Fuzzy relation equations are used to investigate theoretical and applicational aspects of fuzzy set theory, e.g., approximate reasoning, time series forecast, decision making and fuzzy control, etc.. This paper relates these equations to a particular kind of concept lattices. ? Partially supported by the Spanish Science Ministry TIN2009-14562-C05-03 and by Junta de Andaluc a project P09-FQM-5233. ?? Corresponding author.
Recently, multi-adjoint property-oriented concept lattices have been introduced in [16] as a generalization of property-oriented concept lattices [10, 11] to a fuzzy environment. These concept lattices are a new point of view of rough set theory [23] that considers two di erent sets: the set of objects and the set of attributes.
On the other hand, fuzzy relation equations, introduced by E. Sanchez [28], are associated to the composition of fuzzy relations and have been used to investigate theoretical and applicational aspects of fuzzy set theory [22], e.g., approximate reasoning, time series forecast, decision making, fuzzy control, as an appropriate tool for handling and modeling of nonprobabilistic form of uncertainty, etc. Many papers have investigated the capacity to solve (systems) of fuzzy relation equations, e.g., in [1, 8, 9, 25, 26].
In this paper, the multi-adjoint relation equations are presented as a generalization of the fuzzy relation equations [24, 28]. This general environment inherits the properties of the multi-adjoint philosophy, consequently, e.g., several conjunctors and residuated implications de ned on general carriers as lattice structures can be used, which provide more exibility in order to relate the variables considered in the system.
Moreover, multi-adjoint property-oriented concept lattices and systems of multi-adjoint relation equations have been related in order to obtain results that ensure the existence of solutions in these systems. These de nitions and results are illustrated by a toy example to improve the readability and comprehension of the paper.
Among all concept lattice frameworks, we have related the multi-adjoint property-oriented concept lattices to the systems of multi-adjoint relation equations, e.g., the extension and intension operators of this concept lattice can be used to represent multi-adjoint relation equations, and, as a result, the solutions of these systems of relation equations can be related to the concepts of the corresponding concept lattice.
The more important consequence is that this relation provides that the properties given, e.g., in [2{4,12,14,17,18,27] can be applied to obtain many properties of these systems. Indeed, it can be considered that the algorithms presented, e.g., in [5, 6, 15] obtain solutions for these systems.
The plan of this paper is the following: in Section 2 we will recall the multiadjoint property-oriented concept lattices as well as the basic operators used and some properties; later, in Section 3, an example will be introduced to motivate the multi-adjoint relation equations. Once these equations have been presented, in Section 4 the multi-adjoint property-oriented concept lattices and the systems of multi-adjoint relation equations will be related in order to obtain results which ensure the existence of solutions in these systems; the paper ends with some conclusions and prospects for future work. 2
The basic operators in this environment are the adjoint triples, which are formed by three mappings: a non-commutativity conjunctor and two residuated implications [13], which satisfy the well-known adjoint property.
De nition 1. Let (P1; 1), (P2; 2), (P3; 3) be posets and & : P1 P2 ! P3, . : P3 P2 ! P1, - : P3 P1 ! P2 be mappings, then (&; .; -) is an adjoint triple with respect to P1; P2; P3 if: 1. & is order-preserving in both arguments. 2. . and - are order-preserving on the rst argument1 and order-reversing on the second argument. 3. x 1 z . y i x & y 3 z i y 2 z - x, where x 2 P1, y 2 P2 and z 2 P3.
Example of adjoint triples are the Godel, product and Lukasiewicz t-norms together with their residuated implications.
Example 1. Since the Godel, product and Lukasiewicz t-norms are commutative, the residuated implications satisfy that .G=-G, .P =-P and .L=-L. Therefore, the Godel, product and Lukasiewicz adjoint triples are de ned on [0; 1] as: &P (x; y) = x y &G(x; y) = min(x; y) z -P x = min(1; z=x) z -G x = (1 if x z z otherwise &L(x; y) = max(0; x + y 1) z -G x = min(1; 1 x + z) 1 Note that the antecedent will be evaluated on the right side, while the consequent will be evaluated on the left side, as in logic programming framework. In [19] more general examples of adjoint triples are given.
The basic structure, which allows the existence of several adjoint triples for a given triplet of lattices, is the multi-adjoint property-oriented frame. De nition 2. Given two complete lattices (L1; 1) and (L2; 2), a poset (P; ) i and adjoint triples with respect to P; L2; L1, (&i; . ; -i), for all i = 1; : : : ; l, a multi-adjoint property-oriented frame is the tuple
(L1; L2; P; 1; 2; ; &1; .1; -1; : : : ; &l; .l; -l) Multi-adjoint property-oriented frames are denoted as (L1; L2; P; &1; : : : ; &l). Note that the notation is similar to a multi-adjoint frame [18], although the adjoint triples are de ned on di erent carriers.
The de nition of context is analogous to the one given in [18].
De nition 3. Let (L1; L2; P; &1; : : : ; &l) be a multi-adjoint property-oriented frame. A context is a tuple (A; B; R; ) such that A and B are non-empty sets (usually interpreted as attributes and objects, respectively), R is a P -fuzzy relation R : A B ! P and : B ! f1; : : : ; lg is a mapping which associates any element in B with some particular adjoint triple in the frame.2
From now on, we will x a multi-adjoint property-oriented frame and context, (L1; L2; P; &1; : : : ; &l), (A; B; R; ).
Now we de ne the following mappings " : L2B ! L1A and #N : L1A ! L2B as
g" (a) = supfR(a; b) & (b) g(b) j b 2 Bg
f #N (b) = infff (a) - (b) R(a; b) j a 2 Ag
(
Clearly, these de nitions3 generalize the classical possibility and necessity operators [11] and they form an isotone Galois connection [16]. There are two dual versions of the notion of Galois connetion. The most famous Galois connection, where the maps are order-reversing, is properly called Galois connection, and the other in which the maps are order-preserving, will be called isotone Galois connection. In order to make this contribution self-contained, we recall their formal de nitions:
Let (P1; 1) and (P2; 2) be posets, and # : P1 ! P2, " : P2 ! P1 mappings, the pair ("; #) forms a Galois connection between P1 and P2 if and only if: " and # are order-reversing; x 1 x#", for all x 2 P1, and y 2 y"#, for all y 2 P2.
The one we adopt here is the dual de nition: Let (P1; 1) and (P2; 2) be posets, and # : P1 ! P2, " : P2 ! P1 mappings, the pair ("; #) forms an isotone Galois connection between P1 and P2 if and only if: " and # are order-preserving; x 1 x#", for all x 2 P1, and y"# 2 y, for all y 2 P2. 2 A similar theory could be developed by considering a mapping : A ! f1; : : : ; lg which associates any element in A with some particular adjoint triple in the frame. 3 From now on, to improve readability, we will write &b, -b instead of & (b), - (b).
A concept, in this environment, is a pair of mappings hg; f i, with g 2 LB; f 2 LA, such that g" = f and f #N = g, which will be called multi-adjoint propertyoriented formal concept. In that case, g is called the extension and f , the intension of the concept. The set of all these concepts will be denoted as M N [16]. De nition 4. A multi-adjoint property-oriented concept lattice is the set
M N = fhg; f i j g 2 L2B; f 2 L1A and g" = f; f #N = gg in which the ordering is de ned by hg1; f1i f1 1 f2). hg2; f2i i g1 2 g2 (or equivalently
The pair (M N ; ) is a complete lattice [16], which generalize the concept lattice introduced in [7] to a fuzzy environment. 3
This section begins with an example that motivates the de nition of multiadjoint relation equations, which will be introduced later. 3.1
Multi-adjoint logic programming A short summary of the main features of multi-adjoint languages will be presented. The reader is referred to [20, 21] for a complete formulation.
A language L contains propositional variables, constants, and a set of logical connectives. In this fuzzy setting, the usual connectives are adjoint triples and a number of aggregators.
The language L is interpreted on a (biresiduated) multi-adjoint lattice,4 1 n hL; ; . ; -1; &1; : : : ; . ; -n; &ni, which is a complete lattice L equipped with i a collection of adjoint triples h&i; . ; -ii, where each &i is a conjunctor intended to provide a modus ponens -rule with respect to .i and -i.
A rule is a formula A .i B or A -i B, where A is a propositional symbol (usually called the head) and B (which is called the body) is a formula built from propositional symbols B1; : : : ; Bn (n 0), truth values of L and conjunctions, disjunctions and aggregations. Rules with an empty body are called facts.
A multi-adjoint logic program is a set of pairs hR; i, where R is a rule and is a value of L, which may express the con dence which the user of the system has in the truth of the rule R. Note that the truth degrees in a given program are expected to be assigned by an expert.
Example 2. Let us to consider a multi-adjoint lattice h[0; 1]; ;
G; &G;
P ; &P ; ^Li 4 Note that a multi-adjoint lattice is a particular case of a multi-adjoint propertyoriented frame. where &G and &P are the Godel and product conjunctors, respectively, and
G, P their corresponding residuated implications. Moreover, the Lukasiewicz conjunctor ^L will be used in the program [13].
Given the set of variables (propositional symbols) = flow oil; low water; rich mixture; overheating; noisy behaviour; high fuel consumptiong the following set of multi-adjoint rules form a multi-adjoint program, which may represent the behaviour of a motor.
hhigh fuel consumption
hoverheating hnoisy behaviour hoverheating G low oil; 0:5i
hnoisy behaviour
G low oil; 1i
The usual procedural is to measure the levels of \oil", \water" and \mixture" of a speci c motor, after that the values for low oil, low water and rich mixture are obtained, which are represented in the program as facts, for instance, the next ones can be added to the program:
hlow oil hlow water hrich mixture
P >; 0:2i P >; 0:2i P >; 0:5i
Finally, the values for the rest of variables (propositional symbols) are
computed [20]. For instance, in order to attain the value for overheating(o; w), for
a level of oil, o, and water, w, the rules hoverheating G low oil; #1i and
hoverheating P low water; #2i are considered and its value is obtained as:
overheating(o; w) = (low oil(o) &G #1) _ (low water(w) &P #2)
(
Now, the problem could be to recompute the weights of the rules from experimental instances of the variables, that is, the values of overheating, noisy behaviour and high fuel consumption are known for particular measures of low oil, low water and rich mixture.
Speci cally, given the levels of oil, o1; : : : ; on, the levels of water, w1; : : : ; wn, and the measures of mixture, t1; : : : ; tn, we may experimentally know the values of the variables: noisy behaviour(ti; oi), high fuel consumption(ti; oi) and overheating(oi; wi), for all i 2 f1; : : : ; ng.
Considering Equation (
This system can be interpreted as a system of fuzzy relation equations in which several conjunctors, &G and &P , are assumed. Moreover, these conjunctors could be neither non-commutative nor associative and de ned in general lattices, as permit the multi-adjoint framework.
Next sections introduce when these systems have solutions and a novel method to obtain them using concept lattice theory. 3.2
Systems of multi-adjoint relation equations The operators used in order to obtain the systems will be the generalization of the sup- -composition, introduced in [29], and inf-!-composition, introduced in [1]. From now on, a multi-adjoint property-oriented frame, (L1; L2; P; &1; : : : ; &l) will be xed.
In the de nition of a multi-adjoint relation equation an interesting mapping : U ! f1; : : : ; lg will be considered, which relates each element in U to an adjoint triple. This mapping will play a similar role as the one given in a multiadjoint context, de ned in the previous section, for instance, this map provides a partition of U in preference sets. A similar theory may be developed for V instead of U .
Let U = fu1; : : : ; umg and V = fv1; : : : ; vng be two universes, R 2 L2U V an unknown fuzzy relation, : U ! f1; : : : ; lg a map that relates each element in U to an adjoint triple, and K1; : : : ; Kn 2 P U , D1; : : : ; Dn 2 L1V arbitrarily chosen fuzzy subsets of the respective universes.
A system of multi-adjoint relation equations with sup-&-composition, is the following system of equations _ (Ki(u) &u R(u; v)) = Di(v); i 2 f1; : : : ; ng u2U where &u represents the adjoint conjunctor associated to u by , that is, if s (u) = (&s; . ; -s), for s 2 f1; : : : ; lg, then &u is exactly &s.
If an element v of V is xed and the elements Ki(uj ), R(uj ; v) and Di(v) are
written as kij , xj and di, respectively, for each i 2 f1; : : : ; ng, j 2 f1; : : : ; mg,
then System (
_ k1m &um xm = d1
... ... ...
_ knm &um xm = dn where kij and di are known and xj must be obtained.
Hence, for each v 2 V , if we solve System (
The propositional symbols (variables) will be written in short as: hfc, nb, oh, rm, lo and lw, and the measures of particular cases of the behaviour of the motor will be: hi, ni, ovi, ri, oi, wi, for hfc, nb, oh, rm, lo and lw, respectively, in each case i, with i 2 f1; 2; 3g.
For instance, the next system associated to overheating is obtained from the computation provided in Example 2. where #olho and #olhw are the weights associated to the rules with head oh. Similar systems can be obtained to high fuel consumption and noisy behaviour.
Assuming the multi-adjoint frame with carrier L = [0; 1] and the Godel and product triples, these systems are particular systems of multi-adjoint relational equations. The corresponding context is formed by the sets U = frm; lo; lw; rm^L log, V = fhfc; nb; ohg; the mapping that relates the elements lo, rm ^L lo to the Godel triple, and rm, lw to the product triple; the mappings K1; : : : ; Kn 2 P U , de ned as the values given by the propositional symbols in U on the experimental data, for instance, if u = lo, then K1(lo) = lo(o1), . . . , Kn(lo) = lo(on); and the mappings D1; : : : ; Dn 2 L1V , de ned analogously, for instance, if v = rm, then D1(rm) = rm(r1), . . . , Dn(rm) = rm(rn).
Finally, the unknown fuzzy relation R 2 L2U V is formed by the weights of the rules in the program.
In the system above, oh has been the element v 2 V xed. Moreover, as there do not exist rules with body rm and rm ^L lo, that is, the weights for that hypothetical rules are 0, then the terms (rm(ri) &G 0 = 0 and (rm(ri) ^L lo(oi) &P 0 = 0 do not appear.
Its counterpart is a system of multi-adjoint relation equations with inf--composition, that is, ^ (R(u; v) -uj Kj (v)) = Ej(u); j 2 f1; : : : ; mg v2V considered with respect to unknown fuzzy relation R 2 L1U V , and where K1 ; : : : ; Km 2 P V and E1; : : : ; Em 2 L2U . Note that -uj represents the corresponding adjoint implication associated to uj by , that is, if (uj) = (&s; .s ; -s), for s 2 f1; : : : ; lg, then -uj is exactly -s. Remark that in System 6, the implication -uj does not depend of the element u, but of j. Hence, the implications used in each equation of the system are the same.
If an element u of U is xed, fuzzy subsets K1 ; : : : ; Km 2 P V , E1; : : : ; Em 2
L2U are assumed, such that Kj (vi) = kij, R(u; vi) = yi and Ej(u) = ej, for each
i 2 f1; : : : ; ng, j 2 f1; : : : ; mg, then System (
Systems (
Analyzing these systems, we have that the left side of these systems can be
represented by the mappings CK : Lm 1 ! L2m, de ned as:
2 ! L1n; IK : Ln
CK (x)i = ki1 &u1 x1 _
IK (y)j = y1 -uj k1j ^
_ kim &um xm; for all i 2 f1; : : : ; ng
^ yn -uj knj ; for all j 2 f1; : : : ; mg
(
Hence, Systems (
CK (x1; : : : ; xm) = (d1; : : : ; dn)
IK (y1; : : : ; yn) = (e1; : : : ; em)
(
This section shows that Systems (
First of all, the environment must be xed. Hence, a multi-adjoint context (A; B; S; ) will be considered, such that A = V 0, B = U , where V 0 has the same cardinality as V , will be the mapping given by the systems and S : A B ! P is de ned as S(vi0; uj ) = kij . Note that A = V 0 is related to the mappings Ki, since S(vi0; uj ) = kij = Ki(uj );
Now, we will prove that the mappings de ned at the end of the previous section are related to the isotone Galois connection. Given 2 L2B, such that (uj ) = xj , for all j 2 f1; : : : ; mg, the following equalities are obtained, for each i 2 f1; : : : ; ng:
CK (x)i = ki1 &u1 x1 _
_ kim &um xm = S(vi0; u1) &u1 (u1) _ _ S(vi0; um) &um (um) = supfS(vi0; uj ) &uj (uj ) j j 2 f1; : : : ; mgg = " (vi0)
Therefore, the mapping CK : L2m ! L1n is equivalent to the mapping " : L2B ! L1A, where an element x in L2m can be interpreted as a map in L2B, such that (uj ) = xj , for all j 2 f1; : : : ; mg, and the element CK (x) as the mapping " , such that " (vi0) = CK (x)i, for all i 2 f1; : : : ; ng.
An analogy can be developed applying the above procedure to mappings IK and #N , obtaining that the mappings IK : L1n ! L2m and #N : L1A ! L2B are equivalent.
As a consequence, the following result holds: Theorem 1. The mappings CK : Lm 1 ! L2m, establish an iso2 ! L1n; IK : Ln tone Galois connection. Therefore, IK CK : L2m ! L2m is a closure operator and CK IK : L1n ! L1n is an interior operator.
As (CA; IK ) is an isotone Galois connection, any result about the solvability of one system has its dual counterpart.
The following result explains when these systems can be solved and how a solution can be obtained.
Theorem 2. System (
Similarly, System (
The main contribution of the relation introduced in this paper is not only the
above consequences, but a lot of other properties for Systems (
Next example studies the system of multi-adjoint relation equations presented in Example 3.
Example 4. The aim will be to solve a small system in order to improve the understanding of the method. In the environment of Example 3, the following system will be solved assuming the experimental data: oh(ov1) = 0:5, lo(o1) = 0:3, lw(w1) = 0:3, oh(ov2) = 0:7, lo(o2) = 0:6, lw(w2) = 0:8, oh(ov3) = 0:4, lo(o3) = 0:5, lw(w3) = 0:2.
oh(ov1) = (lo(o1) &G #loho) _ (lw(w1) &P #lohw) oh(ov2) = (lo(o2) &G #loho) _ (lw(w2) &P #lohw) oh(ov3) = (lo(o3) &G #loho) _ (lw(w3) &P #lohw) where #olho and #olhw are the variables.
The context is: A = V 0 = f1; 2; 3g, the set of observations, B = U = flo; lwg, associates the propositional symbol lo to the Godel triple and lw to the product triple. The relation S : A B ! [0; 1] is de ned in Table 1.
Therefore, considering the mapping oh : A ! [0; 1] associated to the values
of overheating in each experimental case, that is oh(
CK (#oloh; #olwh) = oh Since, by the comment above, there exists 2 [0; 1]B, such that CK (#olho; #olhw) = " , the goal will be to attain the mapping 2 [0; 1]B, such that " = oh, which can be found if and only if (( oh)#N ; oh) is a multi-adjoint propertyoriented concept in the considered context, by Theorem 2.
First of all, we compute ( oh)#N .
( oh)#N (lo) = inff oh(
= inff1; 0:875; 1g = 0:875 Now, the mapping ( oh)#N " is obtained.
( oh)#N " (
Therefore, (( oh)#N ; oh) is not a multi-adjoint property-oriented concept and thus, the considered system has no solution, although if the experimental value for oh had been 0:3 instead of 0:5, the system would have had a solution.
These changes could be considered in several applications where noisy variables exist and their values can be conveniently changed to obtain approximate solutions for the systems. Thus, if the experimental data for overheating are oh(ov1) = 0:3, oh(ov2) = 0:7 and oh(ov2) = 0:4, then the original system will have at least one solution and the values #olho, #olhw will be 0:4, 0:875, respectively for a solution. Consequently, the truth for the rst rule is lower than for the second or it might be thought that it is more determinant in obtaining higher values for lw than for lo. Another possibility is to consider that this conclusion about the certainty of the rules is not correct, in which case another adjoint triple might be associate to lo.
As a result, the properties introduced in several fuzzy formal concept analysis frameworks can be applied in order to obtain solutions of fuzzy relation equations, as well as in the multi-adjoint general framework.
Furthermore, in order to obtain the solutions of Systems (
Multi-adjoint relation equations have been presented that generalize the existing de nitions presented at this time. In this general environment, di erent conjunctors and residuated implications can be used, which provide more exibility in order to relate the variables considered in the system.
A toy example has been introduced in the paper in order to improve its readability and reduce the complexity of the de nitions and results.
As a consequence of the results presented in this paper, several of the properties provided, e.g., in [2{4, 12, 14, 17, 18, 27], can be used to obtain additional characteristics of these systems.
In the future, we will apply the results provided in the fuzzy formal concept analysis environments to the general systems of fuzzy relational equations presented here.