Institute for Research and Applications of Fuzzy Modeling

University of Ostrava

Formal Theories in Fuzzy Logic

Vil m Nov k
Research report No. 1
July 22, 1997 D. Dubois, E. P. Klement and H. Prade (eds.)
Supported by: Submitted/to appear:

Fuzzy Sets, Logics and Arti cial Intelligence. Kluwer 1997

Grant A1086501 of the GA AV R; (partially) grant 201/96/0985 of the GA R

University of Ostrava Institute for Research and Applications of Fuzzy Modeling Br fova 7, 701 03 Ostrava 1, Czech Republic
tel.: +420-69-622 2808 fax: +420-69-22 28 28 e-mail: [email protected]

Formal Theories in Fuzzy Logic

University of Ostrava Inst. for Research & Applications of Fuzzy Modeling Br fova 7, 701 03 Ostrava, Czech Republic Institute of Information and Automation Theory Academy of Sciences of the Czech Republic Pod vod renskou v 4, 186 02 Praha 8, Czech Republic
and

y)

Vil m Nov k

1 Introduction
In this paper we deal with syntactic aspects of two kinds of fuzzy logic, namely fuzzy logic in narrow (FLn) and that in broader sense (FLb). Fuzzy logic in narrow sense is now quite well established though the work is far from being nished. The goal of this logic is to develop means for modeling of the vagueness phenomenon. One of the partial goals is to formulate analogues of most of the classical logic theorems. This makes us possible to clarify the relation of fuzzy logic to classical logic and also, to gain a more profound understanding to both logics. Further work should concentrate on the extension of the results to comprehend better the vagueness phenomenon. One of the interesting problems are open fuzzy theories in FLn. We consider this topic important as it has direct impact to questions of provability and algorithmization and thus, also to applications. Unfortunately, as shown in 6], proving in fuzzy logic is highly computationally ine ective. Hence, we have to seek some sophisticated methods which, of course, may help us in solution of only speci c problems (which, however, may be just those important for applications). Further interesting consequences may be expected in FLb, which is an extension of FLn. The goal is to develop a logic of the commonsense human reasoning whose main characteristic feature is the use of natural language. FLb includes the concept of computing with words, which has been recently introduced by L. A. Zadeh. In this paper, we de ne the concept of formal theory in both fuzzy logics, demonstrate some of their basic properties and mutual connection of FLb and FLn. We will focus especially to syntactic aspects and speci c questions of provability. However, we assume that the reader is, at least partly, acquainted with some of the cited works 7, 11, 16, 19, 23] where precise de nitions of some concepts and proofs of some theorems, which are only recalled in this paper, can be found.

2 Formal theories in fuzzy logic in narrow sense

2.1 Truth values and consequence operation
Recall that the set of truth values is considered to be the residuated lattice L = hL; _; ^; ; !; 0; 1i; (1) where L is either a nite chain, or L = 0; 1] and ; ! is the adjoint couple of product and residuation. As analyzed in detail in 8], we may distinguish three main streams of FLn, namely that of Lukasiewicz style (FLn(L)), G del style (FLn(G)) and product style (FLn(P)). All three logics assume L = 0; 1] and di er in the de nition of the couple of operations and !. In this paper, we will work in FLn(L) (fuzzy
y) This paper has been supported by the grant A1086501 of the GA AV R and partially also by the grant 201/96/0985 of the GA R.

logic in narrow sense of Lukasiewicz style) in which is the Lukasiewicz product and ! the Lukasiewicz implication. The reasons for this are widely discussed in the literature and we will mention some of them also in this paper. Note that in (1) is a particular case of the concept of t-norm (cf. e.g. 5, 9, 25]) and ! is the corresponding residuation. A general feature of FLn can be characterized by the possibility to introduce more kinds of connectives than classical logic. The choice is practically unlimited but t-norms (and tconorms) seem to be most important and interesting. However, we have to cope with the fact that we obtain various logical systems (determined especially by the implication operation) which may not always behave well with respect to our idea. When con ning to continuous t-norms, we come to the three above mentioned fuzzy logics in narrow sense. However, only Lukasiewicz implication is continuous. Therefore, FLn(L) possesses most distinguished properties. The general requirement in any logic is that the connectives should preserve equivalence which in FLn is naturally interpreted by a $ b = (a ! b) ^ (b ! a); (2) n ?! L is logically n=a a; b 2 L. Furthermore, we put a | {z a. The we say that the operation c : L }
tting on L if there are natural numbers k1 > 0; : : :; kn > 0 such that
n

(3) (an $ bn)kn c(a1; : : :; an) $ c(b1; : : :; bn) (a1 $ b1)k1 holds for every a1; : : :; an, b1; : : :; bn 2 L. Using this concept, it is possible to develop fuzzy logic as an open system in which four operations are basic (given by the structure of the residuated lattice (1) and to extend it, if necessary, by some additional operations. Hence, the structure of truth values may be assumed to form an enriched residuated lattice L = hL; _; ^; ; !; fcj j j 2 Jopg; 0; 1i; (4) where fcj j j 2 Jopg is a set of logically tting operations (Jop is some index set). Let us stress that introducing these operations in FLn(L) does not harm the whole logical system (cf. 23]). Note also that in FLn(L), logically tting operations are exactly those being Lipschitz continuous (see 10]). Additional connectives are especially important in FLb where we need them to accomplish interpretation of various natural language connectives, modi ers and, possibly, other linguistic phenomena. We will consider formal language J consisting of variables, constants, predicates, connectives and quanti ers, as de ned, for example in 11, 19]. A speci c feature is introducing symbols a for all truth values a 2 L. However, as demonstrated in 6, 8, 15], this is only a useful technical means. By FJ we denote the set of all the well-formed formulas (de ned in a common way) and by MJ sets of all terms in the language J. The basic connectives are ; ; & and interpreted by the operations ^; _; and !, respectively. The operation of sum is de ned by a b = :(:a :b), a; b 2 L. It can be extended to multiple na. Syntax of fuzzy logic is evaluated by syntactic truth values. This makes us possible to deal with truth values in the syntax. Furthermore, the main task of fuzzy logic in narrow sense has been declared to provide tool for grasping the vagueness phenomenon. Necessary condition for that is equal importance of all the truth values. Evaluated syntax seems to be a suitable means for this purpose. The evaluated formula is a couple A; a] where A 2 FJ and a 2 L. Evaluated formulas are manipulated using the evaluated n-ary inference rules r which are couples r = hrsyn; rsemi (5) where rsyn is syntactic part of the rule r which is a partial n-ary operation on FJ and rsem is its semantic part which is an n-ary operation on L preserving arbitrary non-empty joins in each argument (semicontinuity). We will work with sound inference rules, i.e. those preserving truth evaluations (for precise de nition see 11, 23]).

^_

Let us remark that evaluated formula A; a] can also be seen as a fuzzy singleton f a Ag. Hence, every set of evaluated formulas is at the same time a fuzzy set of formulas. This ambiguity will often be used in the sequel. A question raises where the syntactic truth values come from; how they should be interpreted? As pointed out by P. H jek (cf. 7]), it is natural to understand evaluated formulas as the formulas a A (a is a symbol for truth value a 2 L) which, when being true in the degree 1, means that the truth of A is greater than or equal to a. This understanding has several consequences. First, we may interpret the evaluated formulas as shorts for the latter ones. Second, the evaluated rules of modus ponens A B; rMP : A; a]; a b] b] B;

A; a] rG : (8x)A; a] may be embedded in non-evaluated syntax simply as special cases of classical ones. For example, rMP can be obtained in FLn(L) (with non-evaluated syntax and truth values a, a 2 L, in the language) using the proof, in which transitivity and importation tautologies, and rule of modus ponens are used: a A; b (A B); (b (A B)) (A (b B)); A (b B); (a A) ((A (b B)) (a (b B)); (a (b B); (a (b B)) (a b B); a b B:

and generalization

) ) ) ) ) ) ) ) ) ) ) ) ) ) ) ) ) ) ) ) ) ) ) ) )

Most important is the fact that somewhat technical and not quite natural b-lifting rules A; a] rRb : b A; b ! a] may be omitted as they can easily be replaced by simple proofs which use modus ponens and the transitivity tautology (a A) ((b a) (b A)). Let us stress, however, that it is not reasonable to avoid evaluated formulas completely. Consistent replacement of evaluated formulas by a A would lead to a very cumbersome notation. Furthermore, the primary goal which is graded model of vagueness (fuzzy approach) would disappear. Other reason consists in the evaluated inference rules. The semantical operation rsyn in (5) is required to be only upper semicontinuous and is realized on the syntactic truth evaluations. This opens the way for extension of fuzzy logic by various non-standard inference rules. However, omitting the concept of evaluated formulas would result in the restriction only to the operations de ned a priori in the structure (1) (or (4)). Consequently, the semantical operation rsyn would have to be inherently realized as the interpretation of some logical connective. As an example, let us consider the rule introduced already in 23]: A ; r : A; bb# a] a] where # is the operation de ned by b # a = 0 otherwise if b a a a; b 2 0; 1]. The operation # is discontinuous and thus, not logically tting in the sense of (3). Hence, it cannot be used as a logical connective in FLn(L). To conclude, we may introduce various, in general n-ary, inference rules with discontinuous semantical operation. We expect here the potential, for example, for modeling of the abduction in FLn.

) ) )) ) )

A fuzzy theory T is determined by a triple T = hAL ; AS ; Ri (6) where AL ; AS are sets of evaluated logical and special axioms, respectively (or, equivalently, fuzzy sets AL ; AS FJ ) and R is a set of inference rules containing at least the rules rMP and rG. In general, there may be fuzzy theories with di erent sets of inference rules. The concept of provability is crucial in any logic. Let C : FJ ?! L be a function such that C(r(A1 ; : : :; An )) 2 G whenever C(Ai ) 2 G, i = 1; : : :; n for every inference rule r. Let us call such a function rules preserving. The following de nition can be introduced in classical logic.

De nition 1 A formula A belongs to the set C synX of syntactic consequences of the set of formulas X i C(A) 2 G holds for every rules preserving function such that C(B) 2 G holds for every B 2 X .
Strong de nition of syntactic consequence accepted in classical as well as in many-valued logic requires existence of a proof w of A from X. In classical logic, De nition 1 is equivalent with the strong de nition of syntactic consequence. The main outcome of De nition 1 is the possibility to generalize syntactic consequences to the case of evaluated syntax.

De nition 2 Let R be a set of sound evaluated inference rules. Then the fuzzy set of syntactic consequences of the fuzzy set X FJ is given by
^

(C synX)A = fC(A) j C FJ , C preserves rules r 2 R and AL; X C g.

( is classical inclusion of fuzzy sets).

(7)

An evaluated proof w is a sequence of evaluated formulas A1 ; a1 ]; : : :; An ; an] such that every evaluated formula Ai ; ai ] is either and axiom or it is derived from previous evaluated formulas using some evaluated inference rule. The evaluation an of the last formula in w is called the value of the proof w and denoted by Val(w). If (C synAS )A = a then we write T `a A (a formula A is provable in the degree a in the fuzzy theory T). The following theorem, whose proof is based on (7) (see 23]), holds true.

Theorem 1

(C synAS )A = a = fVal(w) j w is an evaluated proof of A in T g: a = fb j b A is provable in T g

_

(8) (9)

Note that (8) is equivalent with

in a non-evaluated syntax (with truth values). Hence, Theorem 1 generalizes the existence of a proof to supremum of the values of all the possible evaluated proofs. However, this is not equivalent with the strong de nition of syntactic consequence in many-valued logic. To summarize, FLn(L) can be interpreted in many-valued logic if we introduce names for the truth values in its language and translate the evaluated formulas as mentioned above ). However, to obtain all the results mentioned below, we still need to generalize the concept of provability (cf. (7), (8)). Classical logic can be obtained within FLn as its special case because formulas and inference rules in classical logic may always be evaluated by the truth value 1. However, in many-valued logic, the strong
y

y) Note that this may concern only the rational truth values; we can even designate some formulas to serve us as the above names though the language needs not explicitly contain them.

de nition of the consequence operation is used. The bridge between classical logic and FLn is also other reason for keeping the concept of evaluated formulas; it can be more clearly seen that the latter is direct generalization of the former. At the end of this section, we will repeat some de nitions from FLn(L), keeping on mind the discussion above. The fuzzy theory T is given by the triple (6) where AL ; AR FJ are fuzzy sets of logical and special axioms (sets of evaluated formulas). At the same time we may see fuzzy theory T as a fuzzy set of formulas T = (C syn(AS AL )) FJ : The evaluated proof w of a formula A in the theory T is a sequence of evaluated formulas, the value of the last one is the value ValT (w) of the proof w. The provability degree (=evaluation) of an evaluated formula A is supremum of the values of all its evaluated proofs. In the sequel, when de ning a fuzzy theory, we will usually write only fuzzy set of its special axioms, i.e. T = f ai Ai j i 2 I g where I is some index set. By T j=a A we mean that a formula A 2 FJ is true in the degree a in the fuzzy theory T, i.e. a = fD(A) j D j= T g where D j= T means that D is a model of T (for the precise de nition see 11]).
^

2.2 Few theorems of FLn(L)

Fundamental theorems characterizing the provability are the validity and closure ones. The rst one says that the provability degree of a formula cannot exceed its truth. By the second one, we may con ne ourselves to closed formulas, analogously as in the classical logic. If T is a theory and E FJ a fuzzy set of formulas then T = T E is an extension of the theory T, i.e. its fuzzy set AS of special axioms is AS = AS E. Proofs of the following theorems can be found in the cited literature ( 6, 11, 18, 23].
0 0 0

Theorem 2 (deduction) Let A be a closed formula and T = T f 1 Ag. Then to every B there is n such that T `a An )B i T `a B:
0 0

A fuzzy theory is contradictory if there is a formula A 2 FJ (T ) and proofs wA and w A such that ValT (wA ) ValT (w A ) > 0 (other equivalent characterizations of inconsistency can also be introduced).
: :

Theorem 3 A theory T is contradictory i T ` A holds for every formula A 2 FJ T .

( )

Henkin extension TH of a theory T by a fuzzy set of Henkin axioms TH = T f 1 (Ax r] (8x)A(x)) j r is special for (8x)Ag is proved to be conservative.

Theorem 4 A fuzzy theory T f a Ag is contradictory i to every b 2 L and every formula B there is m such that T `b Am )B . Theorem 5 (completeness)
6

(a) A theory T is consistent i it has a model. (b) T `a A i T j=a A.

The equality predicate ful lling the following (common) axioms can be introduced: There are natural numbers m1 > 0; : : :; mn > 0 such that x=x (x1 = y1 )m1 : : : (xn = yn )mn (f(x1 ; : : :; xn) = f(y1 ; : : :; yn ) (x1 = y1 )m1 : : : (xn = yn )mn (p(x1 ; : : :; xn) p(y1 ; : : :; yn) for every n-ary functional symbol f and predicate symbol p. (E1) (E2) (E3)

) ) ) )

) )

0; : : :; mn > 0 such that

0

where A is a formula which is a result of replacing of the terms ti by the term si in A, respectively.

Theorem 6 (equality) Let T `ai ti = si , i = 1; : : :; n. Then there are natural numbers m > amn T `b A,A ; b a m1 n
1
0

Theorem 7 (equivalence) Let A be a formula and B ; : : :; Bn some of its subformulas. Let T `ai
Bi Bi , i = 1; : : :; n. Then there are natural numbers m1 > 0; : : :; mn > 0 such that T `b A A ; b a m1 amn n 1
0

where A is a formula which is a result of replacing of the formulas B1 ; : : :; Bn in A by B1 ; : : :; Bn.

0 0 0

Let ? FJ be a fuzzy set of formulas. By Supp(?) we denote its support, i.e. Supp(?) = fA j ?(A) > 0g. The denotes Lukasiewicz disjunction given by A B := ( A & B).

:: :

Theorem 8 (reduction for the consistency) A theory T = T ? is contradictory i there are natural numbers m > 0; : : :; mn > 0 and A ; : : :; An 2 Supp(?) such that T `c :Am1 5 5:Amn n where ai = ?(Ai ), i = 1; : : :; n and c > :(am1 amn ) or c = 1 if the right-hand side is equal to 1. n
0

and b > ma or b = 1 if ma = 1.

Corollary 1 A theory T = T f :a :Ag is contradictory i T `b mA for some natural number m > 0

0

Given a Henkin fuzzy theory T, the fuzzy set of formulas (T) contains instances of special, equality, Henkin and substitution axioms in the membership degree 1. A formula A is a fuzzy quasitautology in the degree a if j=a B1 & & Bk A where Bi are closed instances of the equality axioms.

Theorem 9 (consistency) Open theory T is contradictory i there are natural numbers p > 0; : : :,
pn > 0 and special axioms A1 ; : : :; An of the theory T such that p Ap1 Ann 1
1

is a fuzzy quasitautology in the degree b where Ai are instances of the special axioms and b > :(ap1 1 apn ) where ai = AS (Ai ) (or b = 1 if the right hand side is equal to 1). n

: 5 5:

This theorem is a basis for a fuzzy analogy of the Herbrand theorem proved in 20]. 7

3 Formal theories in fuzzy logic in broader sense

Unlike FLn, which is graded generalization of the classical logic, FLb can be motivated by modeling of the human deduction in which crucial role is played by the natural language. We develop it as a certain extension of FLn. In our considerations, we con ne to a suitable part of natural language, i.e. we will work with selected linguistic expressions (syntagms ) from some set S . These are translated into formulas of fuzzy logic in narrow sense and nally lead to construction of a special fuzzy theory T of FLn (cf. 19]). In formal part of FLb, we deal with many-sorted language J of FLn (cf. 3]). We consider a nite number of sorts , = 1; : : :; p of variables and the corresponding constants as well as functions. We thus obtain sets of terms M J of the sort in the language J. However, in FLb we will need only that part of M J which does not contain variables. Therefore, M will denote sets of terms of the sort without variables (the subscript J will often be omitted) in the sequel. As in the previous section, FJ is a set of well formed formulas in J. Natural language expressions are, in general, names of some properties ' of objects. The property '(x) of objects x is assigned a formula A(x) 2 FJ . Hence, each syntagm A 2 S is assigned a couple

hA(x); Ai

(10)

where A = f Ax t]; at] j t 2 M g is a set of evaluated formulas being closed instances of A(x), i.e. Ax t] is obtained from A(x) when replacing the variable x (of the sort ) by the term t 2 M . We will call A a multiformula. Recall from the previous section that a multiformula A can at the same time be seen as a fuzzy set of closed instances Ax t] of the formula A(x). This model is motivated by the use of natural language, and also by the potential of fuzzy logic in narrow sense, which enables us to work with fuzzy sets of formulas included in the syntax. This makes us possible to introduce and formalize (at least partly) the distinction between the concepts of intension and extension which are very important in the study of natural language semantics. In this paper, the intension is the multiformula A together with the formula A(x) (see (10)). To simplify the formalism, we will omit A(x) from most of our deliberation because it is inherently present in all elements of A. In intensional logic, intensions are (losely speaking) functions assigning truth values to objects in each possible world and time moment. Possible worlds as well as time are not explicitly included in our formalism. However, they are hidden behind the assignment of truth values to the instances Ax t] in the multiformula A. Hence, the truth values are assigned using certain function

?! fA : AjM ?! L A 2 S ; A 7! Ag where is the set of possible worlds, time and AjM a set of all instances of the form Ax t], t 2 M . In this paper, however, the multiformulas A are considered to be given a priori.
It follows from the previous discussion that the basic scheme in the fuzzy logic in broader scheme is syntagm A 7! formula A; intension A 7! extension D(A) where D is an interpretation (model) of the formal language J in concern. The extension is thus the fuzzy set o n D(A) = D(Ax t]) D(t) t 2 M : It is clear that one intension A may lead to (in nitely) many extensions D(A). Obviously, D(t) is a concrete object and D(Ax t]) is a truth degree in which the object D(t) has the property A(x). In the case of A(x1 ; : : :; xn) where x1 ; : : :; xn are variables of various sorts, we obtain a fuzzy relation
n

D(Ax1 ;:::;xn t ; : : :; tn]) hD(t ); : : :; D(tn)i t 2 M ; : : :; tn 2 Mn :

1 1 1 1

Example. Let A:= Young and M = ft ; : : :; t g be a set of terms representing years. We may de ne a multiformula Young by Young = f Y oung(t ); 1]; : : :; Y oung(t ); 1]; : : :; Y oung(t ); 0:6]; : : :; Y oung(t ); 0:2]; : : :; Y oung(t ); 0]g:
0 100 0 20 30 45 60

This is the logical representation of the intension of the word \Young". The extensions can be, for example, the following: (11) D(Young) = 1 1; : : :; 1 20; : : :; 0:6 30; : : :; 0:2 45; : : :; 0 60 where D(t0 ) = 1; : : :, D(t20 ) = 20; : : :, D(t30) = 30; : : :, D(t45 ) = 45; : : :, D(t60) = 60 are interpretations of the terms from M when representing age of people. When representing age of dogs, we may obtain the following extension of Young : (12) D(Young) = 1 0:1; : : :; 1 4; : : :; 0:7 6; : : :; 0:3 8; : : :; 0 14 where D(t0 ) = 0:1; : : :, D(t20 ) = 4; : : :, D(t30) = 6; : : :, D(t45) = 8; : : :, D(t60) = 14. Note that the truth degrees in (12) are grater than the corresponding ones in (11) to illustrate that only the inequality D(A) a should be ful lled where a is the truth evaluation of the formula A in the evaluated formula A; a]. Note also that fuzzy sets of the form (11) and (12) are introduced in various examples in the literature on fuzzy set theory ) . From our point of view, some concrete, usually not explicitly characterized, extension is considered there. Let us remark that a slightly simpli ed interpretation of fuzzy logic in broader sense which concerns only the logical aspect without linguistics has been proposed in 3]. However, we are convinced that linguistics should not be excluded from fuzzy logic and fuzzy techniques in general (let us remind the concept of soft computing where the main stress is given to \computing with words"). The proof in FLb is a sequence of linguistic statements (syntagms from S ) together with their intensions B1 B1]; : : :; Bn Bn] (13) each of which is a linguistically formulated axiom (logical or special), or it is derived using some inference rule. A formal theory of FLb is given by the set of linguistically expressed special axioms together with their intensions T = fA0 A0]; : : :; Am Am ]g; (14) where Ai 2 S , i = 1; : : :; m. The reasoning uses proofs of the form (13). As these deal with multiformulas, i.e. sets of evaluated formulas, we face a multiple inference in the adjoint fuzzy theory T determined by the intensions Ai, i = 1; : : :; m from (14), i.e. T = A0 Am : (15) In the sequel, we will denote the fuzzy theory in FLb by the script letter T and the adjoint theory of FLn by the italic letter T. To simplify the formalism, we may omit the linguistic statements from (13) and write proof using only the intensions B1 ; : : :; Bn: (16) This has also other side: To consider the corresponding syntagms at each step of (13) would mean that the reasoning proceeds in words all the time. In practice, it would force us to nd a suitable syntagm Bi to each multiformula Bi , i = 1; : : :; m, which is a task of linguistic approximation to be solved at each reasoning step. But this is unrealistic. Hence, realistic view is to begin with natural language,
y

y)

See, e.g. 26] and a lot of other papers and books.

translate its syntagms into multiformulas, then make proofs of the form (16) and translate only the nal multiformula Bn into the syntagm Bn . More precisely, we form a fuzzy theory T in (14) using natural language and then, within the adjoint fuzzy theory T in (15), we realise the reasoning whose result being a multiformula B may be formulated using a corresponding syntagm B 2 S . Intensions of the syntagms should be constructed from the other (simpler) ones. We face here the problem of truth functionality which is subject to a long and still un nished discussion between logicians and linguists. Truth functionality cannot, in general, be accepted in the model of semantics of natural language. However, for some parts of it, the truth functionality holds, or at least may be bypassed by accepting various kinds of connectives in the local cases (recall our discussion about additional operations in Section 2). Let SA ; SB be two disjoint sets of syntagms in the form

hlinguistic modi eri]hadjectiveihnouni

(17)

where each A 2 SA , B 2 SB is assigned a formula A and B, respectively and the intensions being the respective multiformulas A and B (more exactly, A; B are interpreted by the couples (10)). The following construction is important. The linguistic description in FLb is a set of linguistic conditional statements of the form
IF Aj THEN Bj Aj Bj ];
82 m < ^ 4 (Ajx t] : j =1
0

where Aj 2 SA , Bj 2 SB , j = 1; : : :; m. These statements can be joined by the connective AND interpreted using conjunction. The intension of the whole linguistic description (18) is thus a multiformula

j = 1; : : :m
3 9 =

(18)

)Bjy s]); cts =

m ^ j =1

(ajt ! bjs)5 t 2 M1; s 2 M2 ;

where x; y are variables of the sorts 1 and 2, respectively. Recall that a formula A is a variant of A if it is the result of replacing of all subformulas of A of the form (8y)B by formulas (8x)By x] where x is substitutible into A.

Lemma 1 Let T be a fuzzy theory and A be a variant of A. Then T `a A i T `a A :

0 0

This lemma justi es the following concepts. We say that two formulas A and B are independent if no variant or instance of one is a subformula of the other one. Let F0 be a set of evaluated formulas such that, if A; a]; B; b] 2 F0 then A; B are independent and to each A there is at most one a such that A; a] 2 F0 . We will call F0 a set of independent evaluated formulas. We say that F0 is directed, if: (a) If (8x)A; a] 2 F0 and Ax t]; b] 2 F0, then a b, where t 2 M . (b) If A is a logical axiom then A; a] 2 F0 implies a = AL (A). Note that if A(x); B(y) are independent then also all their respective instances are independent. The proof of the following lemma was inspired by the paper of E. Turunen 24]. 10

Lemma 2 Let F be directed set of independent evaluated formulas Let T = a A A; a] 2 F . Then there is a model D j= T such that D(A) = a (19) holds for all A; a] 2 F .
0 0 0

PROOF: We construct a Henkin extension TH of the theory T and a Lindenbaum algebra L(TH ) using the equivalence A B; i T ` A B: By Theorem 13 in 11], L(TH ) is a residuated lattice. Let j j denote the elements from L(TH ). Now, we construct an algebra Q generated by the set Q0 = fjAj j (9A)(9a)( A; a] 2 F0)g fj0jg: The Q is determined by the following conditions:

(a) Q0 Q. (b) If jAj; jB j 2 Q then jAj ! jB j := jA B j 2 Q.

Using the rule of modus ponens, logical axioms and formulas provable in the degree 1 (theorems) we can show that Q is a residuated lattice (analogously as in the proof of Theorem 13 in 11]). Let us now de ne the function f : Q ?! L as follows: (a) f(jAj) = a if A; a] 2 F0. (b) f(j0j) = 0. (c) f(jAj ! jB j) = f(jAj) ! f(jB j). Since F0 is directed set of independent formulas, the function f exists and it is a homomorphism. Using the results of 2], the lattice of truth values L in consideration is injective and thus, f can be extended to homomorphism g : L(TH ) ?! L: Finally, we de ne the truth evaluation H : FJ ?! L by H(A) = g(jAj). Obviously, H(A) = a for every V A; a] 2 F0. We will also show that H((8x)B) = t M H(Bx t]). As TH is Henkin and H is a homomorphism, it follows from the logical and Henkin axioms that H((8x)B) = H(Bx r]) where r is a special constant for (8x)B, both of the same sort . At the same time, H((8x)B) H(Bx t]) holds for every term t of the sort . If H(C) H(Bx t]) holds for all terms t then, H(C) H(Bx r]) = H((8x)B) as a special case, i.e. H((8x)B) is in mum of all the truth evaluations H(Bx t]), t 2 M J . Analogously we proceed for suprema, using the negation. Hence, using H, we can construct a canonical structure D, which is a model of the theory TH and has the property (19). But then D j= T follows from the fact that TH is a conservative extension of T. 2
2

This lemma plays an important role in proving some theorems about approximate reasoning. Using it and the completeness theorem we can prove the following lemma. 11

Lemma 3 Let Aj (x); Bj (y), j = 1; : : : be formulas, x; y variables of the sorts 1 and 2 such that for every j = k, Aj and Ak , as well as Bj and Bk are independent, respectively. Let k, 1 k m be given and 6 n o V T = akt Akx t]; cts m (Ajx t])Bjy s]) t 2 M ; s 2 M j
=1 1 2

be a fuzzy theory. Then

T `bks Bky s]; bks = PROOF: Put

"

t2M1

(akt cts);
#

s 2 M2 :

F0 = f Ajx t]; ajt]; Bjy s]; bjs =

It follows from the assumptions that F0 is a set of independent evaluated formulas which, obviously, is also directed. By Lemma 2, there exists a structure D such that D(Ajx t]) = ajt D(Bjy s]) = bjs: We will show that D j= T. Obviously, _ ajt cts (ajt cts) for all t 2 M1 and s 2 M2 and j. By the adjunction, we obtain _ cts ajt ! (ajt cts) = D(Ajx t]) ! D(Bjy s]) = D(Ajx t] Bjy s]) for all j = 1; : : :m, and thus cts i.e. D j= T. Consider the proofs
t2M1 t2M1

t2M1

(ajt cts) j t 2 M1 ; s 2 M2 ; j = 1; : : :mg:

m ^ j =1

D(Ajx t])Bjy s]) = D( (Ajx t])Bjy s]));

j =1
2 3

m ^

wts := Akx t]; akt]SA ; 4 (Ajx t] Bjy s]); cts 5 ; Bky s]; akt cts]rMPC ;
j =1

m ^

t 2 M1 , s 2 M2 where rMPC is a modi ed rule of modus ponens for the conjunction of implications (cf. 19, 16]). Then _ _ bks ValT (wts) = (akt cts): But at the same time, bks completeness property.

SA

D(Bky s]) =

t2M1

t2M1 t2M1 (akt cts ),

and we obtain the theorem using the 2

This lemma states that for special kinds of formulas, we may obtain the maximal provability degree only on the basis of multiformulas used in the de nition of the fuzzy theory in concern. In other words, the formula used in the generalized modus ponens for FLb gives the maximal possible value if we con ne only to linguistic expressions of the special kind. This is explicitly formulated in the following theorem. Let SA ; SB be two disjoint sets of syntagms (17). Using the translation rules from 19], the intension of each syntagm is a set of closed evaluated instances of a formula of the form c(p(x)) where c is a logically sound unary connective. 12

Theorem 10 Let the theory of FLb

T = Ak Ak ]; AND(IF Aj THEN Bj ) j
=1

h ^m (A j =1 j

)Bj i
9 =

be given using the above syntagms for some k, 1 k and

82 m < ^ Bj ) = : 4 (Ajx t] j =1 ("

m. Let Ak = f Akx t]; akt] j t 2 M1g

^m j =1 (Aj

)Bjy s]); cts5 t 2 M ; s 2 M ; :

1 2

Then we may derive the conclusion Bk with the intension

Bk =

Bky s]; bs =

such that all bs for s 2 M2 in the multiformula Bk are maximal.

t2M1

(akt cts) s 2 M2

(20)

PROOF: The linguistic description determines the fuzzy theory T = : akt Akx t]; cts The theorem then follows from Lemma 3.
8 <

m ^

j =1

(Ajx ( t] Bjy s]) t 2 M1 ; s 2 M2 ; :

9 =

In this theorem, we do not consider modi cation of the premise. However, this is possible when using special inference rule in fuzzy logic in narrow sense (cf. 19]). The following lemma is proved using the same methods. It demonstrates that standard Mamdani's Max-Min rule can be obtained as a consequence of some special axioms (see 14]) in which we consider a new predicate R(x; y) representing some function to be approximated. Furthermore, for the formulas in concern (i.e. those occurring in all the linguistic descriptions so far) the resulting computation formula gives the best possible result in the same sense as above.

Lemma 4 Given a fuzzy theory T = fAj ^Bj j j = 1; : : :; mg = = ff Ajx t]^Bjy s]; ajt ^ bjs] j t 2 M ; s 2 M g j j = 1; : : :; mg
0

where x; y are variables of di erent sorts, M1 ; M2 are the corresponding sets of terms (without variables) and for every j 6= k are Aj Bj and Ak Bk independent. Furthermore, put

T =T
If

f (8x)(8y)((Aj (x)^Bj (y)))R(x; y)); 1] j j = 1; : : :; mg:

0

(21)

T `at Ax t] t 2 M1 where A (x) is either Aj (x) for some j = 1; : : :; m or it is independent on all A1 (x) Bj (y) then T `bs By s] s 2 M2
0 0

where and B (y) := (9x)(A (x) ^ R(x; y)).

0 0

bs =
0

t2M1

(at ^
0

m _

j =1

(ajt ^ bjs))

13

PROOF: Using the instances of the substitution axiom, we obtain the provable evaluated formula (Ajx t] Bjy s]) Rx;y t; s]; 1], from which it follows that T `dts Rxy t; s] t 2 M1 ; s 2 M2 Wm where dts j =1(ajt ^ bjs). Then there is a set of proofs in T

wjt := Ax t]; at] ; Rxy t; s]; ajt ^ bjs]; : : :; Ax t] Rxy t; s]; at ^ ajt ^ bjs]; : : : (9x)(A (x) R(x)y s]); at ^ (ajt ^ bjs)]; t 2 M1 , j = 1; : : :; m where we have used the rule rMP , substitution axiom and its consequences. From it follows that
0 0 0

_
t2M1 j=1;:::;m

ValT (wjt) =

m _ _

t2M1 j =1

(at ^ (ajt ^ bjs)) =

0

t2M1

(at ^
0

m _

j =1

(ajt ^ bjs))

which gives where cs

_

T `cs By s]
0

As the formulas Aj Bj and A , j = 1; : : :; m are independent, there exists a model D j= T such that D (Ajx t] Bjy s]) = ajt ^ bjs; D (Ax t]) = at; j = 1; : : :m; t 2 M1 ; s 2 M2. Let us construct a model D as follows. We put D = D and, furthermore, D(Ajx t] Bjy s]) = D (Ajx t] Bjy s])
0 0 0 0

t2M1

(at ^
0

m _

j =1

(ajt ^ bjs)):

(22)

D(Rxy t; s]) =
0

m _

t 2 M1 , s 2 M2 , j = 1; : : :m and D(C) = D (C) for every formula C containing no instance of R(x; y). Then D(8x)(8y)((Aj (x) Bj (y)) R(x; y))) = ^ = (D(((Ajx t] Bjy s]) Rx;y t; s])) =

j =1

(ajt ^ bjs)

t2M1 ;s2M2 ^

) ^

and thus, D j= T. Finally, D(By s]) = D((9x)(A (x) ^ R(x)y s])) =

0

t2M1 ;s2M2

((ajt ^ bjs) !

m _

j =1

(ajt ^ bjs)) = 1

= i.e.

t2M1

D(Ax t] ^ Rxy t; s]) =

0

t2M1

(at ^
0

m _ j =1

(ajt ^ bjs));

cs

which which together with (22) gives the required equality. On the basis of Lemma 4 and analogously to Theorem 10 we can formulate the following theorem. 14

t2M1

(at ^
0

m _

j =1

(ajt ^ bjs))

Theorem 11 Let the theory of FLb

T = fA A ]; OR(Aj AND Bj ) j
0 0

=1

h _m i (Aj ^ Bj ) g j =1
0 0

be given using the same syntagms as in Theorem 10. Furthermore, let the axioms from (21) be added. Then we may derive a conclusion B assigned a formula B (y) := (9x)(A (x) ^ R(x; y)) and having the intension 9 82 3
0

B = : 4By s]; bs =
0 0 0 0

<

_
0

where all bs for s 2 M2 in the multiformula B are maximal.

t2M1

(at ^
0

m _

j =1

(ajt ^ bjs))5 s 2 M2 ;

This theorem, analogously as Theorem 10, explicitly states that for the syntagms widely used in fuzzy control, and provided that we assume (21), the Mamdani's Max-Min rule can be used to derive a conclusion which is the best possible one (in the sense of maximalization of truth values). Hence, we have basically two (from the point of view of truth values) e ective inference procedures: the rst one is based on sound inference rules of fuzzy logic in narrow sense and deals with logical implications, and the second one is based on additional assumptions and deals with conjunctions. Note that the latter is linguistic representation of the concept of fuzzy graph.

4 Conclusion
In this paper, we reviewed some formal aspects of fuzzy logics in narrow as well as in broader sense. The former can be viewed as a special many-valued logic aimed at modelling of the vagueness phenomenon. Therefore, it is slightly modi ed to ful l this goal. Most important is its ability to derive conclusions concerning any truth value, i.e. all the truth values are equally important. As a consequence, we obtain evaluated syntax in which evaluated formulas A; a] are considered. Such formulas, however, can be interpreted as non-evaluated ones of the form a A. We have discussed outcomes of this approach, and also showed that this logic may be considered as a direct generalization of the classical one due to the de nition of the syntactic consequence operation (De nition 1) which is equivalent with the classical de nition. In many-valued logic, however, the requirement that a provable formula must have a proof with the designated truth value is quite strong and restrictive for fuzzy logic in narrow sense. Slight weakening (cf. De nition 2 and Theorem 1) makes possible to keep the syntactico-semantical completeness. Unfortunately, this holds only in the case that we use Lukasiewicz implication since it is continuous. Interesting problem might be to classify fuzzy logics in narrow sense with respect to some \degrees of 1 completeness" which would be based on the general implication 1 ^ (1 ? ap + bp ) p , p 6= 0. The last section is devoted to fuzzy logic in broader sense which should be the logic of commonsense human deduction and thus, it is non-separably connected with linguistics. As natural language inherently encompasses vagueness, FLn becomes its frame and FLb can thus be seen as an extension of FLn. Natural language expressions are translated into multiformulas (sets of evaluated instances of formulas of FLn) which are interpreted as intensions of the former. We have proved two theorems which demonstrate that when con ning ourselves to restricted kinds of syntagms, the formulas widely used for generalized modus ponens both in implicational as well as Mamdani's forms give the best possible values. Godo and H jek 3] derived generalized modus ponens on the basis of purely logical considerations. They present several forms of this rule. Note that it is possible to express their rules also in the aggregated form using the concept of multiformula. Theorem 11 is based on the assumption (21) stating that, roughly speaking, the linguistic description concerns some relation between input and output (predicate R(x; y)). Godo and H jek use weaker assumption but they still keep a condition which inherently assumes some relation between the premise and the consequent.

15

References
1] Dubois, D. and H. Prade: Possibility Theory. An Approach to Computerized Processing of Uncertainty. Plenum, New York, 1988. 2] Glushankof, D. Prime deductive systems and injective objects in the algebras of Lukasiewicz in nitevalued calculi. Algebra Universalis 29(1992), 354{377. 3] L. Godo and P. H jek: On Deduction in Zadeh's Fuzzy Logic. Proc. Conf. IPMU'95, Granada 1995. 4] Gottwald, S.: Mehrwertige Logik. Akademie-Verlag, Berlin, 1989. 5] Gottwald, S.: Fuzzy Sets and Fuzzy Logic. Vieweg: Braunschweig, and Tecnea: Toulouse, 1993. 6] H jek P.: Fuzzy logic and arithmetical hierarchy. Fuzzy Sets and Systems 73(1995), 359{363. 7] H jek P.: Fuzzy logic as logic, In: G. Coletti et al., ed.: Mathematical Models of Handling Partial Knowledge in Arti cial Intelligence, Proc. Erice (Italy), Pergamon Press 1996. 8] H jek, P.: Metamathematics of fuzzy logic. Manuscript. To be published by Kluwer. 9] E. P. Klement, R. Mesiar, E. Pap: Additive generators of t-norms which are not necessarily continuous Proc. EUFIT'96, Aachen, 1996, pp. 70{73. 10] Mesiar, R. and V. Nov k: On Fitting Operations. Proc. of VIIth IFSA World Congress, Academia, Prague 1997. 11] Nov k, V.: On the Syntactico-Semantical Completeness of First{Order Fuzzy Logic. Part I | Syntactical Aspects; Part II | Main Results. Kybernetika 26(1990), 47{66; 134{154. 12] Nov k, V.: Logical Basis of Approximate Reasoning with Quanti ers. Proc. 12th Seminar on Fuzzy Set Theory, Johanes Kepler Universit t, Linz 1990. 13] Nov k, V.: On the logical basis of approximate reasoning, in V. Nov k, J. Ram k, M. Mare , M. ern and J. Nekola, Eds.: Fuzzy Approach to Reasoning and Decision Making. Academia, Prague and Kluwer, Dordrecht 1992. 14] Nov k, V.: Fuzzy Control from the Point of View of Fuzzy Logic. Fuzzy Sets and Systems 66(1994), 159{173. 15] Nov k, V.: Ultraproduct Theorem and Recursive properties of Fuzzy Logic. In: H hle, U. and E. P. Klement (eds.), Non-Classical Logics and Their Applications to Fuzzy Subsets. A Handbook of the Mathematical Foundations of Fuzzy Set Theory, Kluwer, Dordrecht 1995, 341{370. 16] Nov k, V.: Towards Formalized Integrated Theory of Fuzzy Logic. In: Bien, Z, and K. Min (eds.), 17] Nov k, V.: Linguistically Oriented Fuzzy Logic Controller and Its Design. Int. J. of Approximate Reasoning 1995, 12(1995), 263{277. 18] Nov k, V. A New Proof of Completeness of Fuzzy Logic and Some Conclusions for Approximate Reasoning. Proc. Int. Conference FUZZ-IEEE/IFES'95, Yokohama 1995, 1461{1468. 19] Nov k, V.: Paradigm, Formal Properties and Limits of Fuzzy Logic. Int. J. of General Systems 24(1996), 377{405. 20] Nov k, V: Open Theories, Consistency and Related Results in Fuzzy Logic. Int. J. of Approximate Reasoning 1997 (submitted). 21] Nov k, V. and I. Per lieva: On Logical and Algebraic Foundations of Approximate Reasoning. FUZZIEEE'97, Barcelona (to appear). 16

Fuzzy Logic and Its Applications to Engineering, Information Sciences, and Intelligent Systems, Kluwer, Dordrecht 1995, 353{363.

22] Nov k, V. and I. Per lieva: On Model Theory in Fuzzy Logic in Broader Sense. Proc. Int. Conference EUFIT'97, Verlag Mainz, Aachen 1997 (to appear). 23] Pavelka, J.: On fuzzy logic I, II, III, Zeit. Math. Logic. Grundl. Math. 25(1979), 45{52; 119{134; 447{464. 24] Turunen, E.: Well-De ned Fuzzy Sentential Logic. Math. Logic Quaterly 41(1995), 236{248. 25] B. Schweizer, A. Sklar: Probabilistic metric spaces North Holland, New York, 1983. 26] Zadeh, L.A.: The concept of a linguistic variable and its application to approximate reasoning I, II, III, Inf. Sci., 8(1975), 199{257, 301{357; 9(1975), 43{80.

17