Title: A New Frontier in ComputationComputation with Information Described in Natural Language
1A New Frontier in ComputationComputation with
Information Described in Natural Language Lotfi
A. Zadeh Computer Science Division Department
of EECSUC Berkeley SMC06 Taipei,
Taiwan October 9, 2006 URL http//www-bisc.cs.be
rkeley.edu URL http//www.cs.berkeley.edu/zadeh/
Email Zadeh_at_eecs.berkeley.edu
2PREAMBLE
- Humans have a remarkable capability to reason and
make decisions in an environment of imprecision,
uncertainty and partiality of knowledge, truth
and class membership. One of the principal
objectives of AI is formalization/mechanization
of this capability. - Computation with information described in natural
language, or NL-Computation for short, is a step
toward achievement of this objective.
3A HISTORICAL NOTE
- NL-Computation is a culmination of my
long-standing interest in exploring what I have
always believed to be a central issue in fuzzy
logic, namely, the relationship between fuzzy
logic and natural languages. My 1971 paper
entitled Quantitative Fuzzy Semantics, was the
first in a long series.
4WHAT IS NL-COMPUTATION?
Here are a few examples
- Trip planning
- I am planning to drive from Berkeley to Santa
Barbara, with stopover for lunch in Monterey.
Usually, it takes about two hours to get to
Monterey. Usually it takes about one hour to have
lunch. It is likely that it will take about six
hours to get from Monterey to Santa Barbara. At
what time should I leave Berkeley to get to Santa
Barbara, with high probability, before about 6 pm?
5CONTINUED
- Maximization
- f is a function from reals to reals described
as If X is small then Y is small if X is medium
then Y is large if X is large then Y is small.
What is the maximum of f? - Balls-in-box
- A box contains about twenty balls of various
sizes. Most are large. What is the number of
small balls? What is the probability that a ball
drawn at random is neither small nor large?
6CONTINUED
- Swedes and Italians
- Most Swedes are much taller than most Italians.
What is the difference in the average height of
Swedes and the average height of Italians? - Flight delay
- Usually most United Airlines flights from San
Francisco leave on time. What is the probability
that my flight will be delayed?
7THE BASIC PROBLEMEVALUATION OF A FUNCTION
- ?Zf(X, Y)
- IML(X) information about X expressed in
mathematical language - INL(X) information about X expressed in natural
language - MLComputation ?Z f(X,Y)
- NLComputation ?Z f(X,Y)
IML
IML
IML
IML
INL
INL
INL
INL
8EQUIVALENCE
- A natural language may be viewed as a system for
describing perceptions. - Example Most Swedes are tall
- Information described in natural language
perception-based information - INL (X) IPB(X)
9PRECISIATION
- A key concept in NL-Computation is that of
precisiation of meaning. Precisiation is a
prerequisite to computation. - NLComputation MNLComputation
- Precisiand may be viewed as a model of meaning.
- Cointension of precisiand measure of goodness
of model of meaning
precisiation
precisiend
precisand
precisiation
10BASIC POINTS
- Much of human knowledge is expressed in natural
language - A natural language is basically a system for
describing perceptions - Perceptions are intrinsically imprecise,
reflecting the bounded ability of sensory organs,
and ultimately the brain, to resolve detail and
store information - Imprecision of perceptions is passed on to
natural languages, resulting in semantic
imprecision - Semantic imprecision of natural languages places
the problem of computation with information
described in natural language beyond the reach of
existing techniques based on bivalent logic and
probability theory.
11BASIC STRUCTURE OF NL-COMPUTATION
Basically, NL-Computation is a system of
computation in which the objects of computation
are predicates and propositions drawn from a
natural language
COMPUTATION
PRECISIATION
NL
Pre1(p)
Pren(p)
p q
information
solution
reduction
ans(q/p)
question
Pre1(q)
Pren(q)
final solution
reduction to a standard problem
bridge from NL to MATH
(generalized-constraint-based)
(generalized-constraint-based)
12KEY IDEAS IN NL-COMPUTATION
- FUNDAMENTAL THESIS
- Information closed generalized constraint
- proposition is a carrier of information
- MEANING POSTULATE
- proposition closed generalized constraint
- predicate open generalized constraint
- In our approach, NL-Computation is reduced to
computation with generalized constraints, that
is, to generalized-constraint-based computation.
NL-Computation is based on fuzzy logic.
NL-Computation is closely related to Computing
with Words (CW)
13FUZZY LOGICKEY POINTS
- Humans have a remarkable capability to reason and
make decisions in an environment of imprecision,
uncertainty and partiality of knowledge, truth
and class membership. The principal objective of
fuzzy logic is formalization/mechanization of
this capability.
14WHAT IS FUZZY LOGIC?
- Fuzzy logic is not fuzzy logic
- Fuzzy logic is a precise logic of reasoning and
decision making based on information which is
imprecise, uncertain, incomplete and partially
true. - The principal distinguishing features of fuzzy
logic are - In fuzzy logic everything is, or is allowed to be
graduated, that is, be a matter of degree or,
equivalently fuzzy - In fuzzy logic everything is, or is allowed to be
granulated, with a granule being a clump of
points drawn together by indistinguishability,
similarity or proximity - Informal versions of Graduation and granulation
have a position of centrality in human cognition
15ANALOGY
- In bivalent logic, one writes and draws with a
ballpoint pen - In fuzzy logic, one writes and draws with a spray
pen which has an adjustable and precisely defined
spray pattern - This simple analogy suggests many mathematical
problems - What is the maximum value of f?
- Precisiation/imprecisiation principle
Y
X
16A BASIC CONCEPT IN NL-COMPUTATION GRANULAR VALUE
A
granular value of X
a
singular value of X
universe of discourse
- singular X is a singleton
- granular X isr A granule
- a granule is defined by a generalized constraint
- example
- X unemployment
- a 7.3
- A high
17KEY POINTS
- A granule may be viewed as a representation of
the state of imprecise, uncertain, incomplete
and/or partially true knowledge about the
singular value of X - Information described in natural language
granular information - Computation with information described in natural
language Granular Computing (Zadeh 1979, 1997,
1998 Lin 1998 Bargiela and Pedrycz 2005).
18GRANULATION OF A VARIABLE
- continuous quantized granulated
-
- Example Age
middle-aged
µ
µ
old
young
1
1
0
0
Age
quantized
Age
granulated
19GRANULATION OF A FUNCTION
Y
f
0
Y
medium x large
f (fuzzy graph)
perception
f f
summary
if X is small then Y is small if X is
medium then Y is large if X is large then Y
is small
0
X
20COMPUTATION WITH GRANULAR VALUESGRANULAR
COMPUTING
- Problem
- If I leave the hotel at about 10am and usually
it takes about an hour to get to the airport,
then at what time will I arrive at the airport? - Protoformal formulation
- ? Z X Y
- precisiation
- granular computing
usually(b)
a
21PRECISIATION OF approximately a, a
?
1
singleton
s-precisiation
0
x
a
?
1
cg-precisiation
interval
0
a
x
fuzzy interval
g-precisiation
?
type 2 fuzzy interval
0
a
x
fuzzy graph
22CONTINUED
probability distribution
g-precisiation
p
bimodal distribution
g-precisiation
0
x
23CONTINUED
?A
A
a is A
u
?B
B
b is B
u
?usually
usually is C
u
0
1
usually (b) p(v)
is usually
24CONTINUED
p(v)
p(v)p(v-u)
v
subject to
25EXTENSION PRINCIPLE (Zadeh 1965, 1975)
Yf(X) singular values
granulation
Yf(X) granular values
example f(X) is A g(X) is B
Bsupu(?A(f(u))
subject to
vg(u)
26MAMDANI
- Yf(X)
- granular f
- f if X is Ai then Y is Bi, i1, , n
- f is Ii Ai?Bi
- X is a
- Y is ?i?Ai(a)?Bi
27NL-CAPABILITY
- NL-capability capability to compute with
information described in natural language - Existing scientific theories do not have
NL-capability - In particular, probability theory does not have
NL-capability
28THE CONCEPTS OF PRECISIATION AND
COINTENSIVE PRECISIATION
29UNDERSTANDING VS. PRECISIATION
- Understanding precedes precisiation
- I understand what you said, but can you be more
precise - Beyond reasonable doubt
- Use with adequate ventilation
- Unemployment is high unemployment is
over 5 - Where do you draw the line? Paraphrase The US
Constitution is an invitation to argue over where
to draw the line - Where to draw the line is a key issue in legal
arguments
precisiation
30WHAT IS PRECISE?
PRECISE
v-precise
m-precise
- precise value
- p X is a Gaussian random variable with mean m
and variance ?2. m and ?2 are precisely defined
real numbers - p is v-imprecise and m-precise
- p X is in the interval a, b. a and b are
precisely defined real numbers - p is v-imprecise and m-precise
precise meaning
m-precise mathematically well-defined
31PRECISIATION AND IMPRECISIATION
v-imprecisiation
1
1
0
0
a
v-precisiation
a
x
x
m-precise
m-precise
v-precise
v-imprecise
1
If X is small then Y is small If X is medium then
Y is large If X is large then Y is small
v-imprecisiation
0
x
v-imprecise
v-precise
m-imprecise
m-precise
32v-IMPRECISIATION
v-imprecisiation
forced
deliberate
- forced V is not known precisely
- deliberate V need not be known precisely
- v-imprecisiation principle Precision carries a
cost. If there is a tolerance for imprecision,
exploit it by employing v-imprecisiation to
achieve lower cost, robustness, tractability,
decision-relevance and higher level of confidence
33IMPRECISIATION/ SUMMARIZATION OF FUNCTIONS
L
If X is small then Y is small If X is medium then
Y is large If X is large then Y is small
v-imprecisiation
M
summarization
S
0
S
M
L
If X is small then Y is small If X is medium then
Y is large If X is large then Y is small
(X, Y) is small ? small medium ? large large
? small
mm-precisiation
fuzzy graph
34SUMMARIZATION OF T-NORMS
S
M
L
S
M
L
- To facilitate the chore of an appropriate t-norm,
each t-norm should be associated with a summary
35APPROXIMATION VS. SUMMARIZATION
- summarization may be viewed as a form of
imprecisiation
y
y
approximation
0
0
x
x
y
summarization
0
x
36V-PRECISIATION
A
X variable
g-precisiation
a
X
s-precisiation
s-precisiation
7.3
unemployment
high
g-precisiation
- s-precisiation is used routinely in scientific
theories and especially in probability theory - defuzzification may be viewed as an instance of
s-precisiation
37PRECISIATION/IMPRECISIATION PRINCIPLE
(Zadeh 2005)
- a approximately a
- simple version
f(a) f(a)
Y
Y
X
X
38PRECISE SOLUTION
level set
undominated
39MODALITIES OF m-PRECISIATION
m-precisiation
mh-precisiation
mm-precisiation
machine-oriented
human-oriented
m-precisiation
l precisiand of l (Pre(l))
40PRECISIATION AND DISAMBIGUATION
- Examples
- Overeating causes obesity most of those who
overeat become obese Count(become.obese/ov
ereat) is most - Obesity is caused by overeating most of
those who are obese were overeating
Count(were.overeating/obese) is most
41PRECISIATION/ DISAMBIGUATION
- P most tall Swedes P(A) is?
POPULATION (Swedes)
tall Swedes
A
P1 most of tall Swedes P2 mostly tall
Swedes
mh-precisiation
P
mm-precisiation
P1 Count(A/tall.Swedes) is most P2 Count(tall.
Swedes/A) is most
mm-precisiation
42BASIC STRUCTURE OF DEFINITIONS
definiens
definiendum (idea/perception)
concept
mh-precisiand
mm-precisiand
mh-precisiation
mm-precisiation
cointension
cointension
cointension wellness of fit of meaning
Declining market with expectation of further
decline We classify a bear market as a 30
percent decline after 50 days, or a 13 percent
decline after 145 days. (Robert Shuster)
mh-precisiation
bear market
mm-precisiation
43EXAMPLES MOUNTAIN, CLUSTER, STABILITY
mh-precisiation
A natural raised part of the earths surface,
usually rising more or less abruptly, and larger
than a hill
mountain
mm-precisiation
?
44CONTINUED
mh-precisiation
A number of things of the same sort gathered
together or growing together bunch
cluster
mm-precisiation
?
- the concepts of mountain and cluster are
PF-equivalent, that is, have the same deep
structure
mh-precisiation
The capacity of an object to return to
equilibrium after having been displaced
stability
mm-precisiation
Lyapuonov definition
mm-precisiation
fuzzy stability definition
45GRANULATION REVISITED
- Granulation plays a key role in human cognition
- In human cognition, v-imprecisiation is followed
by mh-precisiation. Granulation is
mh-precisiation-based - In fuzzy logic, v-imprecisiation is followed by
mm-precisiation. Granulation is
mm-precisiation-based
- mm-precisiation-based granulation is a major
- contribution of fuzzy logic. No other logical
- system offers this capability
46DIGRESSIONEXTENSION VS. INTENSION
- extension and intension are concepts drawn from
logic and linguistics - basic idea
- object (name (attribute1, value1), ,
(attribute n, value n)) - more compactly
- object (name, (attribute, value))
- n-ary n-ary
name
attribute name
attribute value
object
attribute name
attribute value
47OPERATIONS ON OBJECTS
function
name-based
extensional
definition
object
intensional
definition
attribute-based
(algorithmic)
object (Michael, (gender, male), , (age,
25)) son (Michael) Ron
48PREDICATE (PROPERTY, CONCEPT, SET, MEMBERSHIP
FUNCTION, INPUT-OUTPUT RELATION)
- A predicate, P, is a truth-valued function
- Denotation of P D(P) XP(X)
- Extension of P Ext(P) D(P) if P(X) is
name-based - Intension of P Int(P) D(P) if P(X) is
attribute-based - P(X) open predicate (X is a free variable)
- P(a) closed predicate (X is a bound variable(P
is grounded))
U universe of discourse
generic object
X
P
D(P)
49EXAMPLE
U population
X
P bachelor
D(bachelor)
Ext (bachelor) Xbachelor (X) Int(bachelor)
50PRINCIPAL MODES OF DEFINITION
- Extension name-based meaning
- Intension attribute-based meaning
- Extensional Pu1, , un e-meaning of P
- Ostensive Pu, uk, ul o-meaning of P
- Intensional PuP(u), i-meaning of P
exemplars
51PROPOSITION (TENTATIVE)
- A proposition, p, is a sentence which may be
expressed as P(object). Equivalently, p, is a
closed predicate. Equivalently, p object is P - very simple example
- p Valentina is young
- example
- p most Swedes are tall
- P most
- object Count(tall.Swedes/Swedes)
- p most(Count(tall.Swedes/Swedes))
- i-meaning of p is associated with i-meaning of P
and i-meaning of object - Question D(most.tall.Swedes)?
young(Valentina), e-meaning
extensional
young(Age(Valentina), i-meaning
intensional
52SUMMARY
- p proposition or predicate
- Extension of p name-based meaning of p
- Intension of p attribute-based meaning of p
53THE CONCEPT OF COINTENSION
- p, q are predicates or propositions
- CI(p,q) cointension of p and q degree of match
between the i-meanings of p and q - q is cointensive w/n to p if GI(p, q) is high
- A definition is cointensive if CI(definiendum,
definiens) is high - In practice, CI(p,q) is frequently associated
with o-meaning of p and i-meaning of q - The o-meaning of the definiendum is
perception-based
54THE CONCEPTS OF COINTENSION AND RESTRICTIVE
COINTENSION
U universe of discourse
q
D(q)
p
D(p)
restriction
R
- CI(p,q)degree of proximity of D(p) and D(q)
- Cointension of q relative to pdegree of
subsethood of D(q) in D(p) - Restricted cointension U is restricted to R
55THE CONCEPT OF COINTENSIVE PRECISIATION
- Precisiation of a concept or proposition, p, is
cointensive if Pre(p) is cointensive with p. - Example bear market
- We classify a bear market as a 30 percent
decline after 50 days, or a 13 percent decline
after 145 days. (Robert Shuster) - This definition is clearly not cointensive
56KEY POINTS
- Precisiandmodel of meaning
- In general, p, may be precisiated in many
different ways, resulting in precisiands Pre1(p),
, Pren(p), each of which is associated with the
degree, CIi, of cointension of Prei(p), i 1, ,
n. In general, CIi is context-dependent. - Precisiation is necessary but not sufficient
- To serve its pupose, precisiation must be
cointensive - Cointensive precisiation is a key to
mechanization of natural language understanding
precisiation1
Pre1(p) C1
precisiation2
p
Pre2(p) C2
precisiationn
Pren(p) Cn
57AN IMPORTANT IMPLICATION FOR SCIENCE
- It is a deep-seated tradition in science to
employ the conceptual structure of bivalent logic
and probability theory as a basis for formulation
of definitions of concepts. What is widely
unrecognized is that, in reality, most concepts
are fuzzy rather than bivalent, and that, in
general, it is not possible to formulate a
cointensive definition of a fuzzy concept within
the conceptual structure of bivalent logic and
probability theory.
58EXAMPLES OF FUZZY CONCEPTS WHOSE STANDARD,
BIVALENT-LOGIC-BASED DEFINITIONS ARE NOT
COINTENSIVE
- stability
- causality
- relevance
- bear market
- recession
- mountain
- independence
- stationarity
- cluster
- grammar
- risk
- linearity
59A GLIMPSE INTO THE FUTURE
- To formulate a cointesive definition of a fuzzy
concept it is necessary to employ fuzzy logic - Replacement of bivalent-logic-based definitions
with fuzzy-logic-based definitions is certain to
take place but it will be a slow process - Fuzzy-logic-based definitions will be targeted
(customized)
60ANALOGY
S
M(S)
system
model
modelization
l
Pre(l)
lexeme
precisiand
precisiation
- input-output relation intension (test-score
function) - system analysis semantical analysis
(Freges Principle - of Compositionality)
- degree of match between M(S) and S cointension
- In general, it is not possible to construct a
cointensive - model of a nonlinear system from linear
components
61CHOICE OF PRECISIAND
- Cointension and tractability are contravariant
- To be tractable, precisiation should not be
complex - An optimal choice is one which achieves a
compromise between tractability and cointension
cointension
tractability
complexity
62THE KEY IDEA MEANING POSTULATE
- In NL-computation, a proposition, p, is
precisiated by expressing its meaning as a
generalized constraint. In this sense, the
concept of a generalized constraint serves as a
bridge from natural languages to mathematics.
NL
Mathematics
p
p (GC(p))
precisiation
generalized constraint
- The concept of a generalized constraint is the
- centerpiece of NL-computation
63TEST-SCORE SEMANTICS (ZADEH 1982)
Prinicipal Concepts and Ideas
- Test-score semantics has the same conceptual
structure as systems analysis - In test-score semantics, a lexeme, p, is viewed
as a composite constraint - Each constraint is associated with a test-score
function which defines the degree to which the
constraint is satisfied given the values of
constraint variables - Semantic analysis involves computation of the
test-score function associated with p in terms of
the test-score functions associated with f
components of p - the operation of composition and the resulting
test-score function constitute the meaning of p
64CONTINUED
- Constraints are represented as relations
- The system of relations associated with p
constitutes an explanatory database ED - ED may be viewed as a description of a possible
world - Test-score semantics has a much higher expressive
power than possible-world semantics
65EXAMPLE
- p young men like young women
- p most young men like most young women
- ED POPULATION Name Gender Age
- LIKES Name1 Name2 ?
- YOUNG Age ?
- MOST Proportion ?
POPULATION
ED (explanatory database)
men
women
possible world
Namei
Namej
young
66CONTINUED
- P likes mostly young women
- P(Namei) Count ((POPULATION Name Gender is F
Age is young)/ LIKES Name is Namei Name2
Gender is F Age) is most - ts(p) Count (POPULATIONName is P/ POPULATION
Name Gender is M Age is young) is most
women liked by Namei
young women
Namei
67THE CONCEPT OF A GENERALIZED CONSTRAINT
68PREAMBLE
- The centerpiece of fuzzy logic is the concept of
a generalized constraint. Constraints are
ubiquitous. In scientific theories,
representation of constraints is generally over
simplified. Over simplification of constraints is
a necessity because existing constrained
definition languages have a very limited
expressive power. The concept of a generalized
constraint is intended to provide a basis for
construction of a maximally expressive constraint
definition language which can also serve as a
meaning representation/precisiation language for
natural languages.
69GENERALIZED CONSTRAINT (Zadeh 1986)
- Bivalent constraint (hard, inelastic,
categorical)
X ? C
constraining bivalent relation
- Generalized constraint on X GC(X)
GC(X) X isr R
constraining non-bivalent (fuzzy) relation
index of modality (defines semantics)
constrained variable
r ? ? ? ? blank p v u rs
fg ps
bivalent
primary
- open GC(X) X is free (GC(X) is a predicate)
- closed GC(X) X is instantiated (GC(X) is a
proposition)
70CONTINUED
- constrained variable
- X is an n-ary variable, X (X1, , Xn)
- X is a proposition, e.g., Leslie is tall
- X is a function of another variable Xf(Y)
- X is conditioned on another variable, X/Y
- X has a structure, e.g., X Location
(Residence(Carol)) - X is a generalized constraint, X Y isr R
- X is a group variable. In this case, there is a
group, G (Name1, , Namen), with each member of
the group, Namei, i 1, , n, associated with an
attribute-value, hi, of attribute H. hi may be
vector-valued. Symbolically
71CONTINUED
- G (Name1, , Namen)
- GH (Name1/h1, , Namen/hn)
- GH is A (µA(hi)/Name1, , µA(hn)/Namen)
-
- Basically, GH is a relation and GH is A is a
fuzzy restriction of GH - Example
- tall Swedes SwedesHeight is tall
72SIMPLE EXAMPLES
- Check-out time is 1 pm, is an instance of a
generalized constraint on check-out time - Speed limit is 100km/h is an instance of a
generalized constraint on speed - Vera is a divorcee with two young children, is
an instance of a generalized constraint on Veras
age
73GENERALIZED CONSTRAINTMODALITY r
X isr R
r equality constraint XR is abbreviation of
X isR r inequality constraint X
R r? subsethood constraint X ? R r
blank possibilistic constraint X is R R is the
possibility distribution of X r v veristic
constraint X isv R R is the verity distributio
n of X r p probabilistic constraint X isp R R
is the probability distribution of X
Standard constraints bivalent possibilistic,
bivalent veristic and probabilistic
74CONTINUED
r bm bimodal constraint X is a random
variable R is a bimodal distribution r rs
random set constraint X isrs R R is the set-
valued probability distribution of X r fg fuzzy
graph constraint X isfg R X is a function
and R is its fuzzy graph r u usuality
constraint X isu R means usually (X is R) r g
group constraint X isg R means that R constrains
the attribute-values of the group
75PRIMARY GENERALIZED CONSTRAINTS
- Possibilistic X is R
- Probabilistic X isp R
- Veristic X isv R
- Primary constraints are formalizations of three
basic perceptions (a) perception of possibility
(b) perception of likelihood and (c) perception
of truth - In this perspective, probability may be viewed as
an attribute of perception of likelihood
76EXAMPLES POSSIBILISTIC
- Monika is young Age (Monika) is young
- Monika is much younger than Maria
- (Age (Monika), Age (Maria)) is much younger
- most Swedes are tall
- ?Count (tall.Swedes/Swedes) is most
X
R
X
R
R
X
77EXAMPLES PROBABILISITIC
- X is a normally distributed random variable with
mean m and variance ?2 - X isp N(m, ?2)
- X is a random variable taking the values u1, u2,
u3 with probabilities p1, p2 and p3, respectively - X isp (p1\u1p2\u2p3\u3)
78EXAMPLES VERISTIC
- Robert is half German, quarter French and quarter
Italian - Ethnicity (Robert) isv (0.5German
0.25French 0.25Italian) - Robert resided in London from 1985 to 1990
- Reside (Robert, London) isv 1985, 1990
79STANDARD CONSTRAINTS
- Bivalent possibilistic X ? C (crisp set)
- Bivalent veristic Ver(p) is true or false
- Probabilistic X isp R
- Standard constraints are instances of generalized
constraints which underlie methods based on
bivalent logic and probability theory
80GENERALIZED CONSTRAINTSEMANTICS
A generalized constraint, GC, is associated with
a test-score function, ts(u), which associates
with each object, u, to which the constraint is
applicable, the degree to which u satisfies the
constraint. Usually, ts(u) is a point in the unit
interval. However, if necessary, it may be an
element of a semi-ring, a lattice, or more
generally, a partially ordered set, or a bimodal
distribution. example possibilistic constraint,
X is R X is R Poss(Xu) µR(u) ts(u) µR(u)
81TEST-SCORE FUNCTION
- GC(X) generalized constraint on X
- X takes values in U u
- test-score function ts(u) degree to which u
satisfies GC - ts(u) may be defined (a) directly (extensionally)
as a function of u or indirectly (intensionally)
as a function of attributes of u - intensional definitionattribute-based
definition - example (a) Andrea is tall 0.9
- (b) Andreas height is 175cm µtall(175)0.9
Andrea is 0.9 tall
82CONSTRAINT QUALIFICATION
- p isr R means r-value of p is R
- in particular
- p isp R Prob(p) is R (probability
qualification) - p isv R Tr(p) is R (truth (verity)
qualification) - p is R Poss(p) is R (possibility
qualification) - examples
- (X is small) isp likely ProbX is small
is likely - (X is small) isv very true VerX is small
is very true - (X isu R) ProbX is R is usually
83GENERALIZED CONSTRAINT LANGUAGE (GCL)
- GCL is an abstract language
- GCL is generated by combination, qualification,
propagation and counterpropagation of generalized
constraints - examples of elements of GCL
- X/Age(Monika) is R/young (annotated element)
- (X isp R) and (X,Y) is S)
- (X isr R) is unlikely) and (X iss S) is likely
- If X is A then Y is B
- the language of fuzzy if-then rules is a
sublanguage of GCL - deduction generalized constraint propagation and
counterpropagation
84CONSTRAINTS
generalized constraints
primary constraints
standard constraints
- generalized X isr R , r possibilistic,
probabilistic, veristic, random
set, usuality, group, - primary possibilistic, probabilistic, veristic
- standard bivalent possibilistic, probabilistic,
bivalent veristic - existing scientific theories are based on primary
constraints
85PRECISIATION TRANSLATION INTO GCLBASIC
STRUCTURE
NL
GCL
p
p
precisiation
precisiand of p GC(p)
translation
generalized constraint
- annotation
- p X/A isr R/B GC-form of p
- example
- p Carol lives in a small city near San
Francisco - X/Location(Residence(Carol)) is R/NEARCity ?
SMALLCity
86v-PRECISIATION
s-precisiation
g-precisiation
- conventional (degranulation)
- a a
-
- approximately a
-
-
GCL-based (granulation)
precisiation
a
precisiation
X isr R
p
proposition
GC-form
common practice in probability theory
- cg-precisiation crisp granular precisiation
87PRECISIATION OF approximately a, a
?
1
singleton
s-precisiation
0
x
a
?
1
cg-precisiation
interval
0
a
x
fuzzy interval
g-precisiation
?
type 2 fuzzy interval
0
a
x
fuzzy graph
88CONTINUED
probability distribution
g-precisiation
p
bimodal distribution
g-precisiation
0
x
89EXAMPLE
- p Speed limit is 100 km/h
poss
cg-precisiation r blank (possibilistic)
p
speed
100
110
poss
g-precisiation r blank (possibilistic)
p
100
110
prob
g-precisiation r p (probabilistic)
p
speed
100
110
90CONTINUED
prob
g-precisiation r bm (bimodal)
p
100
110
120
speed
If Speed is less than 110, Prob(Ticket) is
low If Speed is between 110 and 120,
Prob(Ticket) is medium If Speed is greater than
120, Prob(Ticket) is high
91GC-BASED DEFINITION OF GRANULAR VALUE
- X is a singular value
- X is A granular value
- A is defined as a generalized constraint
- example
- X is small granular value
singleton
granule
fuzzy set
92GRANULAR COMPUTING (GrC) REVISITED
- The objects of computation in granular computing
are granular values of variables and parameters - Granular computing has a position of centrality
in fuzzy logic - Granular computing plays a key role in
precisiation and deduction - Informally
- granular computingballpark computing
93GRANULAR COMPUTING AND DEDUCTION
- The principal rule of deduction in fuzzy logic is
the Extension Principle (Zadeh 1965, 1975).
f(X) is A g(X) is B
subject to
94CONTINUED
- Generalized extension principle
-
- Zf(X,Y) singular values
- Zf(X,Y) granular values
extension
95EXAMPLE
- p most Swedes are tall
- p ?Count(tall.Swedes/Swedes) is most
- further precisiation
- X(h) height density function (not known)
- X(h)du fraction of Swedes whose height is in h,
hdu, a ? h ? b
96PRECISIATION AND CALIBRATION
- µtall(h) membership function of tall (known)
- µmost(u) membership function of most (known)
?height
?most
1
1
0
0
height
fraction
0.5
1
1
X(h)
height density function
0
h (height)
b
a
97CONTINUED
- fraction of tall Swedes
- constraint on X(h)
is most
granular value
98DEDUCTION
q How many Swedes are short q is ?
Q deduction is
most given
is ? Q needed
- Frege principle of compositionalityprecisiated
version - precisiation of a proposition requires
precisiations - (calibrations) of its constituents
99CONTINUED
deduction
given
is ? Q needed
solution
subject to
100CONTINUED
q What is the average height of Swedes? q
is ? Q deduction is most
is ? Q
101LOOKAHEAD--PROTOFORMAL DEDUCTION
- Example
- most Swedes are tall 1/n?Count(GH is R)
is Q
Height
102PROTOFORMAL DEDUCTION RULE
1/n?Count(GH is R) is Q
1/n?Count(GH is S) is T
?i µR(hi) is Q
?i µS(hi) is T
µT(v) suph1, , hn(µQ(?i µR(hi))
subject to
v ?i µS(hi)
values of H h1, , hn
103PROTOFORM LANGUAGE AND PROTOFORMAL DEDUCTION
PFL
104THE CONCEPT OF A PROTOFORM
PREAMBLE
- As we move further into the age of machine
intelligence and automated reasoning, a daunting
problem becomes harder and harder to master. How
can we cope with the explosive growth in
knowledge, information and data. How can we
locateand infer fromdecision-relevant
information which is embedded in a large
database. - Among the many concepts that relate to this
issue there are four that stand out in
importance search, precisiation and deduction.
In relation to these concepts, a basic underlying
concept is that of a protoforma concept which is
centered on the confluence of abstraction and
summarization
105WHAT IS A PROTOFORM?
- protoform abbreviation of prototypical form
- informally, a protoform, A, of an object, B,
written as APF(B), is an abstracted summary of B - usually, B is lexical entity such as proposition,
question, command, scenario, decision problem,
etc - more generally, B may be a relation, system,
geometrical form or an object of arbitrary
complexity - usually, A is a symbolic expression, but, like B,
it may be a complex object - the primary function of PF(B) is to place in
evidence the deep semantic structure of B
106CONTINUED
object space
object p
protoform space
summary of p
protoform
summarization
abstraction
S(p)
A(S(p))
PF(p)
- PF(p) abstracted summary of p
- deep structure of p
- protoform equivalence
- protoform similarity
107PROTOFORMS
object space
protoform space
PF-equivalence class
- at a given level of abstraction and
summarization, objects p and q are PF-equivalent
if PF(p)PF(q) - example
- p Most Swedes are tall Count (A/B) is Q
- q Few professors are rich Count (A/B) is Q
108EXAMPLES
instantiation
- Monika is young Age(Monika) is young A(B) is C
- Monika is much younger than Robert
- (Age(Monika), Age(Robert) is much.younger
- D(A(B), A(C)) is E
- Usually Robert returns from work at about 615pm
- ProbTime(Return(Robert) is 615 is usually
- ProbA(B) is C is D
abstraction
usually
615
Return(Robert)
Time
109CONTINUEDEXTENSION VS INTENSION
Q As are Bs
(attribute-free extension)
Count(GH is A) is Q
(attribute-based intension)
110EXAMPLES
gain
Alan has severe back pain. He goes to see a
doctor. The doctor tells him that there are two
options (1) do nothing and (2) do surgery. In
the case of surgery, there are two possibilities
(a) surgery is successful, in which case Alan
will be pain free and (b) surgery is not
successful, in which case Alan will be paralyzed
from the neck down. Question Should Alan elect
surgery?
2
1
0
option 2
option 1
Y
Y
object
i-protoform
X
0
X
0
111PROTOFORMAL DEDUCTION
NL
GCL
PFL
p q
p q
p q
precisiation
summarization
precisiation
abstraction
WKM
DM
r
World Knowledge Module
a
answer
deduction module
112PROTOFORMAL DEDUCTION
- Rules of deduction in the Deduction Database
(DDB) are protoformal - examples (a) compositional rule of inference
-
X is A (X, Y) is B Y is AB
symbolic
computational
(b) Extension Principle
X is A Y f(X) Y f(A)
Subject to
symbolic
computational
113RULES OF DEDUCTION
- Rules of deduction are basically rules governing
generalized constraint propagation - The principal rule of deduction is the extension
principle -
X is A f(X,) is B
Subject to
computational
symbolic
114GENERALIZATIONS OF THE EXTENSION PRINCIPLE
information constraint on a variable
f(X) is A g(X) is B
given information about X
inferred information about X
subject to
115CONTINUED
f(X1, , Xn) is A g(X1, , Xn) is B
Subject to
(X1, , Xn) is A gj(X1, , Xn) is Yj , j1,
, n (Y1, , Yn) is B
Subject to
116EXAMPLE OF DEDUCTION
- p Most Swedes are much taller than most Italians
- q What is the difference in the average height
of Swedes and Italians? - Solution
- Step 1. precisiation translation of p into GCL
- S S1, , Sn population of Swedes
- I I1, , In population of Italians
- gi height of Si , g (g1, , gn)
- hj height of Ij , h (h1, , hn)
- µij µmuch.taller(gi, hj) degree to which Si is
much taller than Ij
117CONTINUED
- Relative ?Count of Italians in relation to
whom Si is much taller - ti µmost (ri) degree to which Si is much
taller than most Italians - v Relative ?Count of Swedes who are
much taller than most Italians - ts(g, h) µmost(v)
- p generalized constraint on S and I
- q d
118CONTINUED
- Step 2. Deduction via Extension Principle
subject to
119DEDUCTION PRINCIPLE
- Precisiate query
- Precisiate query-relevant information
- Employ constraint propagation (Extension
Principle) to deduce the answer to query - example
- q What is the average height of Swedes?
- Assume that P is a population of Swedes,
P(Name1, , Namen), with hiHeight(Namei), i1,
, n.
120CONTINUED
- q (h1hn)
- (qri) I Most Swedes are tall
- I (µtall(h1)µtall(hn) is most
- GC(h) (µmost( (?iµtall(hi)) , h (hi, ,
hn)
121CONTINUED
- constraint propagation
- (µmost( (?iµtall(hi))
-
- Ave(h) ?ihi
- Extension Principle
- (µAve(h)(v) suph(µmost( ?iµtall(hi)) ,
(h1hn) - subject to
- v ?ihi
122DEDUCTION PRINCIPLEGENERAL FORMULATION
- Point of departure question, q
- Data D (X1/u1, , Xn/un)
- ui is a generic value of Xi
- Ans(q) answer to q
- If we knew the values of the Xi, u1, , un, we
could express Ans(q) as a function of the ui - Ans(q)g(u1, ,un) u(u1, , un)
- Our information about the ui, I(u1, , un) is a
generalized constraint on the ui. The constraint
is defined by its test-score function - f(u)f(u1, , un)
123CONTINUED
- Use the extension principle
subject to
124MODULAR DEDUCTION DATABASE
POSSIBILITY MODULE
PROBABILITY MODULE
FUZZY ARITHMETIC MODULE
agent
SEARCH MODULE
FUZZY LOGIC MODULE
EXTENSION PRINCIPLE MODULE
125PROBABILITY MODULE
126THE CONCEPT OF BIMODAL DISTRIBUTION (ZADEH 1979)
X isbm R
bimodal distribution
random variable
- A bimodal distribution is a collection of ordered
pairs of the form - R (P1, A1), , (Pn, An)
- or equivalently
- ?i(Pi \Ai) , i1, , n
- where the Pi are fuzzy probabilities and the Ai
are fuzzy sets
127CONTINUED
- Special cases
- The Pi are crisp the Ai are fuzzy
- The Pi are fuzzy the Ai are crisp
- The Pi are crisp the Ai are crisp
- The Demspter-Shafer theory of evidence is
basically a theory of crisp bimodal distributions
128EXAMPLE FORD STOCK
- I am considering buying Ford stock. I ask my
stockbroker, What is your perception of the
near-term prospects for Ford stock? He tells me,
A moderate decline is very likely a steep
decline is unlikely and a moderate gain is not
likely. My question is What is the probability
of a large gain?
129CONTINUED
- Information provided by my stockbroker may be
represented as a collection of ordered pairs - Price ((unlikely, steep.decline),
- (very.likely, moderate.decline),
- (not.likely, moderate.gain))
- In this collection, the second element of an
ordered pair is a fuzzy event or, equivalently, a
possibility distribution, and the first element
is a fuzzy probability. - The importance of the concept of a bimodal
distribution derives from the fact that in the
context of human-centric systems, most
probability distributions are bimodal
130BIMODAL DISTRIBUTIONS
- Bimodal distributions can assume a variety of
forms. The principal types are Type 1, Type 2 and
Type 3. Type 1, 2 and 3 bimodal distributions
have a common framework but differ in important
detail
131BIMODAL DISTRIBUTIONS (Type 1, 2, 3)
U
A2
A1
An
A
- Type 1 (default) X is a random variable taking
values in U - A1, , An, A are events (fuzzy sets)
- pi Prob(X is Ai) , i 1, , n
- ?ipi is unconstrained
- pi is Pi (granular probability)
- BMD bimodal distribution ((P1, A1), , (Pn,
An)) - X isbm (P1\A1
Pn\An) - Problem What is the probability, p, of A? In
general, this probability is
fuzzy-set-valued, that is, granular
132CONTINUED
- Type 2 (fuzzy random set) X is a
fuzzy-set-valued random variable with
values A1, , An (fuzzy
sets) - pi Prob(X Ai), i 1, , n
- BMD X isrs (p1\A1 pn\An)
- ?ipi 1
- Problem What is the probability, p, of A? p is
not definable. What are definable are (a) the
expected value of the conditional
possibility of A given BMD, and (b) the
expected value of the conditional necessity
of A given BMD
133CONTINUED
- Type 3 (augmented random set Dempster-Shafer)
- X is a set-valued random variable taking the
values X1, , Xn with respective probabilities
p1, , pn - Yi is a random variable taking values in Ai, i
1, , n - Probability distribution of Yi in Ai, i 1, ,
n, is not specified - X isp (p1\X1pn\Xn)
- Problem What is the probability, p, that Y1 or
Y2 or Yn is in A? Because probability
distributions of the Yi in the Ai are not
specified, p is interval-valued. What is
important to note is that the concepts of upper
and lower probabilities break down when the Ai
are fuzzy sets
134IS DEMPSTER SHAFER COINTENSIVE?
- In applying Dempster Shafer theory, it is
important to check on whether the data fit Type 3
model. - Caveat In many cases the cointensive
(well-fitting) precisiand (model) of a problem
statement is bimodal distribution of Type 1
rather than Type 3 (Demspter-Shafer)
Bimodal Type 1
precisiation
NL description of problem
precisiation
Bimodal Type 2
precisiation
Dempster-Shafer
Bimodal Type 3
135BASIC BIMODAL DISTRIBUTION (BMD) (Type
1)(PERCEPTION-BASED PROBABILITY DISTRIBUTION)
X is a real-valued random variable
probability
P3
g
P2
P1
X
0
A2
A1
A3
BMD P(X) Pi(1)\A1 Pi(2)\A2
Pi(3)\A3 Prob X is Ai is Pj(i)
P(X) low\small high\medium low\large
136INTERPOLATION OF A BASIC BIMODAL DISTRIBUTION
(TYPE 1)
P
g(u) probability density of X
p2
p
p1
pn
X
0
A1
A2
A
An
pi is Pi granular value of pi , i1, , n (Pi ,
Ai) , i1, , n are given A is given (?P, A)
137INTERPOLATION MODULE AND PROBABILITY MODULE
Prob X is Ai is Pi , i 1, , n Prob X is
A is Q
subject to
138EXAMPLE
- Probably it will take about two hours to get from
San Francisco to Monterey, and it will probably
take about five hours to get from Monterey to Los
Angeles. What is the probability of getting to
Los Angeles in less than about seven hours? - BMD (probably, 2) (probably, 5)
X
Y
Z XY
w
v
u
139CONTINUED
query is ?A
qri
subject to
140TEST PROBLEMS (PROBABILITY THEORY)
- X is a real-valued random variable. What is known
about X is a(usually X is much larger than
approximately a b usually X is much smaller than
approximately b, where a and b are real numbers
with a lt b. What is the expected value of X? - X and Y are random variables. (X,Y) takes values
in the unit circle. Prob(1) is approximately 0.1
Prob(2) is approximately 0.2 Prob(3) is
approximately 0.3 Prob(4) is approximately 0.4.
What is the marginal distribution of X?
Y
1
4
0
X
2
3
141CONTINUED
- function if X is small then Y is large
- (X is small, Y is large)
- probability distribution low \ small low \
medium high \ large - Count \ attribute value distribution 5 \ small
8 \ large - PRINCIPAL RATIONALES FOR F-GRANULATION
- detail not known
- detail not needed
- detail not wanted
142OPERATIONS ON BIMODAL DISTRIBUTIONS
P(X) defines possibility distribution of g
problem a) what is the expected value of X
143EXPECTED VALUE OF A BIMODAL DISTRIBUTION
Extension Principle
subject to
144PERCEPTION-BASED DECISION ANALYSIS
ranking of bimodal probability distributions
PA
0
X
PB
0
X
maximization of expected utility ranking of
fuzzy numbers
145USUALITY CONSTRAINT PROPAGATION RULE
X random variable taking values in U g
probability density of X
X isu A Prob X is B is C
X isu A
Prob X is A is usually
subject to
146PROBABILITY MODULE (CONTINUED)
X isp P Y f(X) Y isp f(P)
Prob X is A is P Prob f(X) is B is Q
X isp P (X,Y) is R Y isrs S
X isu A Y f(X) Y isu f(A)
147PNL-BASED DEFINITION OF STATISTICAL INDEPENDENCE
Y
contingency table
L
?C(M/L)
L/M
L/L
L/S
3
M
?C(S/S)
M/M
M/S
M/L
2
S
X
S/S
S/M
S/L
1
0
1
2
3
S
M
L
?C (M x L)
? (M/L)
?C (L)
- degree of independence of Y from X
- degree to which columns 1, 2, 3 are identical
PNL-based definition
148WHAT IS A RANDOM SAMPLE?
- In most cases, a sample is drawn from a
population which is a fuzzy set, e.g., middle
class, young women, adults - In the case of polls, fuzziness of the population
which is polled may reflect the degree
applicability of the question to the person who
is polled - example (Atlanta Constitution 5-29-95)
- Is O.J. Simpson guilty?
- Random sample of 1004 adults polled by phone.
- 61 said yes. Margin of error is 3
- to what degree is this question applicable to a
person who is n years old?
149USUALITY SUBMODULE
150CONJUNCTION
X is A X is B X is A B
X isu A X isu B X isr A B
- determination of r involves interpolation of a
bimodal distribution
151USUALITY CONSTRAINT
X isu A X isu B X isp P (A B) ispv Q
X is A X is B X is A B
g probability density function of X ?(g)
possibility distribution function of g
subject to
subject to
152USUALITY QUALIFIED RULES
X isu A X isun (not A)
X isu A Yf(X) Y isu f(A)
153USUALITY QUALIFIED RULES
X isu A Y isu B Z f(X,Y) Z isu f(A, B)
154SUMMATION
- The concept of GC-computation is the centerpiece
of NL-computation. The point of departure in
NL-computation is the key idea of representing
the meaning of a proposition drawn from a natural
language, p, as a generalized constraint. This
mode of representation may be viewed as
precisiation of p, with the result of
precisiation being a precisiand, p, of p. Each
precisiand is associated with a measure, termed
cointension, of the degree to which the intension
of p is a good fit to the intension of p. A
principal desideratum of precisiation is that the
resulting precisiand be cointensive. The concept
of cointensive precisiation is a key to
mechanization of natural language understanding. - The concept of NL-computation has wide-ranging
ramifications, especially within human-centric
fields such as economics, law, linguistics and
psychology
155APPENDIX
156DEDUCTION THE BALLS-IN-BOX PROBLEM
- Version 1. Measurement-based
- A flat box contains a layer of black and white
balls. You can see the balls and are allowed as
much time as you need to count them - q1 What is the number of white balls?
- q2 What is the probability that a ball drawn at
random is white? - q1 and q2 remain the same in the next version
157DEDUCTION
- Version 2. Perception-based
- You are allowed n seconds to look at the box. n
seconds is not enough to allow you to count the
balls - You describe your perceptions in a natural
language - p1 there are about 20 balls
- p2 most are black
- p3 there are several times as many black balls
as white balls - PTs solution?
158MEASUREMENT-BASED
PERCEPTION-BASED
version 2
version 1
- a box contains 20 black and white balls
- over seventy percent are black
- there are three times as many black balls as
white balls - what is the number of white balls?
- what is the probability that a ball picked at
random is white?
- a box contains about 20 black and white balls
- most are black
- there are several times as many black balls as
white balls - what is the number of white balls
- what is the probability that a ball drawn at
random is white?
159COMPUTATION (version 2)
- measurement-based
- X number of black balls
- Y2 number of white balls
- X ? 0.7 20 14
- X Y 20
- X 3Y
- X 15 Y 5
- p 5/20 .25
- perception-based
- X number of black balls
- Y number of white balls
- X most 20
- X several Y
- X Y 20
- P Y/N
160FUZZY INTEGER PROGRAMMING
Y
X most 20
XY 20
X several y
x
1