
be programmed
to have a conversation with a human interrogator for
five
minutes
and would have a 30 per cent chance of
deceiving the interrogator
that it
was a
human

The Turing Test


Computer passes the test if interrogators cannot distinguish the machine from a
human on the basis of the answers to their questions.


Original Game:


First phase


Interrogator, a man and a woman are each placed in separate rooms and can communicate only via a
neutral medium such as a remote terminal.


Interrogator’s objective is to work out who is the man and who is the woman by questioning them.


Man should attempt to deceive the interrogator that he is the woman, while the woman has to
convince the interrogator that she is the woman.


Second phase


Man is replaced by a computer programmed to deceive the interrogator as the man did.


Programmed to make mistakes and provide fuzzy answers in the way a human would.


If the computer can fool the interrogator as often as the man did, we may say this computer has
passed the intelligent behaviour test.


Interrogator does not see, touch or hear the computer and is therefore not
influenced by its appearance or voice


Annually The Lobner Prize
-

http://www.loebner.net/Prizef/loebner
-
prize.html


McCulloch & Pitts


Proposed model
of artificial neural networks in which
each neuron was postulated
as being
in binary state, that
is, in either on or off
condition


D
emonstrated
that their neural network model was, in
fact, equivalent
to the Turing machine, and proved that any
computable
function could
be computed by some
network of connected
neurons


McCulloch & Pitts


S
timulated
both theoretical and
experimental work
to
model the brain in the laboratory.


Experiments clearly demonstrated
that the binary model
of neurons was not correct

Von Neumann


Part of the Manhattan Project


A
dviser
for the
Electronic Numerical
Integrator and
Calculator (ENIAC) project at the University
of Pennsylvania


First general purpose computer


Helped
to design the Electronic Discrete Variable
Automatic
Computer
(EDVAC), a stored program machine.


Binary rather than decimal

History of AI (Great Expectations)

(mid 50’s


late 60’s)


John McCarthy


Inventor of LISP


AdviceTaker


first complete knowledge
-
based system


Marvin Minsky


Focus on formal logic


Developed anti
-
logic outlook on knowledge representation and reasoning


Frames


McCulloch & Pitts


Continuing work on neural networks


Learning methods improved


Newell & Simon


General Problem Solver(GPS)


simulate human problem solving


Based on technique of means
-
end analysis


Choose and apply operators to achieve goal state


Focus on general problem solving, weak AI

History of AI (Great Expectations)

(mid 50’s


late 60’s)


Newell & Simon


Attempts to separate problem solving from data


Proposed that a problem to be solved could be defined in
terms of states.


Means
-
ends analysis was used to determine a difference
between the current state and the desirable state or the goal
state of the problem, and to choose and apply operators to
reach the goal state.


If the goal state could not be immediately reached from the
current state, a new state closer to the goal would be
established and the procedure repeated until the goal state was
reached.


The set of operators determined the solution plan.


History of AI (Great Expectations)

(mid 50’s


late 60’s)


Newell & Simon


GPS failed to solve complicated problems.


Program
was
based on
formal logic and therefore could
generate an infinite number of
possible operators
, which
is inherently inefficient.


The
amount of computer time
and memory
that GPS
required to solve real
-
world problems led to the project
being abandoned
.

History of AI (Reality Strikes)

(late 60’s


early 70’s)


AI researchers were developing general methods for
broad classes of problems


Programs contained little or no knowledge about
problem domain


Applied a search strategy by trying different combinations
of steps until right one found


Problems chosen too broad and too difficult

History of AI (Expert Systems)

(early 70’s


mid 80’s)


Realisation that problem domain must be restricted


Feigenbaum & Buchanan


DENDRAL program developed at Stanford to analyse chemicals


Incorporated knowledge of expert into program to perform at human
expert level


Shift from weak methods


Difficult


knowledge acquisition


Shortliffe


MYCIN


rule
-
based expert system for the diagnosis of infectious
diseases


Rules reflected uncertainty

History of AI (Making Machines Learn)

(mid 80’s
-

)


Expert Systems require more than rules


Rebirth of neural networks


Technology assisted


Evolutionary computing


Learning by doing


Ongoing since 70s


Natural intelligence is product of evolution


Based on computational models of natural selection and genetics


Simulate populate, evaluate performance, generate new population


Concept introduced by John Holland in 1975

History of AI (Making Machines Learn)

(1980’s onwards)


Knowledge Engineering


Computing with Words


Handling Uncertainty


Improved computational power


Improved cognitive modelling


The ability to represent multiple experts


Today


Topics in AI are much the same


Language now not so near the centre but it was at the
centre in the 70s


Roots now much further from logic and theorem proving


Neural nets and machine learning now more central


AI Approaches transitioned to main stream

What has AI achieved in real world ?


Robots in manufacturing


Diagnosis of illness: screen lab tests, diagnose blood
infections, identify tumors


Run airports: e.g. assign baggage gates, direct re
-
fuelling


Reasonable machine translation


Search systems like Google


efficient information
retrieval


Computer games


Deep Blue beat Kasparov in 1997



Key Lessons


Intelligence =

ability to learn and understand, to solve
problems and
to make
decisions
.


Goal of AI =
making
machines do
things that would
require intelligence if done by humans.


A
machine is thought intelligent if it can achieve human
-
level performance
in some
cognitive task.


To
build an intelligent machine, we have to
capture,
organise
and use human expert knowledge in some
problem area.

Negnevitsky M 2005, Artificial Intelligence, A guide to
intelligent systems design, 2
nd

Edition, Addison Welsey

Why Representation?


Humans need words (or symbols) to communicate
efficiently


Mapping of words to things is only possible indirectly


Create concepts that refer to things



What is knowledge representation?


What is representation?


Representation refers to a symbol or thing which represents
(’refers to’, ’stands for’) something else.


When do we need to represent?


We need to represent a thing in the natural world when we
don’t have, for some reason, the possibility to use the original
’thing’.


Example: Planning ahead


how will our actions affect the
world, and how will we reach our goals?


The object of knowledge representation is to express the
problem in computer
-
understandable form

Aspects of KR


Syntactic


Possible (allowed) constructions


Each individual representation is often called a sentence.


For example: color(my_car, red), my_car(red), red(my_car), etc.


Semantic


What does the representation mean (maps the sentences to the
world)


For example:


color(my_car, red)


??






‘my car is red’, ‘paint my car red’, etc.


Inferential


The interpreter


Decides what kind of conclusions can be drawn


For example: Modus ponens (P, P

Q, therefore Q)

Well
-
defined syntax/semantics


Knowledge representation languages should have precise
syntax and semantics.


You must know exactly what an expression means in
terms of objects in the real world.


Representation

of facts in the world

New

conclusions

Real World

Map to

KR language

Map back to

real world

Inference

Real World

Declarative vs. Procedural


Declarative knowledge (facts about the world)


A set of declarations or statements.


All facts stated in a knowledge base fall into this category of
knowledge.


In a sense, declarative knowledge tells us what a problem (or
problem domain) is all about


Procedural knowledge (how something is done)


Something that is not stated but which provides a mean of
dynamically (usually at run
-
time) arriving at new facts.


Declarative example


Information about items in a store



cheaper(coca_cola, pepsi)


tastier(coca_cola, pepsi)



if (cheaper(x,y) && (tastier(x,y) )

buy(x)

Procedural example


Shopping script:


Make a list of all items to buy


Walk to the shop


For each item on the list, get the item and add it to the
shopping basket


Walk to the checkout counter


Pack the items


Pay


Walk home

Types of knowledge


Domain knowledge:


What we reason about


Structural knowledge


Organization of concepts


Relational knowledge


How concepts relate



Strategic knowledge:


How we reason


At representation level
, rather than at implementation level


(e.g. at implementation level


control knowledge, for resolving
conflicting situations)

What is a Knowledge Representation?


“What is a Knowledge Representation?”


(Davis, Shrobe &Szolovits) AI Magazine, 14(1):17
-
33, 1993
http://groups.csail.mit.edu/medg/ftp/psz/k
-
rep.html


Defines the five roles the knowledge representation plays


Each role defines characteristics a KR should have


These roles provide a framework for comparison and
evaluating KRs

Role I: A KR is a Surrogate


A KR is used to model objects in the world.


Substitute for direct interaction with the world.


Cannot possibly represent everything in the world, a KR must
necessarily focus on certain objects and properties while
ignoring others.


As a result only objects and properties that are relevant to reasoning
are modeled.


Consequences:


Representation is not perfect


will have errors (at least by omission) and we may even introduce
new artifacts which not present


At least some unsound reasoning will occur

Role I: A KR is a Surrogate



The only complete accurate representation of an object is
the object itself.



All other representations are inaccurate.


Role II: A KR is a set of

Ontological Commitments


All representations are approximations to reality and they
are invariably imperfect.


Therefore
we need to focus on only some parts of the
world, and ignore the others.


Ontological commitments

determine what part of
the world we need to look at, and how to view it.

Role II: A KR is a set of

Ontological Commitments


The ontological commitments are accumulated in layers:


First layer


representation technologies.


For example, logic or semantic networks (entities and relations) vs.
frames (prototypes)


Second layer


how will we model the world.


Example from a frame
-
based system:


“The KB underlying INTERNIST system is composed of two basic types of
elements: disease entities and manifestations […] It also contains a hierarchy of
disease categories organised primarily around the concept of organ systems
having at the top level such categories as ’liver disease’, ’kidney disease’, etc”


Commits to model prototypical diseases which will be organised in a taxonomy
by organ failure


Third layer (conceptual)


which objects will be modelled.


What is considered a disease (abnormal state requiring cure), e.g.
alcoholism, chronic fatigue syndrome?

Role III : A KR is a Fragmentary

Theory of Intelligent Reasoning


“What is intelligent reasoning?”


The views of intelligence normally come from fields outside of
AI: mathematics, psychology, biology, statistics and economics.


Fragmentary


the representation typically incorporates only part of the
insight or belief that motivated it


that insight or belief is in turn only a part of the complex and
multi
-
faceted phenomenon of intelligent reasoning.

Role III : A KR is a Fragmentary

Theory of Intelligent Reasoning


There are three components:


the representation's fundamental conception of intelligent
inference

(What does it mean to reason intelligently?)


the set of inferences the representation
sanctions

(What can we infer from what we know?)


the set of inferences it
recommends


(What ought we to infer from what we know?)


Role IV: A KR is a medium for

efficient computation


The knowledge representation should make
recommended inferences efficient.


The information should be organized in such a way to
facilitate making those inferences.


There is usually a tradeoff between


the power of expression (how much can be expressed and
reasoned about in a language) and


how computationally efficient the language is.

Role V: A KR is a medium of

human expression


A representation is a language in which we communicate.


How well does the representation function as a medium of
expression?


How general is it?


How precise?


Does it provide expressive adequacy?


How well does it function as a medium of
communication?


How easy is it for us to ‘talk’ or think in that language?

Consequences of this KR


The spirit should be indulged, not overcome




KRs
should be used only in ways that they are intended to be
used, that is the source of their power.


Representation and reasoning are
intertwined


a
recommended method of inference is needed to make sense
of a set of facts.


Some researchers claim equivalence between KRs, i.e.
“frames are just a new syntax for first
-
order logic
”.



However, such claims ignore the important ontological
commitments and computational properties of a
representation.


All five roles of a KR
matter

Randall Davis, Howard Shrobe, Peter Szolovits MIT Lab

Requirements for KR languages


Representation adequacy


should to allow for representing all the required knowledge


Inferential adequacy


should allow inferring new knowledge


Inferential efficiency



inferences should be efficient


Clear syntax and semantics


unambiguous and well
-
defined syntax and semantics


Naturalness


easy to read and use

Some Knowledge Representation
Formalisms


Production systems, expert systems


Semantic networks


Frames


Case
-
based reasoning


Biologically inspired approaches


neural networks


genetic algorithms