# minimal_undefinedness_for_fuzzy_answer_sets__01a9966b.pdf Minimal Undefinedness for Fuzzy Answer Sets Mario Alviano, Giovanni Amendola Department of Mathematics and Computer Science University of Calabria, Italy {alviano,amendola}@mat.unical.it Rafael Pe naloza KRDB Research Centre Free University of Bozen-Bolzano, Italy rafael.penaloza@unibz.it Fuzzy Answer Set Programming (FASP) combines the nonmonotonic reasoning typical of Answer Set Programming with the capability of Fuzzy Logic to deal with imprecise information and paraconsistent reasoning. In the context of paraconsistent reasoning, the fundamental principle of minimal undefinedness states that truth degrees close to 0 and 1 should be preferred to those close to 0.5, to minimize the ambiguity of the scenario. The aim of this paper is to enforce such a principle in FASP through the minimization of a measure of undefinedness. Algorithms that minimize undefinedness of fuzzy answer sets are presented, and implemented. Introduction Fuzzy Answer Set Programming (FASP) (Nieuwenborgh, Cock, and Vermeir 2007; Janssen et al. 2012a; 2012b; Blondeel et al. 2013; Lee and Wang 2014; Mushthofa, Schockaert, and Cock 2015) is a successfully combination of Fuzzy Logic (Cintula, H ajek, and Noguera 2011) and Answer Set Programming (ASP) (Gelfond and Lifschitz 1991; Marek and Truszczy nski 1999; Niemel a 1999), which resulted in a declarative framework supporting non-monotonic reasoning on propositional formulas interpreted by truth degrees in the interval [0,1]. As in ASP, reasoning on unknown knowledge is eased by the use of default negation, whose semantics is elegantly captured by the notion of answer set, or stable model: in a model, truth of unknown knowledge may be assumed as soon as there is no evidence of the contrary, and the model is discarded when the truth of some propositions is not necessary in order to satisfy the input program under the assumption for the unknown knowledge provided by the model itself. Moreover, as in Fuzzy Logic, non-Boolean truth degrees are useful to handle vague information, but also to deal with inconsistencies that may arise from mathematical abstractions of real phenomena. In this respect, the truth degree 0.5 is analogous to the truth value undefined used by many paraconsistent logics and paracoherent answer set semantics (Przymusinski 1991; You and Yuan 1994; Sakama and Inoue 1995; Eiter, Leone, and Sacc a 1997; Amendola et al. 2016). Fuzzy answer sets satisfy two fundamental principles shared by several semantics for logic programs: justifia- Copyright c 2017, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. bility (J, or foundedness), and the closed world assumption (C, or CWA). Briefly, for normal ASP programs, justifiability requires that every true atom is derived from the input program under the assumption provided for the negated formulas, and the CWA constrains to false any atom whose defining rules have false bodies (You and Yuan 1994; Eiter, Leone, and Sacc a 1997). For FASP programs, these two principles can be recast in terms of truth degrees: (JC ) Any truth degree x (0,1] for an atom p is derived from the input program under the assumption provided for the negated formulas. Minimal undefinedness (U) is another fundamental principle introduced in the context of paraconsistent reasoning, imposing a minimization on the number of undefined atoms (You and Yuan 1994). For example, the FASP program Π = {p q, q p} has three answer set candidates, namely I = {(p,1),(q,0)}, J = {(p,0),(q,1)}, and K = {(p,0.5),(q,0.5)}, but K has to be discarded because it contains two undefined atoms. In terms of truth degrees, minimal undefinedness imposes the following preference: (U ) Truth degrees close to 0 or 1 should be preferred to those close to 0.5. However, minimal undefinedness is not enforced by the current notion of fuzzy answer set, as for example K is among the fuzzy answer sets of the program Π above. In fact, Π has uncountably many fuzzy answer sets of the form {(p,x),(q,1 x)}, x [0,1]. I and J should be the preferred two fuzzy answer sets for being undefinedness-free. The previous example also highlights that fuzzy answer sets do not possess another important principle known as congruence (Co): The extension of a semantics must coincide with the original semantics on coherent theories. In terms of fuzzy ASP, congruence can be stated as follows: (Co ) Fuzzy answer sets of coherent ASP programs coincide with crisp answer sets. Principles (U ) and (Co ) are useful in abduction processes involving fuzzy circuits. For example, the designer of a fuzzy controller may be interested in computing input configurations producing a given output. This abduction process can be improved by focusing on minimal undefined interpretations, as those are the simplest to explain in the real world. Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence (AAAI-17) Example 1. Consider a fuzzy controller with temperature and humidity as input, and fan speed as output. Let t1,t2,t3 and h1,h2,h3 be atoms representing different classes of temperatures and humidities (e.g., t1 is cool, t2 is normal, and t3 is warm; h1 is humidity low, h2 is humidity normal, and h3 is humidity high) and s1,s2,s3 represent different classes of fan speeds (e.g., s1 is off, s2 is normal speed, and s3 is maximum speed). The fuzzy controller rules are Πc = {s1 t1, s2 t2 h3, s3 t3 (h2 h3)}; i.e., the fan is turned off when the temperature is cool, set to normal speed when the temperature is normal and the humidity is high, and set to maximum speed when the temperature is warm and the humidity is normal or high. A possible input for the controller is Πin = {t1 0,t2 0.8,t3 0.2,h1 0,h2 0.1, h3 0.9}, representing a slightly warm temperature and a high humidity. In this case the controller sets the fan speed to a value slightly higher than normal. Indeed, as will be clear after formally introducing the semantics in the next section, the truth degrees of s1,s2,s3 are respectively 0, 0.7, and 0.2. In this context, the designer of the fuzzy controller may be interested in checking the existence of input configurations such that all output variables s1,s2,s3 are assigned the truth degree 0, hence leaving the fan speed completely undetermined. For example, in Πc this is the case when h1,t3 are 1, and all other input variables are 0. In this paper the notion of fuzzy answer sets is refined by means of a measure of undefinedness, which results into the definition of minimal undefinedness fuzzy answer set. Principles (JC ), (U ), and (Co ) are satisfied by FASP after this refinement. Algorithms to iteratively compute fuzzy answer sets that decrease the measure of undefinedness are also presented, among them binary and progression search. The algorithms are implemented in a prototype system extending FASP2SMT (Alviano and Pe naloza 2015), a FASP solver using Z3 (de Moura and Bjørner 2008) as backend. The performance of these algorithms is compared, on the same solver, with the use of the MINIMIZE instruction of Z3. Preliminaries Let B be a fixed set of propositional variables. A fuzzy atom (atom for short) is either a variable from B, or a numeric constant in the interval [0,1]. Fuzzy expressions are defined inductively as follows: every atom is a fuzzy expression; if α and β are fuzzy expressions and { , , , } is a connective, then α and α β are fuzzy expressions, where denotes default negation. The connectives , are called Łukasiewicz, and , are the G odel connectives. A positive expression is a fuzzy expression not using . A rule is of the form α β, with α an atom, and β a fuzzy expression. This rule is positive if β is a positive expression. A (general) FASP program is a finite set of rules. A positive FASP program is a FASP program where every rule is positive. We use At(Π) to denote the set of atoms occurring in Π. The semantics of FASP programs requires a set of truth degrees. For the scope of this paper we consider the real interval L = [0,1] and the sets Ln = { i n 1|i = 0,...,n 1}, for n 2. Note that L2 = {0,1} is the classical Boolean set. Henceforth, L denotes an arbitrary but fixed such set, unless explicitly stated otherwise. A FASP program Π is an L -program if all constants occurring in Π are in L . An L -interpretation I is a function I : B L mapping each atom of B into a truth degree in L . We will often denote such a function as the set {(p,I(p)) | p B}. I is extended to fuzzy expressions as follows: I(c) = c for all c [0,1]; I( α) = 1 I(α); I(α β) = max{I(α) + I(β) 1,0}; I(α β) = min{I(α) + I(β),1}; I(α β) = max{I(α),I(β)}; and I(α β) = min{I(α),I(β)}. I satisfies the rule α β (I |= α β) if I(α) I(β); it is an L -model of an L -program Π (I |= Π) if I |= r for all r Π. Given L -interpretations I,J, we write I J if I(p) J(p) for all p B. If I J and I = J, we write I < J. A minimal L -model of an L -program Π is an L -model of Π such that there is no L -model J satisfying J < I. The reduct of an L -program Π w.r.t. an L -interpretation I, denoted ΠI, is obtained by replacing each occurrence of α by the constant 1 I(α). Let Π be an L -program, and I be L -interpretation. I is an L -answer set of Π if I is a minimal L -model of ΠI. FAS(L ,Π) is the set of L -answer sets of Π. Π is L -coherent if FAS(L ,Π) = /0, and L -incoherent otherwise. Example 2. Let Π = {a b, b c, b 0.4}, and I = {(a,0),(b,1),(c,0)}. I FAS(L ,Π) because I is a minimal L -model of ΠI = {a 0,b 1,b 0.4}. ASP programs can be seen as special cases of FASP L2-programs using only the connective . The classical answer sets of such programs are FAS(L2,Π). Justifiability and CWA Let us first introduce the immediate consequence operator TΠ for a positive L -program Π: TΠ(I) : α max{I(β) | α β Π}, α B (1) where I is an L -interpretation. The operator is monotonic, and thus has a least fixpoint lfp(TΠ). Given an L -program Π and an L -interpretation I, an atom p B is justified in Π and I if I(p) = lfp(TΠI)(p). Note that the immediate consequence operator is applied to the reduct ΠI, so that lfp(TΠI)(p) actually derives the truth degree of p from the input program under the assumption provided by I for the negated expressions in Π. Theorem 1 (Justifiability). Let Π be an L -program, and I FAS(L ,Π). All atoms p B such that I(p) (0,1] are justified in Π and I. Proof. We show that lfp(TΠI) equals I. The reduct ΠI has the unique L -answer set lfp(TΠI) (Lukasiewicz 2006). Hence, there is no J < lfp(TΠI) with J |= (ΠI)I. Since (ΠI)I = ΠI, we have that lfp(TΠI) is a minimal L -model of ΠI. It follows that lfp(TΠI) = I because I FAS(L ,Π) by assumption. Consequently, atoms not occurring in the left-hand side of any rule of Π, among them those in B \ At(Π), are associated with truth degree 0 in any L -answer set of Π. Corollary 1 (CWA). Let Π be an L -program, p B, and I FAS(L ,Π). If I(β) = 0 for all p β Π, then I(p) = 0. As stated in the introduction, justifiability and CWA are important properties of fuzzy answer sets, but insufficient to produce simple explanations in abductive reasoning. The next example clarifies this aspect. Example 3. Let Πabd be Πc from Example 1 extended with the following rules: p p p {t1,t2,t3,h1,h2,h3} (2) 0 (t1 t2 t3) (3) 0 (h1 h2 h3) (4) 0 s1 s2 s3 (5) Fuzzy answer sets of Πabd correspond to input configurations of the fuzzy controller in Example 1 such that all output variables are assigned 0. Indeed, (2) guesses an input configuration I, (3) (4) enforce I(t1)+I(t2)+I(t3) 1 and I(h1)+I(h2)+I(h3) 1 (some restrictions are omitted for simplicity), and (5) discards I if I(s1)+I(s2)+I(s3) > 0. Two fuzzy answer sets of Πabd are {(t1,0),(t2,0.5),(t3,0.5),(h1,0.5),(h2,0.5),(h3,0),(s1,0), (s2,0),(s3,0)} and {(t1,0),(t2,0),(t3,1),(h1,1),(h2,0), (h3,0),(s1,0),(s2,0),(s3,0)}. The latter is a simpler explanation, and should be preferred to the former. Minimal Undefined Fuzzy Answer Sets Fuzzy answer sets can be seen as a kind of imprecise answer set, where the interpretation of some of the atoms may not be fully defined. We want to restrict our attention only to those fuzzy answer sets that minimize the undefinedness according to an appropriate measure. Following De Luca and Termini (1972), a measure of undefinedness is a function U mapping every interpretation I to a non-negative real number U(I) R such that: (P1) U(I) = 0 if and only if I(p) {0,1} for all p B; (P2) if for every p B either (i) J(p) I(p) 0.5, or (ii) J(p) I(p) 0.5, then U(I) U(J); and (P3) let I(p) := 1 I(p) for all p B; then U(I) = U( I). Intuitively, these properties state, respectively, that classical interpretations are fully defined; interpretations closer to the extreme degrees are more defined (or less undefined); and undefinedness is symmetric w.r.t. complementary interpretations. Given a measure of undefinedness U,U(J) |J(q) 0.5|, then U(I) ε do 5 b := (lb+ub)/2; 6 (sat,I) := solve(Π,U({(p,xp) | p At(Π)}) b); 7 if sat then (I ,ub) := (I,U(I)); 8 else lb := b; 9 return I ; to discard any interpretation I such that U(I) > b. Moreover, for an L -program Π and a formula of the form (7), let solve(Π,ϕ) denote the invocation of a function computing a fuzzy answer set I of Π satisfying ϕ, if it exists; in this case, the function returns (true,I), and otherwise it returns (false, ). Abusing of notation, we also use solve(Π) to invoke function solve on Π alone, with no bound on the measure of undefinedness. The first algorithm we consider is based on binary search (Algorithm 1). Its input is an L -program, a measure of undefinedness U, and a precision threshold ε R+. The algorithm returns either the string incoherent, if FAS(L ,Π) = /0 (lines 1 2), or an L -answer set I FAS(L ,Π) such that U(I ) U(J) < ε for all J MUFAS(L ,Π,U). If Π is coherent, the algorithm initializes a lower bound lb and an upper bound ub (line 3): the upper bound is the measure of undefinedness of the current solution, while the lower bound represents a measure of undefinedness that cannot be achieved by any L -answer set of Π. Then, the algorithm searches for an L -answer set whose measure of undefinedness is at most (lb + ub)/2 (lines 5 6); if one is found, the current solution and the upper bound are updated (line 7), otherwise the lower bound is set to (lb + ub)/2: there is no I FAS(L ,Π) with U(I) (lb + ub)/2 (line 8). The process is repeated until the possible improvement of the upper bound is below the given precision threshold ε. Example 6. Consider the program Π = {a b,b a} from Example 4, with the measure UD, and ε = 0.1. solve(Π) returns a fuzzy answer set of Π; say I0 = {(a,0.6),(b,0.4)}. Then lb 0.1 and ub 0.8. The algorithm then finds a new answer set I1 with U(I1) 0.35; say I1 = {(a,0.1), (b,0.9)}, and updates the upper bound ub 0.2. At the next iteration we get I2 = {(a,0.05),(b,0.95)}, so ub 0.1. The algorithm sets b 0, and the answer set I3 = {(a,1),(b,0)} is retrieved and returned. The second method (Algorithm 2) uses progression search to find MUFAS. In this case, the undefinedness is bound to ub pr, where pr is the required improvement to the current solution. The variable pr is initially set to the precision threshold ε (line 3), doubled at each iteration (line 6), and reset to ε when it becomes too large (line 5). Example 7. Using the input from Example 6, Algorithm 2 finds the fuzzy answer set I0 = {(a,0.6), (b,0.4)}, initializing (lb,ub, pr) ( 0.1,0.8,0.1). It then searches for a new fuzzy answer set I1 with UD(I1) 0.7; say I1 = {(a,0.3), Algorithm 2: Progression Search(Π, U, ε) 1 (sat,I ) := solve(Π); 2 if not sat then return incoherent; 3 (lb,ub, pr) := ( ε,U(I ),ε); 4 while ub lb > ε do 5 if ub pr lb then pr := ε; 6 (b, pr) := (ub pr,2 pr); 7 (sat,I) := solve(Π,U({(p,xp) | p At(Π)}) b); 8 if sat then (I ,ub) := (I,U(I)); 9 else lb := b; 10 return I ; (b,0.7)}, updating ub 0.6 and pr 0.2. The next iteration restricts answer sets to have a measure of at most 0.4, yielding e.g. I2 = {(a,0.1), (b,0.9)} and updating ub 0.2, pr 0.4. Since ub pr 0, pr is reset to 0.1, and a new solution with UD(I3) 0.1 is seeked. The next iterations find I3 = {(a,0.05),(b,0.95)} and I4 = {(a,1),(b,0)}. At this point, ub lb = ε, and I4 is given as a solution. Finally, we consider a third algorithm which we call ε-improvement. This method is obtained from Algorithm 1 by replacing line 5 to update b as follows: 5 b := ub ε Intuitively, the modified algorithm minimally improves the measure of undefinedness of the current solution, until an incoherence arises. Example 8. Consider the input from Example 6, yielding the first solution I0 = {(a,0.6), (b,0.4)} and lb 0,up 0.8. The next solution should have a measure of at most 0.7; hence, the algorithm retrieves e.g. I1 = {(a,0.3), (b,0.7)}. At the next iteration, b is set to 0.5, which yields a new solution like I2 = {(a,0.1), (b,0.9)}. The next solution should then be bounded by 0.1. Thus, I3 = {(a,0.05),(b,0.95)} and I4 = {(a,1),(b,0)} are retrieved. The latter is returned. Implementation and Evaluation The three algorithms presented above were implemented in a prototype system extending FASP2SMT (Alviano and Pe naloza 2015) with the distance function UD as the measure of undefinedness. Other measures of undefinedness can be easily accommodated in the prototype, but left for future extensions. Briefly, FASP2SMT parses a symbolic L -program, which is then rewritten into ASP Core 2.0 (Alviano et al. 2013) and grounded by GRINGO (Gebser et al. 2011). The output of GRINGO encodes a propositional L -program Π, which is parsed and rewritten into SMT-Lib (Barrett, Stump, and Tinelli 2010), and processed by Z3 (de Moura and Bjørner 2008) to compute an L -answer set, if it exists. The new prototype keeps Z3 online, adds a formula of the form (6), and asks for a new model. Then, the formula (6) is removed, and this iteration is repeated until the desired precision threshold is reached. Besides the three algorithms presented here, FASP2SMT can use the minimize instruction of Z3, which however is experimental and not part of the SMT-Lib format. In this case Z3 runs silently to completion, without intermediate solutions provided to the user. The prototype system was tested empirically on satisfiable instances of Hamiltonian Cycle from the literature (Mushthofa, Schockaert, and Cock 2014; Alviano and Pe naloza 2015). Each method was tested with threshold ε {0.1,0.01,0.001}, on an Intel Xeon CPU 2.4 GHz with 16 GB of RAM. Time and memory were limited to 600 seconds and 15 GB, respectively. The results of the experiment are reported in Table 1. Binary search exhibits the best performance: it solves all instances if the required precision ε is set to 0.1, and anyhow the majority of the testcases if ε is smaller. The value of ε significantly affects the algorithm based on progression. Indeed, this algorithm reached the best performance when ε was set to 0.01. The larger value 0.1 resulted in several expensive invocations of function solve returning incoherent, while the smaller value 0.001 slowed down the search in some cases. Even if the algorithm is often unable to terminate, it can still provide to the user fuzzy answer sets with a good guarantee on the measure of undefinedness (the difference between lower and upper bound is usually lesser than 1). Algorithm ε-improvement reports a very bad performance, with several timeouts and not negligible bound differences. Using the minimize construct of Z3 is not an option here, as its performance was similar to ε-improvement. Related Work FAS and MUFAS are related to paracoherent answer set semantics. First we consider 3-valued stable models semantics (one of the best known approximations of answer sets) introduced in (Przymusinski 1991). Each 3-valued stable model (where false stands for 0, true for 1, and undefined for 0.5) corresponds to a fuzzy answer set. However, the reverse statement does not hold. For example, the only 3-valued stable models of the program Π from Example 4 are {a, b}, { a,b}, and /0, which correspond to {(a,1), (b,0)}, {(a,0), (b,1)} and {(a,0.5),(b,0.5)}, respectively. On the other hand, Π has infinitely many fuzzy answer sets. However, if we restrict to the L3 semantics for programs using only the Table 1: Experimental results on Hamiltonian Cycle instances: solved instances and average bound difference after 600 seconds; average running time and memory consumption on solved instances is also reported. ε binary progress ε-impr. min. solved 100% 22% 0% 6% ub lb 0.073 0.409 7.157 time (s) 149 206 10 mem. (MB) 60 59 68 solved 94% 67% 6% 0% ub lb 0.013 0.111 7.834 time (s) 227 295 419 0 mem. (MB) 61 62 74 0 solved 83% 44% 0% 0% ub lb 0.005 0.565 8.399 time (s) 230 358 0 mem. (MB) 61 62 0 G odel connectives, the 3-valued stable models and the class of fuzzy answer sets coincide. Moreover, Eiter, Leone, and Sacc a (1997) note that 3-valued stable models leave more atoms undefined than necessary. Thus, they characterized 3-valued stable models in terms of Partial stable (P-stable) models and introduced the subclass of Least undefinedstable (L-stable) models. Intuitively, L-stable model semantics selects those 3-valued stable models, where a smallest set of atoms is undefined. Thus, MUFAS, restricted to the G odel connectives, coincide to the L-stable models on FASP programs interpreted over L3. Finally, among other paracoherent answer set semantics, we consider Semi-stable models (Sakama and Inoue 1995) and Semi-equilibrium models (Amendola et al. 2016). These paracoherent semantics satisfy three desiderata (see (Amendola et al. 2016)): (i) every consistent answer set of a program corresponds to a paracoherent model (answer set coverage); (ii) if a program has some (consistent) answer set, then its paracoherent models correspond to answer sets (congruence); (iii) if a program has a classical model, then it has a paracoherent model (classical coherence). In general, FAS semantics (and, thus, MUFAS semantics) does not satisfy this last property. For example, the program Π = {0 a b c, a b, b c, c a} has no fuzzy answer set, while it has some classical model (for instance, setting a, b, and c to true). Measure of undefinedness are often associated with the notion of entropy in information (Kapur and Kesavan 1992; Kullback 1959; Jayne 1957; Bhandari and Pal 1993; Li 2015; Li and Liu 2007; Wang, Dong, and Yan 2012), and applied in several areas: machine learning and decision trees (Hu et al. 2010; Vagin and Fomina 2011; Wang 2011; Wang and Dong 2009; Wang, Zhai, and Lu 2008; Yi, Lu, and Liu 2011; Zhai 2011); portfolio selection and optimization models (Qin, Li, and Ji 2009; Haber, del Toro, and Gajate 2010; Xie et al. 2010); clustering, image processing and computer vision (De Luca and Termini 1974; Yager 1979; 1980; Xie and Bedrosian 1984; Kosko 1986; Pal and Pal 1992; Shang and Jiang 1997; Szmidt and Kacprzyk 2001; Parkash, Sharma, and Mahajan 2008). Conclusions We have studied the notion of minimal undefinedness for fuzzy answer set programming as a means to identify solutions that satisfy additional desired properties. Intuitively, we are interested in solutions that are as close to being classical as possible. This study is motivated by previous work on paraconsistent and paracoherent logical formalisms, but also by an attempt to enhance abduction processes in fuzzy circuits. More precisely, minimally undefined fuzzy answer sets yield the most precise, and easiest to understand explanations for an observed output. Minimally undefined fuzzy answer sets, along with the measures of undefinedness used to define them, satisfy many properties that have been considered in the literature. In particular, they satisfy justifiability, the closed world assumption, and are coherent. Moreover, the distance function UD is a strict measure of undefinedness. We implemented and evaluated four different methods for computing MUFAS based on the distance function, by ex- tending the FASP2SMT system. Our evaluation shows that binary search provides the best strategy in this setting, while the internal (experimental) minimize instruction of Z3 yields the worst results. As future work we intend to extend the capabilities of our prototype to handle other measures of undefinedness and apply it to more realistic instances for abduction in fuzzy circuits. Moreover, the implementation can be improved by employing approximation operators (Alviano and Pe naloza 2013) to properly initialize lower bounds to values greater than ε. For example, for Π = {a 0.1 b, b 0.8 a} and I FAS(L ,Π), I(a) [0.1,0.2] and I(b) [0.8,0.9] hold. Hence, lb in Algorithms 1 3 can be safely initialized to 0.2 ε. Alviano, M., and Dodaro, C. 2016. Anytime answer set optimization via unsatisfiable core shrinking. TPLP 16(56):533 551. Alviano, M., and Pe naloza, R. 2013. Fuzzy answer sets approximations. TPLP 13(4-5):753 767. Alviano, M., and Pe naloza, R. 2015. Fuzzy answer set computation via satisfiability modulo theories. TPLP 15(45):588 603. Alviano, M.; Calimeri, F.; Charwat, G.; and et al. 2013. The fourth answer set programming competition: Preliminary report. In Cabalar, P., and Son, T. C., eds., LPNMR 2013, volume 8148 of LNCS, 42 53. Alviano, M.; Dodaro, C.; and Ricca, F. 2014. Anytime computation of cautious consequences in answer set programming. TPLP 14(4-5):755 770. Amendola, G.; Eiter, T.; Fink, M.; Leone, N.; and Moura, J. 2016. Semi-equilibrium models for paracoherent answer set programs. Artif. Intell. 234:219 271. Barrett, C.; Stump, A.; and Tinelli, C. 2010. The SMT-LIB Standard: Version 2.0. In Gupta, A., and Kroening, D., eds., SMT 10. Bhandari, D., and Pal, N. R. 1993. Some new information measures for fuzzy sets. Inf. Sci. 67(3):209 228. Blondeel, M.; Schockaert, S.; Vermeir, D.; and Cock, M. D. 2013. Fuzzy answer set programming: An introduction. In Soft Computing: State of the Art Theory and Novel Applications. 209 222. Cintula, P.; H ajek, P.; and Noguera, C., eds. 2011. Handbook of Mathematical Fuzzy Logic, volume 37 38 of Studies in Logic. College Publications. De Luca, A., and Termini, S. 1972. A definition of a nonprobabilistic entropy in the setting of fuzzy sets theory. Information and Control 20(4):301 312. De Luca, A., and Termini, S. 1974. Entropy of L-fuzzy sets. Information and Control 24(1):55 73. de Moura, L. M., and Bjørner, N. 2008. Z3: an efficient SMT solver. In TACAS 2008, 337 340. Eiter, T.; Leone, N.; and Sacc a, D. 1997. On the partial semantics for disjunctive deductive databases. Ann. Math. Artif. Intell. 19(1-2):59 96. Gebser, M.; Kaminski, R.; K onig, A.; and Schaub, T. 2011. Advances in gringo series 3. In LPNMR 2011, 345 351. Gelfond, M., and Lifschitz, V. 1991. Classical negation in logic programs and disjunctive databases. New Generation Comput. 9(3/4):365 386. Haber, R. E.; del Toro, R. M.; and Gajate, A. 2010. Optimal fuzzy control system using the cross-entropy method. A case study of a drilling process. Inf. Sci. 180(14):2777 2792. Hu, Q.; Pan, W.; An, S.; Ma, P.; and Wei, J. 2010. An efficient gene selection technique for cancer recognition based on neighborhood mutual information. Int. J. Machine Learning & Cybernetics 1(1-4):63 74. Janssen, J.; Schockaert, S.; Vermeir, D.; and Cock, M. D. 2012a. Answer Set Programming for Continuous Domains - A Fuzzy Logic Approach, volume 5 of Atlantis Computational Intelligence Systems. Atlantis Press. Janssen, J.; Vermeir, D.; Schockaert, S.; and Cock, M. D. 2012b. Reducing fuzzy answer set programming to model finding in fuzzy logics. TPLP 12(6):811 842. Jayne, E. 1957. Information theory and statistical mechanics. Physical Reviews 106(4):620 630. Kapur, J. N., and Kesavan, H. K. 1992. Entropy optimization principles and their applications. In Entropy and Energy Dissipation in Water Resources. Springer Netherlands. 3 20. Kosko, B. 1986. Fuzzy entropy and conditioning. Inf. Sci. 40(2):165 174. Kullback, S. 1959. Information Theory and Statistics. New York: Wiley. Lee, J., and Wang, Y. 2014. Stable models of fuzzy propositional formulas. In JELIA 2014, 326 339. Li, X., and Liu, B. 2007. Maximum entropy principle for fuzzy variables. Int. J. of Uncertainty, Fuzziness and Knowledge-Based Systems 15(Supplement-2):43 52. Li, X. 2015. Fuzzy cross-entropy. J. of Uncertainty Analysis and Applications 3(1):1 6. Lukasiewicz, T. 2006. Fuzzy description logic programs under the answer set semantics for the semantic web. In Rule ML 2006, 89 96. IEEE Computer Society. Marek, V. W., and Truszczy nski, M. 1999. Stable Models and an Alternative Logic Programming Paradigm. In The Logic Programming Paradigm A 25-Year Perspective. Springer Verlag. 375 398. Mushthofa, M.; Schockaert, S.; and Cock, M. D. 2014. A finite-valued solver for disjunctive fuzzy answer set programs. In ECAI 2014, 645 650. Mushthofa, M.; Schockaert, S.; and Cock, M. D. 2015. Solving disjunctive fuzzy answer set programs. In LPNMR 2015, 453 466. Niemel a, I. 1999. Logic programs with stable model semantics as a constraint programming paradigm. Ann. Math. Artif. Intell. 25(3-4):241 273. Nieuwenborgh, D. V.; Cock, M. D.; and Vermeir, D. 2007. An introduction to fuzzy answer set programming. Ann. Math. Artif. Intell. 50(3-4):363 388. Pal, N., and Pal, S. 1992. Higher order fuzzy entropy and hybrid entropy of a set. Inf. Sci. 61(3):211 231. Parkash, O.; Sharma, P.; and Mahajan, R. 2008. New measures of weighted fuzzy entropy and their applications for the study of maximum weighted fuzzy entropy principle. Inf. Sci. 178(11):2389 2395. Przymusinski, T. C. 1991. Stable semantics for disjunctive programs. New Generation Comput. 9(3/4):401 424. Qin, Z.; Li, X.; and Ji, X. 2009. Portfolio selection based on fuzzy cross-entropy. J. Comput. App. Math. 228(1):139 149. Sakama, C., and Inoue, K. 1995. Paraconsistent stable semantics for extended disjunctive programs. J. Log. Comput. 5(3):265 285. Shang, X., and Jiang, W. 1997. A note on fuzzy information measures. Pattern Recognition Letters 18(5):425 432. Szmidt, E., and Kacprzyk, J. 2001. Entropy for intuitionistic fuzzy sets. Fuzzy Sets and Systems 118(3):467 477. Vagin, V. N., and Fomina, M. V. 2011. Problem of knowledge discovery in noisy databases. Int. J. Machine Learning & Cybernetics 2(3):135 145. Wang, X., and Dong, C. 2009. Improving generalization of fuzzy IF-THEN rules by maximizing fuzzy entropy. IEEE Trans. Fuzzy Systems 17(3):556 567. Wang, X.; Dong, L.; and Yan, J. 2012. Maximum ambiguitybased sample selection in fuzzy decision tree induction. IEEE Trans. Knowl. Data Eng. 24(8):1491 1505. Wang, X.; Zhai, J.; and Lu, S. 2008. Induction of multiple fuzzy decision trees based on rough set technique. Inf. Sci. 178(16):3188 3202. Wang, L. 2011. An improved multiple fuzzy NNC system based on mutual information and fuzzy integral. Int. J. Machine Learning & Cybernetics 2(1):25 36. Xie, W., and Bedrosian, S. 1984. An information measure for fuzzy sets. IEEE Transactions on systems, man, and cybernetics 14(1):151 156. Xie, H.; Zheng, Y.; Guo, J.; and Chen, X. 2010. Cross-fuzzy entropy: A new method to test pattern synchrony of bivariate time series. Inf. Sci. 180(9):1715 1724. Yager, R. R. 1979. On the measure of fuzziness and negation part i: Membership in the unit interval. Int. J. of General Systems 5(4):221 229. Yager, R. R. 1980. On the measure of fuzziness and negation. II. Lattices. Information and Control 44(3):236 260. Yi, W.; Lu, M.; and Liu, Z. 2011. Multi-valued attribute and multi-labeled data decision tree algorithm. Int. J. Machine Learning & Cybernetics 2(2):67 74. You, J., and Yuan, L. 1994. A three-valued semantics for deductive databases and logic programs. J. Comput. Syst. Sci. 49(2):334 361. Zhai, J. 2011. Fuzzy decision tree based on fuzzy-rough technique. Soft Comput. 15(6):1087 1096.