Download as pdf or txt
Download as pdf or txt
You are on page 1of 5

Fuzzy Logic, Logic Programming, and Linear Logic:

Towards a New Understanding of Common Sense


H.T. NGUYEN
Department of Mathematical Sciences, New Mexico State University, Las Cruces, NM 88003
email hunguyenQnmsu . edu

V. KREINOVICH
Department of Computer Science, University of Texas at El Paso, El Paso, TX 79968
email v l a d i k Q c s. u t ep . edu

Abstract
Fuzzy logic was originally proposed as a tool for describing human reasoning. Currently, the main area
of applications of fuzzy logic is an fuzzy control, where
the choice of a logic is usually motivated not b y logical considerations (i.e., not b y what best describes how
people actually think), but b y purely pragmatic, engineering considerations: what logic would lead t o the
best control.
In this paper, we try to return t o the original meaning of fuzzy logic: a tool for describing human reasoning. We analyze why the existing formalisms are not
always adequate, and describe possible modifications
of fuzzy logic. Our analysis shows that there are deep
similarities between the descriptions of common sense
reasoning in three different fields: fuzzy logic, logic
programming, and linear logic. Thus, the future formalism for describing human reasoning will probably
be a synthesis of these three.

Two Meanings of the Term Fuzzy


Lsgic: an Engineering Tool (as in
Fuzzy Control), and a Description of
Human Reasoning
1.1 Fuzzy Logic Started as a Description
of Human Reasoning
Fuzzy Logic is Proposed as a Family of Possible
Formalisms. Fuzzy set theory and fuzzy logic were
1

originally proposed to describe how people actually


think and reason [25]. For this purpose, L. Zadeh proposed to represent degrees of belief in different statements by numbers from the interval [0,1], and suggested several alternative reasonable operations with
these degrees of belief that correspond to logical connectives and , ~ r not
,
, etc.
Clarification: Degrees of Belief are Not TruthFunctional, but for Suficiently Large Knowledge
Bases, We Need And, Or, and Not Operations
with Degrees of Belief. It is well known that the degrees of belief are not truth-functionalin the sense that

0-7803-3225-3-6/96 $5.00 0 1996 IEEE

if we only known the degrees of belief d(A) and d(B)


in the statements A and B , then, depending on the
relationship between A and B , we can have different possible values of the degree of belief in A&B.
For example, if for some statement A , we know that
d(A) = 0.5 and d ( i A ) = 0.5, then for B = A , we
have A&B = A , so in this case, d(B) = 0.5 and
d A&B) = d(A) = 0.5, while for B = T A , we have
d [ B ) = 0.5 and d(A&B) = d(A&-A) = 0 . So, ideally, to describe the experts degrees of belief in different logical combinations of the original statements,
we have to describe the degrees of belief in all these
logical combinations. In real-life knowledge bases, the
numbers of statements s is reasonably large (often, s
is about several hundred), so, it is practically impossible to get the experts degrees of belief in all 2 2
(2 2 M lo3) logical combinations of these statements, For such knowledge bases, we will, therefore,
often encounter the situations in which for some statements A and B , we know the experts degree of belief
d(A) and d(B) in these statements, but we do not
know the experts degree of belief in, say, A&B. In
this situation, if a user asks a query about A&B, we
have to estimate the degree of belief in A&B based on
the known degrees of belief in A and B . This estimate
is a function of d ( A ) and d(B) that we have denoted
) ,B ) ) .
by f & ( d ( A 4
To find the function f& that is the most adequate
for a given area, we do the following:
First, we ask an expert, for different pairs of statements ( A ,B ) , to estimate his degree of belief in A , B ,
and A&B;
Then, we apply some data processing method to the
resulting set of triples (d(A),d ( B ) ,d(A&B)),and find
a function f & for which on average , the approximation d(A&B)M f & ( d ( A ) , d ( B )is) the best for these
triples.
The (Unsuccessful) Search for the Actual
Fuzzy Logic. At first, some researchers thought that
it would be possible to find the actual operations with
degrees of belief that best describe human reasoning in
all possible areas; for example, an operation f&(a,.b)
that best describes how the experts degree of belief

546

d(A&BJ in the conjunction A&B is related to the experts egrees of belief d(A) and d ( B ) in the statements A and B : d(A&B)M f & ( d ( A ) ,d ( B ) ) .
This idea was used by the designers of the first
successful expert system MYCIN (for diagnosing rare
blood diseases), who spent several years interviewing
medical experts and trying to find functions f k , fv ,
etc., that best describe the reasoning of these experts
[4]. As a result of this analysis, they found (and implemented) some reasonably complicated operations.
The same operations were describing experts that
belonged to different schools, to different countries,
etc.; therefore, it was originally thought that they may
be the desired universal operations that describe human reasoning. The authors of MYCIN were so sure
about the universality of their operations that they
even designed an expert system shell (called EMYCIN,
short for Empty MYCIN) that implemented MYCINs
operations and that could be filled with knowledge
from different areas.
Unfortunately, the very first application of
EMYCIN to another field (geophysics) was very unsuccessful. It turned out that to adequately describe how geophysicists think we need &- and
V-operations different from the ones used by medical doctors. This difference can be easily explained by
the fact that doctors have to be very cautious in their
statements and decisions: indeed, it is normal for a
geophysicist to have several wrong predictions of, say,
oil, as long as he is on average successful, but a doctor
can afford only a few mistakes.

A General Formalism of Fuzzy Logic: &- and


V-Operations ( t - N o r m s and t-Conorms). Since
it turned out that in different fields, reasoning is described by different operations, the original goal of describing the most adequate fuzzy logic was replaced by
the goal of describing the class of fuzzy logics that can
be adequate for different areas. Such general formalism was presented in terms of axioms that describe &and V-operations (also called t-norms and t-conorms;
for details, see, e.g., [13]).
The idea was: when we start analyzing a new
field, we must first find, among all possible &- and
V-operations, the operations that best describe this
particular field.

1.2

Enter Fuzzy Control

In his pioneer 1971 paper [26], Zadeh proposed to


use fuzzy logic to describe the reasoning of operators
who control real-life system, and thus design an automatic controller. This idea, first implemented by
Mamdani, started the field of fuzzy control, or fuzzy
logac control, that by now, has become the most successful application area for fuzzy logic [13].
1.3 Evolution of Fuzzy Logics Used in

tions that describe the expert knowledge, and then,


used the resulting (rather complicated) operations.
It soon turned out that the initial careful analysis
was not only unnecessary, but often even misleading:
Indeed, first, the quality of the resulting fuzzy control
turned out to be rather independent on the fine details of membership functions, etc., and second, when
the designers of fuzzy cont,rollers did some fine-tuning
to find the membership functions and operations that
lead to the best control, they found out that the resulting (pragmatically best) operations are quite different
from the ones that the original expert operators used.
As a result of this discovery, at present, the design
of fuzzy controllers starts with very simple membership functions and operations, that are later fine-tuned
to get the best control. The term fuzzy logic is still
used, but this logic is simply an engineering tool
for designing controllers (see also [lo7 91). This new
meaning of the term fuzzy logic is most evident in
Mike Smiths adaptive fuzzy control, in which different
&- and V-operations are chosen on different stages of
control (see, e.g., [23] and i~eferencestherein). Clearly,
we do not switch our thinking patterns that fast.
For this engzneerzng fumy logic, there are several
results about the choice of the best operations (for
the basis results of this type, see, e.g., [23, 151).
Formulation of the Prloblem. This use of engzneerang logic rather than llogic that describes human
thinking is natural in contirol applications: in control,
our goal is not to simulate the human control (the
control that is often far from perfect), but to desi n
an automated controller that is the best possible tin
particular, that is superior to the original human operator).
However, in other (non-control applications, we
still need to describe how people t ink. In this case,
we must revisit the same question: what logic is the
most adequate for describing how humans think. In
this paper, we will descrilbe why the existing fuzzy
logic is not always adequate, and what changes need
to be made.

The First Reason Why Simple Fuzzy


Logic is Not a Perfect Description of
Human Reasoning: Too Few
Degrees of Be1ie.f

The fact that the simple fuzzy logic described above


may not be the most adequate formalism for describing human reasoning was first noticed by Zadeh himself. For example, one of the most natural ways to describe an experts degree of belief d(A) in a statement
A is to ask an expert to estimate his/her degree of
belief on a scale, say, from 0 to 10. If an expert marks
6, we take d ( A ) = 0.6(= 6/10). Theoretically, the
more points of the scale, the better the estimates. In
practice, however, an expert usually cannot describe
his degree of belief too precisely: e.g., hardly anyone is able to distinguish between 600 and 601 on a
scale from 0 to 1000. So, we cannot find the exact
degree of belief $(A);at best, we can find an anterval
d(A) = [&(A), d+(A)] of possible degrees of belief.

Fuzzy Control: From Description of


Human Reasoning to Pragmatic

Engineering Tool
The first papers on fuzzy logic followed the abovedescribed pattern: first, the researchers carefully described the membership functions and logical opera-

547

Numerical degrees of belief d A ) can be represented


as degenerate intervals [d(A),( A ) ] ;so, we can say
that this interval-valued fuzzy logic is obtained by
a d d i n g new degrees of belief (non-degenerate intervals)
to the numbers from the interval [ a l l ] . Since we had
t o a d d new degrees of belief to get a better description
of human reasoning, we can say that the problem with
the simple fuzzy logic is that it has t o o few degrees of
belief.
This idea of interval-valued fuzzy logic was proposed and successfully applied by I. B. Tiirkgen, L.
Kohout, and others (see, e.g., interval sections of
NAFIPS 1994 [17], or a survey [20]).
Intervals may also not be sufficient: e.g., an expert
can not only say that his degree of belief corresponds
t o either 5 or 6 on a scale from 0 to 10, but he can also
add that it is rather 5 than 6. To describe this complicated statement, we must quantize what the author
means by this rather ... then; if we use fuzzy logic to
describe this rather, then we end up with a degree
of belief that is, in itself, a fuzzy set. These complicated degrees of belief include not only the original
beliefs (described by real numbers), but also degrees
of meta-belief (belief in belief). Such second-order
fuzzy sets have been proposed long ago (see, e.g., [13]);
they are not yet practically useful, but they seem t o
be a reasonable step t o go. It has also been suggested
that we can go one step further and describe third ord e r fuzzy sets, etc.
The higher we go in this hierarchy, the more complicated the description becomes; so we get very complicated and mathematically interesting theorems, and
hardly any algorithms. So fuzzy logic, that started as
a simple, natural, and computer-implementable tool
to describe uncertainty, turns into a complicated nonalgorithmic mathematical complexity. Is this complexity really necessary?
The main reason for this complexity is that we have
too many degrees of belief, so many, that it is difficult
t o represent them in the computer: e.g., if a degree of
belief is a number, we just store this number; if it is
a fuzzy set (i.e., a function), we need to store many
values of this function, etc. So, our only hope to avoid
this complexity is t o somehow decrease the number of
possible degrees of belief. We do not propose t o go
back t o the simple fuzzy logic: some n e w degrees of
belief (especially the interval ones) are truly necessary.
What we propose to do is t o analyze whether all old
degrees of belief were necessary.
This is what we will do in the next section.

The Second Reason Why Simple


zzy Logic is Not a Perfect
Description of Human Reasoning:
Too Many Degrees of Belief

We will give two arguments in favor of out thesis


that in simple fuzzy logic, there are too many degrees
of belief. The first argument is not original, it has
often been proposed, and it is based on common sense;
for this argument, we will not need any logic at all.
Our second argument will be more complicated and
based on the analysis of the resulting logic.

548

3.1

The First Argument: Degrees of


Belief Correspond to Words, and
There Are Only Finitely Many Words

We are trying t o formalize experts degrees of belief, and these degrees of belief are represented by
words and phrases of natural language. There are
only finitely many different words and phrases, and
therefore, there are only finitely many degrees of belief. Hence, out of infinitely many value E [0, l],only
finitely many can serve as actual degrees of belief.
In control application, degrees of belief get transformed into control values. So, if we want the control
to be continuous, we must require that the set of all
degrees of belief be continuous; hence, in the e n g i n e e r ing logic, we must consider all d E [0,1].
However, when we are describing human reasoning,
we do not need to add anything t o the finite set of
degrees of belief. Such finite-valued logics have been
successfully used in expert systems [l].
The use of these logics improves the size and response time of the corresponding expert systems:
since there are fewer degrees of belief than all real
numbers from [0, 11, we need fewer bits t o store each
degree of belief; thus, we can process them faster. To
achieve the biggest possible advantage, we must appropriately modify traditional fuzzy formalisms; this
modfication is still in the works. For example, a membership function p can no longer be a continuous function; for linearly ordered finite set D = (a1 < . . . <
a,} of degrees of belief, we can describe a membership function p : X -+ D as a finite sequence of embedded nested) a-cuts, i.e., as a sequence of sets
Si = {z p ( z ) 2 ai} for which SI _> ... _> S,. The
survey of the recent results can be found in [21].

The Second Argument: With Respect


to Logical Properties, Interval-Valued
Logic is Equivalent to 4-Valued One
The Main Result. In control applications, we need
3.2

the actual values of degrees of belief, because these


values are directly transformed into control values.
In formalizing human reasoning, the values are just
a technical tool.
What we are interested in are questions like: when
are given l o g i c a l f o r m u l a s P and Q (i.e., formulas built
from variables A , B, ..., by using logical connectives,
e.g., A&(-A V B ) ) , equivalent? (i.e., when do the
computed degrees of belief d ( P ) and d(&) in P and
Q coincide for all possible degrees of belief d ( A ) , ...,
in A , B , ...7) When does P imply Q? (i.e., when
d ( P ) 2 d(Q) for all d(A), ...?).
Let us first consider fuzzy logics in which A&A = A
and A V A = A , i.e., in which $& = min and fv =
max. Without loss of generality, we can assume that
f-,(u) = 1 - U. It turns out that for the intervalvalued logic with these operations, formulas P and
Q are equivalent (i e d P = d Q for all possible
intervals d(A), ...) iff d[P] = d[Q] for all cases in
which each variable is assigned one of the following
four intervals: 0(= [ O , O ] > , 1, 0.5, .and [0,1]. In other
words, P and Q are equivalent in interval-valued logic
iff they are equivalent in a 4-valued logic (for details

we have AandA = A. Crudely speaking, and correspond t o min(a, b ) , while both looks more like a . b .
This fine distinction may explain the necessity to use
several different &-operations in fuzzy control [23]:
as the control situation changes, we are not changing
the way we think (that would be impossible), we are
just changing the meaning of the word and. Our understanding of fuzzy control would thus improve if we
combine this distinction with fuzzy logic.

Acknowledgments
This work was supported in part by NSF Grant No.
EEC-9322370, and by NASA Research Grant No. 9757. The authors are thankful t o Petr Hijek and the
anonymous referees for valuable suggestions.

References
[l] J . Agusti et al. Structured local fuzzy logics in
MILORD, In: L. Zadeh and J. Kacrpzyk (eds.)
Fuzzy Logic for the Management of Uncertainty,
Wiley, N.Y., 1992, pp. 523-551.
[a] B. R. Barmish, New tools for robustness of h e a r
systems, McMillan, N.Y., 1994.
[3] N. D. Belnap, The logic of questions and answers,
Yale Univ. Press, New Haven, CT, 1976.
[4] B. G. Buchanan and E. H. Shortliffe. Rule-based
expert systems, Addison-Wesley, 1984.
[5] D. E. Cooke, R. Duran, A. Gates, and V.
Kreinovich, Bag languages, concurrency, Horn
logic programs, and linear logic, Proc. 6-th In27
Conference on Software Engineering and Knowledge Engineering SEKE94, June 1994, Jurmala,
Latvia, IEEE Computer Society, Skokie, IL, 1994,
pp. 289-297.
[6] L. A. Cooper and V. Ya. Kreinovich, Using Linear Logic t o represent uncertainty of our knowledge, University of Texas at El Paso, Computer
Science Department, Technical report UTEP-CS90-2, May 1990.
[7] M. Gelfond and H. Przymusinska, Definitions
in epistemic specifications, In: A . Nerode, W.
Marek, V. S. Subrahmanian, Logic Programming
a n d Non-Monotonic Reasoning, Proc. of the 1st
Intl. Workshop, 1991, pp. 245-259.
[a] J.-Y. Girard, Linear logic , Theoretical Computer Science, 1987, Vol. 50, pp. 1-102.
[9] P. Hcijek, Fuzzy Logic from the Logical Point
of View, Institute of Computer Science, Czech
Academy of Sciences, Prague, preprint, 1996.
[lo] P. Hajek, T. Havrcinek, and R. Jirousek, Uncertain Information Processing in Expert Systems,
CRC Press, Boca Raton, FL, 1992.
[11] R. B. Kearfott and V. Kreinovich (eds.), Applications of Interval Computations, Kluwer, Dordrecht, 1996.
[12] V. L. Kharitonov, Asymptotic stability of an
equilibrium position of a family of systems of linear differential equation, Differential Egzlations,
1978, Vol. 14, pp. 2086-2088.

[13] G. Klir and B. Yuan, Fuzzy sets and fuzzy logic:


theory and applications, Prentice Hall, Upper
Saddle River, N J , 1995.
[14] 0. Kosheleva, V. Kreinovich, and H. T . Nguyen,
Mamdanis Rule: a weird use of and i ~ simplication justified by modern logic , Sixth IFSA
World Congress, San Paulo, Brazil, 1995, Vol. 1,
pp. 229-232.
151 V. Kreinovich e t al., What non-linearity t o
choose? Mathematical foundations of fuzzy control, Proc. 1992 Intl Conf. on Fuzzy Systems
and Intelligent Control, Louisville, KY, 1992, pp.
349-412.
161 R. E. Moore, Interval analysis, Prentice Hall, Englewood Cliffs, NJ, 1966.
[17] NAFIPS/IFIS/NASA 94, San Antonio, December 1994, IEEE, Piscataway, NJ, 1994.
[IS] H. T. Nguyen, 0. M. Kosheleva, and V.
Kreinovich, ((Is the success of fuzzy logic really
paradoxical? Or: Towards the actual logic behind expert systems, Intl J. of Intelligent Systems, 1996 (to appear).
[19] H. T. Nguyen andV. Kreinovich, How stable is a
fuzzy linear system?, FUZZ-IEEE94, Orlando,
FL, 1994, Vol. 2, pp. 1023-1027.
[20] H. T. Nguyen and V. Kreinovich, (Towards theoretical foundations of soft computing applications , Intl J. on Uncertainty, Fuzziness, and
Knowledge-Based Systems, 1995, Vol. 3, No. 3,
pp. 341-373.
[21] H. T . Nguyen and V. Kreinovich, Nested Intervals and Sets: Concepts, Relations t o Fuzzy Sets,
and Applications, In [ll],pp. 245-290.
[22] T. Przymusinski, The well-founded semantics
coincides with the three-valued stable semantics,
Fundamenta Informaticae, 1989.
[23] M. H. Smith and V. Kreinovich, (Optimal strategy of switching reasoning methods in fuzzy control, in H. T. Nguyen et al. (eds.), Theoretical
aspects of fuzzy control, J . Wiley, N.Y., 1995, pp.
117-146.
[24] A. S. Troelstra, Lectures on linear logic, CSLI,
St anford , 1992.
[25] L. Zadeh, Fuzzy sets, Information and control,
1965, Vol. 8, pp. 338-353.
[26] L. A. Zadeh, (Towards a theory of fuzzy systems, In: R. E. Kalman and N. DecLaris (eds.),
Aspects of network and systems theory, Holt,
Rinehart, Winston, 1971.

550

we have AandA = A. Crudely speaking, and correspond t o min(a, b ) , while both looks more like a . b .
This fine distinction may explain the necessity to use
several different &-operations in fuzzy control [23]:
as the control situation changes, we are not changing
the way we think (that would be impossible), we are
just changing the meaning of the word and. Our understanding of fuzzy control would thus improve if we
combine this distinction with fuzzy logic.

Acknowledgments
This work was supported in part by NSF Grant No.
EEC-9322370, and by NASA Research Grant No. 9757. The authors are thankful t o Petr Hijek and the
anonymous referees for valuable suggestions.

References
[l] J . Agusti et al. Structured local fuzzy logics in
MILORD, In: L. Zadeh and J. Kacrpzyk (eds.)
Fuzzy Logic for the Management of Uncertainty,
Wiley, N.Y., 1992, pp. 523-551.
[a] B. R. Barmish, New tools for robustness of h e a r
systems, McMillan, N.Y., 1994.
[3] N. D. Belnap, The logic of questions and answers,
Yale Univ. Press, New Haven, CT, 1976.
[4] B. G. Buchanan and E. H. Shortliffe. Rule-based
expert systems, Addison-Wesley, 1984.
[5] D. E. Cooke, R. Duran, A. Gates, and V.
Kreinovich, Bag languages, concurrency, Horn
logic programs, and linear logic, Proc. 6-th In27
Conference on Software Engineering and Knowledge Engineering SEKE94, June 1994, Jurmala,
Latvia, IEEE Computer Society, Skokie, IL, 1994,
pp. 289-297.
[6] L. A. Cooper and V. Ya. Kreinovich, Using Linear Logic t o represent uncertainty of our knowledge, University of Texas at El Paso, Computer
Science Department, Technical report UTEP-CS90-2, May 1990.
[7] M. Gelfond and H. Przymusinska, Definitions
in epistemic specifications, In: A . Nerode, W.
Marek, V. S. Subrahmanian, Logic Programming
a n d Non-Monotonic Reasoning, Proc. of the 1st
Intl. Workshop, 1991, pp. 245-259.
[a] J.-Y. Girard, Linear logic , Theoretical Computer Science, 1987, Vol. 50, pp. 1-102.
[9] P. Hcijek, Fuzzy Logic from the Logical Point
of View, Institute of Computer Science, Czech
Academy of Sciences, Prague, preprint, 1996.
[lo] P. Hajek, T. Havrcinek, and R. Jirousek, Uncertain Information Processing in Expert Systems,
CRC Press, Boca Raton, FL, 1992.
[11] R. B. Kearfott and V. Kreinovich (eds.), Applications of Interval Computations, Kluwer, Dordrecht, 1996.
[12] V. L. Kharitonov, Asymptotic stability of an
equilibrium position of a family of systems of linear differential equation, Differential Egzlations,
1978, Vol. 14, pp. 2086-2088.

[13] G. Klir and B. Yuan, Fuzzy sets and fuzzy logic:


theory and applications, Prentice Hall, Upper
Saddle River, N J , 1995.
[14] 0. Kosheleva, V. Kreinovich, and H. T . Nguyen,
Mamdanis Rule: a weird use of and i ~ simplication justified by modern logic , Sixth IFSA
World Congress, San Paulo, Brazil, 1995, Vol. 1,
pp. 229-232.
151 V. Kreinovich e t al., What non-linearity t o
choose? Mathematical foundations of fuzzy control, Proc. 1992 Intl Conf. on Fuzzy Systems
and Intelligent Control, Louisville, KY, 1992, pp.
349-412.
161 R. E. Moore, Interval analysis, Prentice Hall, Englewood Cliffs, NJ, 1966.
[17] NAFIPS/IFIS/NASA 94, San Antonio, December 1994, IEEE, Piscataway, NJ, 1994.
[IS] H. T. Nguyen, 0. M. Kosheleva, and V.
Kreinovich, ((Is the success of fuzzy logic really
paradoxical? Or: Towards the actual logic behind expert systems, Intl J. of Intelligent Systems, 1996 (to appear).
[19] H. T. Nguyen andV. Kreinovich, How stable is a
fuzzy linear system?, FUZZ-IEEE94, Orlando,
FL, 1994, Vol. 2, pp. 1023-1027.
[20] H. T. Nguyen and V. Kreinovich, (Towards theoretical foundations of soft computing applications , Intl J. on Uncertainty, Fuzziness, and
Knowledge-Based Systems, 1995, Vol. 3, No. 3,
pp. 341-373.
[21] H. T . Nguyen and V. Kreinovich, Nested Intervals and Sets: Concepts, Relations t o Fuzzy Sets,
and Applications, In [ll],pp. 245-290.
[22] T. Przymusinski, The well-founded semantics
coincides with the three-valued stable semantics,
Fundamenta Informaticae, 1989.
[23] M. H. Smith and V. Kreinovich, (Optimal strategy of switching reasoning methods in fuzzy control, in H. T. Nguyen et al. (eds.), Theoretical
aspects of fuzzy control, J . Wiley, N.Y., 1995, pp.
117-146.
[24] A. S. Troelstra, Lectures on linear logic, CSLI,
St anford , 1992.
[25] L. Zadeh, Fuzzy sets, Information and control,
1965, Vol. 8, pp. 338-353.
[26] L. A. Zadeh, (Towards a theory of fuzzy systems, In: R. E. Kalman and N. DecLaris (eds.),
Aspects of network and systems theory, Holt,
Rinehart, Winston, 1971.

550

You might also like