Towards Smooth Monotonicity in Fuzzy Inference System Based On Gradual Generalized Modus Ponens

Download as pdf or txt
Download as pdf or txt
You are on page 1of 8

8th Conference of the European Society for Fuzzy Logic and Technology (EUSFLAT 2013)

Towards Smooth Monotonicity


in Fuzzy Inference System based on
Gradual Generalized Modus Ponens
Phuc-Nguyen Vo1 Marcin Detyniecki1
1
Laboratoire d’Informatique de Paris VI
Université Pierre et Marie Curie
4, place Jussieu, 75005 Paris, France
Email: {phuc-nguyen.vo, marcin.detyniecki}@lip6.fr

Abstract hicle to avoid obstacles [2], there is a monotonic re-


lationship between the distance (from the vehicle to
Monotonic relationship between input and output the obstacle) and the turning degree of the vehicle:
often is an inherent property of fuzzy systems. For “The nearer the obstacle is, the more the vehicle
example, for a cart-pole system, there is a gradual should be turned.”
relationship between the pole angle and the cart In recent years, several research works have fo-
velocity of the form: “The more the pole deviates cused on designing monotonic fuzzy inference sys-
from the upright position, the higher velocity the tems (FIS). In [3] and [4], K. M. Tay et al. propose
cart must gain”. an approach to build a fuzzy inference system that
Thus, we propose an approach, that we call Grad- preserves the monotonicity property. The authors
ual GMP, which is able to integrate the graduality, introduce a fuzzy re-labeling technique to re-order
when this underlying hypothesis exists, into the gen- the consequences of fuzzy rules in the database and
eralized modus ponens. a monotonicity index. This approach is able to
In this paper, we proposed to study in a simple overcome several restrictions; in particular it uses a
case (crisp observations) the monotonic response similarity-based reasoning scheme to design mono-
of what could be the adaptation to this case of tonic multi-input FIS models. There are also some
GGMP-based FIS. We showed that for a single rule studies ([5], [6]) on the monotonicity of single in-
FIS, the GGMP induces strict monotonocity, to be put rule modules (SIRMs) connected fuzzy inference
compared with static response in the standard case method where the output of the rule is generally
(here Mamdani controller). Further experimental simple, i.e., a precise value. The authors study the
results, on single and double input FIS, show that properties of the inference method and especially
the proposed method implies a smooth monotone the conditions for a monotonicity of the inference
behavior even when in presence of several linguistic results.
values for each variable. In the case of controllers, especially Mamdani-like
Finally, this work reveals that the only way to guar- models [7], Broekhoven et al. ([8], [9]) show that,
antee strict monotonicity is to have an inference ordered linguistic values for all input variables and
that has a gradual behavior. for all output variables, plus a set of rules describing
a monotone system are not enough to guarantee a
Keywords: Fuzzy Inference System, Fuzzy Con- monotone input-output behavior. They state that
trol, Generalized Modus Ponens, Monotonicity the choice of the mathematical operators used when
calculating the model output and the properties of
1. Introduction the membership functions are also of crucial impor-
tance. These constraints have been observed as well
In most of the applications of fuzzy control, mono- as in another work by M. Štěpnička et al. [10].
tonicity is an inherent property of the system out- All these works focused on the post factum analy-
put (control action) with respect to the input. For sis of the inference system and on the interaction of
example, consider a simple cart-pole system [1] in the rules. In this paper, we propose to modify the
which the pole angle and its angular velocity can be underlying inference mechanism by using, what we
set as two state variables and the cart velocity as the called, the Gradual GMP that allows to integrate
system output. An appropriate controller for the the graduality - in a sense of monotonicity - in a
cart-pole system must have the monotonicity prop- GMP framework. Based on the GGMP, described
erty between the given pole angle (as well as the an- in next section, an inference system may be built,
gular velocity) and the desired cart velocity: “The as described in section 3. Moreover, we present a
more the pole deviates from the upright position, set of experiments showing how, the introduction
the higher velocity the cart must gain.” In the same of the GGMP in the inference step, can lead to a
way, in case of controlling an automated guided ve- smooth monotone behavior, compared to a partially

© 2013. The authors - Published by Atlantis Press 788


static classical fuzzy inference system (here Mam-
dani controller).

2. Gradual Generalized Modus Ponens

Generalized modus ponens (GMP) is a key inferring


mechanism in inference systems. As defined in [11],
for X and Y , two variables in the universes of dis-
course U and V , A and A0 two fuzzy subsets of U ,
and B a fuzzy subset of V . The most general form
of GMP is given by:
If X is A then Y is B
X is A0
Y is B 0
where the membership function of B 0 is defined by,
for each y ∈ V :

µB 0 (y) = sup T (µA0 (x), I(µA (x), µB (y))) (1)


x∈U

with T is a triangular norm, and I is a fuzzy impli-


cation operator. Figure 1 shows some examples of
inferences obtained using GMP.
To integrate the graduality into GMP, the ba-
sic idea of our approach is to separate the uni-
verse of the premise A (respectively the consequence
B) into three parts, named Smaller, Greater and
Indistinguishable. By doing so, the model fo-
cuses its action in the corresponding parts of A
and B. In other words, only the Smaller (re-
spectively Greater, Indistinguishable) part of A
is used to infer the Smaller (respectively Greater,
Indistinguishable) part of B.

Membership requirements
In this study, the membership functions are as- Figure 1: Examples of inferences using the Gener-
sumed to be convex, normalized and continuous, alized Modus Ponens (t-norm Lukasiewicz, impli-
and their supports are required to be bounded. cation Reichenbach). We observe that if there is
Moreover, the convexity notion is restricted to strict gradual relationship between premise and conclu-
convexity as discussed in [10], i.e., the fuzzy sets sion (what we call the gradual hypothesis) then it
have no partially constant membership functions. is not modeled by the GMP: we obtain the same
In this sense, fully continuous fuzzy numbers can conclusion for observations clearly smaller or larger
be used with the proposed method. Examples of than the premise. If the gradual hypothesis is han-
viable membership functions are shown in Figure 2. dled we expect to obtain an inferred conclusion
smaller (or larger) than the rule’s conclusion.

Universe partitioning
We define membership functions of three parts
Smaller, Greater and Indistinguishable of fuzzy
set A, using the kernel and complement of A (Fig-
ure 3).
Let us refer to the kernel of a fuzzy set A by
[AL , AR ], i.e., the interval so that ∀x ∈ [AL , AR ]
µA (x) = 1.
The membership function of the Smaller part of
A is defined as follows: Figure 2: Examples of strictly convex, normalized
 and continuous membership functions
µA(x) = 1 − µA (x) x < AL
ϕSmallerA (x) =
0 otherwise
(2)

789
The membership function of the Greater part of
A is:

µA(x) = 1 − µA (x) AR < x
ϕGreaterA (x) =
0 otherwise
(3)
The membership function of the
Indistinguishable part of A is:

1 x ∈ [AL , AR ]
ϕIndistinguishableA (x) = (4)
0 otherwise

Figure 3: Partitioning of the universe of discourse


based on the complement and kernel of the fuzzy
set

Formal definition of the GGMP


In order to induce a gradual 1 behavior in the GMP,
we propose to compute conclusion’s membership
function B 0 as follows, for each y ∈ V :

µB 0 (y) = sup T (µA0 (x), I(µA (x), µB (y))) (5)


x∈ψ(y)

where
ψ(y) = {x ∈ U | Figure 4: Gradual Generalized Modus Ponens im-
ϕPA (x) = ϕPB (y) and ϕPB (y) > 0, proves the specificity of the inferred conclusion by
P ∈ {Smaller, Greater, taking advantage of the gradual relationship (t-
Indistinguishable}} norm product, implication Brouwer-Godel)

Figure 4 shows the conclusion B 0 inferred using


GGMP, that should be compared with what would
• If µA0 (x) = 0
have been obtained using normal GMP, as shown
then T (µA0 (x), I(µA (x), µB (y))) = 0
on Figure 1.
• If µA0 (x) = 1
then T (µA0 (x), I(µA (x), µB (y))) = 1
Crisp observation
In practice, input values of fuzzy inference systems
That means, the conclusion of GGMP is crisp. 
often are crisp (without any fuzziness). The follow-
ing theorem shows that when the observation is a
crisp value, the proposed inference method, GGMP,
infers a crisp conclusion. Compatible t-norms and implications
Theorem 1: If the observation is a crisp value,
the conclusion of GGMP is a crisp value. The compatibility [13] between a t-norm and impli-
Proof: cation for the generalized modus ponens translates
For each y, from the definition of ψ(y) in formula the requirement that if the observation is identical
(5), I(µA (x), µB (y)) = 1 for all x ∈ ψ(y). with the premise (A0 ≡ A), the inferred conclusion
Because A0 is crisp, for any x ∈ U , µA0 (x) takes should also be identical with the consequence of the
one of two values 0 or 1: rule (B 0 ≡ B). Table 1 lists some implications and
1 This is not the “gradual” meant by some authors, who t-norms which are compatible for the Gradual GMP.
focus on the graduality of the truth-values as in [12]

790
T-norm Definition U which is the chosen point to infer µB 0 (y0 ), i.e.,
Lukasiewicz TL (u, v) = max(0, u + v − 1) x0 = arg max T (µA0 (x), I(µA (x), µB (y0 )))
Minimum TM (u, v) = min(u, v) x∈ψ(y0 )

Product TP (u, v) = uv First, we will show that inf{Bα } ≤ inf{Bα0 }.

Implication T-norm i) If x0 is in the Indistinguishable or Greater


Rescher-Gaines part of A then y0 is also in the corresponding
( Indistinguishable or Greater part of B, re-
1 x≤y TL , TM , TP spectively.
IRG (x, y) =
0 x>y Thus we have inf{Bα } ≤ y0 = inf{Bα0 }
Brouwer-Godel (since inf{Bα } belongs to the Smaller part of
( B).
1 x≤y TL , TM , TP
IBG (x, y) = ii) If x0 is in the Smaller part of A then y0 is in
y x>y the Smaller part of B.
Goguen In case of using implications IRG , IBG , IG , IL ,
( we have µA0 (x0 ) = µB 0 (y0 ) = α
1 x≤y TL , TM , TP
IG (x, y) = then inf{Aα } ≤ x0
y/x x > y (from the hypothesis that inf{Aα } ≤ inf{A0α }).
Lukasiewicz Therefore, µB (y0 ) = µA (x0 ) ≥ α
TL , TM , TP (by non-decreasing property of convex fuzzy set
IL (x, y) = min(1 − x + y, 1)
in the Smaller part),
Table 1: T-norms and Implications compatible for this implies inf{Bα } ≤ y0 = inf{Bα0 }.
the Gradual GMP In case of using implication IM with t-norm
minimum, from the definition of x0 ,
µB 0 (y0 ) = min(µA0 (x0 ), µA (x0 )) = α.
Preserving the ordering by Gradual GMP Here, we have µA0 (x0 ) = α and µA (x0 ) ≥ α
Ordering is a notion which is closely related to (otherwise, if µA (x0 ) = α and µA0 (x0 ) = α1 >
monotonicity. Preserving the ordering of the con- α,
clusion with regard to the observation is a prop- then inf{Aα1 } > x0 , this is inconsistent with
erty of GGMP. This property plays a key role when the hypothesis that inf{Aα1 } ≤ inf{A0 α1 }).
inducing smooth monotonicity of fuzzy systems, Therefore µB (y0 ) = µA (x0 ) ≥ α,
which are discussed in the sections below. it implies inf{Bα } ≤ inf{Bα0 }.
For the purpose of this paper, we focus, here, on Similarly, we have sup{Bα } ≤ sup{Bα0 }, i.e.,
the fuzzy max order introduced by J. Ramík and J. B m B 0 .
Římánek in [14]: In the same way, we could prove that:
Definition: (fuzzy max order) Fuzzy set A is said
to be below or equal to fuzzy set B according to the If A0 m A then B 0 m B
fuzzy max order (A m B) if ∀α ∈ (0, 1]:

inf{Aα } ≤ inf{Bα } and sup{Aα } ≤ sup{Bα } (6)
Order preservation of the GGMP is even stronger
in which Aα is the α-cut of A defined by than what is presented by this property, which fo-
cuses on comparing observation with the premise
Aα = {x ∈ U |µA (x) ≥ α} and obtained conclusion with rule’s consequence. In
fact, for GGMP, if one observation is smaller than
Theorem 2: Let A and B be the antecedent and
another observation, then the corresponding conclu-
consequence of a fuzzy rule; A0 and B 0 be the ob-
sions also fulfill this order, i.e., if A01 is smaller than
servation and conclusion of GGMP.
A02 then B10 is smaller than B20 (Figure 5). From the
If A m A0 then B m B 0 definition of GGMP and above theorem, it can be
deducted the following corollary about the ordering
Proof: of the conclusions according to the ordering of the
From the conditions of the membership functions observations.
mentioned above, there exist both the minimal and Corollary: Let A and B be the antecedent and
maximal elements of the α-cut of a fuzzy set for all consequence of a fuzzy rule; B10 and B20 respectively
α ∈ (0, 1], i.e., be the conclusions according to the observations A01
and A02 using GGMP.
inf{Aα } = min{Aα } and sup{Aα } = max{Aα }
If A01 m A02 then B10 m B20 .
If A m A0 , for each α ∈ (0, 1], we determine
y0 ∈ V such that y0 = inf{Bα0 }. There exists x0 ∈

791
3.1. Towards GGMP-based FIS
Several difficulties have to be solved when replac-
ing, in a classical FIS, the GMP by the introduced
gradual version (equation 5). Most of these chal-
lenging questions are due to the high interaction of
the different rules of the FIS. Thus, one of the adap-
tations we propose is to restrict the rules to be fired,
to the ones concerned by the observation. In other
words, if the observation is out of the support of the
premise, the inference mechanism is not applied to
deduce the conclusion. This is obtained with the fol-
lowing formulation: For the inference step, we pro-
pose that for given input vector X = (x1 , x2 , ..., xn ),
for rule Ri , the result for each input value xj , called
0
Bi j , should be calculated separately by:
µB 0 j (y) = sup T (µxj (x), I(µAj (x), µBi (y)))
i i
Figure 5: GGMP preserves the ordering of conclu- x∈ψij (y)

sions (B10 m B20 ) with regard to ordered observa- (8)


tions (A01 m A02 ) where
ψij (y) = {x ∈ supp(Aji )|
ϕP j (x) = ϕPBi (y) and ϕPBi (y) > 0,
3. Studying Fuzzy Inference Systems based A
i

on GGMP P ∈ {Smaller, Greater,


Indistinguishable}}
Before introducing a GGMP-based FIS, we recall Another set of challenges, this time related to
some fundamentals, while introducing notations, of the augmented precision induced by the gradual hy-
Inference Models based on GMP. pothesis has to be solved in order to be able to have
Let us consider a system with a set of m fuzzy a fully functional system (for any observed fuzzy
rules, each of them (Ri ) of the form: set). In particular, this is the question of how to
aggregate and defuzzify the conclusions.
If X1 is A1i1,s AND X2 is A2i2,s ... Xn is Anin,s Before attempting to propose a general model, in
then Y is Bis this paper, we propose to study the induced gradual
behavior, by comparing the FIS in the case of crisp
where (X1 , X2 , ..., Xn ) are n input variables and observation. Notice that this is very often in the
(A1i1,s , A2i2,s , ..., Anin,s ) linguistic values of the an- case of controllers, their observation are often pre-
tecedents in universes of discourse (U1 , U2 , ..., Un ); cise physical readings. In this context we propose
Bis (with i ∈ {1, ..., m}) are linguistic values of the the following formula to obtain the aggregated and
consequences in universe V . defuzzified conclusion y0 :
The first stage of FIS is to infer the result of each Pm Pn j j j
j=1 µBi (yi )µB j (yi )yi
0
rule. For rule Ri , the degree of fitness hi of input i=1
i
y0 = Pm Pn j j
(9)
vector X = (x1 , x2 , ..., xn ) to the antecedent parts
j=1 µBi (yi )µB j (yi )
0
i=1
(A1i1,s , A2i2,s , ..., Anin,s ) is given by: i

with m being the number of rules and n the number


0
of input variables; yij denotes the point at which Bi j
O O O
hi = µA1i (x1 ) µA2i (x2 ) ... µAni (xn )
n,s
1,s 2,s
has positive membership degree.
(7)
N This formulation is analogous to a conjunctive ag-
where stands for a conjunction operator (such as
gregation and center of gravity defuzzification.
minimum or product operator). Then the inference
result Bi0 is deduced by using GMP.
Experimental Results and Discussion
Next, the overall consequence B 0 is aggregated
from B10 , B20 , ..., Bm
0
. Depending on the type of rules In the following subsections, we study the proposed
a t-norm (conjunction operator) or a t-conorm (dis- method based on some experiments. In particular,
junction operator) is used to aggregate B 0 . the behavior of the GGMP-based FIS will be com-
pared with, what can be considered as a reference
The final stage is a defuzzification which con- in terms of monotonic behavior ([8], [9] [10]), the
cludes the control action. There are many meth- Mamdani’s inference system with Center-of-Gravity
ods to obtain the representative point y0 for the re- method as defuzzification.
sulting fuzzy set B 0 , such as the Center-of-Gravity, The tests consist in, for each input universe, in-
Mean-of-Maxima and Center-of-Sums. crease the precise observations gradually (notice

792
that there may be several inputs). The output of explained by the fact that there is no gradual be-
the system is plotted and compared with similar havior in areas, of the input universe, where only
plot obtained using the reference system (Mamdani one rule applies. And since Mamdani inference (as
controller). In other words, for a system with one well as others) does not integrate a gradual behavior
input variable, we compare two curves, while for one (when this hypothesis can be retained), we observe
with two input variables, we compare two surfaces. plateaus in the response.
The linguistic values of the antecedents as well
as of the consequences of the rules are ordered as
illustrated in Figure 2, i.e., for universe Ui :

Ai1 m Ai2 m ... m Ain

Figure 7: Effect of using GGMP, compared to Mam-


dani’s method, in the case of an inference system
with one input variable

3.4. Double input FIS


Figure 6: Smooth strict monotonic behavior in-
duced by the GGMP, for a single rule, to be com- In order to study the influence that may have an
pared with the constant response of Mamdani’s input over another one, we set up several tests with
method FIS containing double inputs with several linguistic
values each. Figure 8 shows the results of systems
having two inputs and a rule base consisting of three
3.2. Single input FIS with one rule rules of the form:
The simplest possible FIS is a single input FIS with
If X1 is A1i AND X2 is A2i then Y is Bi
just one rule. There is no aggregation, it is just the
inference and the defuzzification. Figure 6, com- With Mamdani’s method, there is a plateau on the
pares the behavior of a Mamdani FIS versus that surface of system output. Despite the increasing
of a GGMP-base FIS. We observe that the latter of input values, the output of Mamdani’s method
has a smooth strict monotonic behavior. The con- does not change while that of the proposed method
sistently increasing curve translates the fact that consistently increases.
when the observation value is increased (in the scale For systems with a more complex rule base,
defined in universe U ) the inferred decision strictly as shown on Figure 9 where there are 19 rules
increases (in the scale of universe V ). This is to be corresponding to the combinations of the under-
compared with what is obtained with Mamdani’s lying linguistic values, the proposed method still
method, where the increase of observations value has smoother responses than those of Mamdani’s
has no influence on the conclusion. method.

3.3. Single input FIS with multiple rules 4. Conclusion


In a system with a single input variable with several
linguistic values, the fuzzy rules Ri have the form: Monotonic relationship between input and output
often is an inherent property of fuzzy systems. For
If X is Ai then Y is Bi example, for a cart-pole system, there is a gradual
relationship between the pole angle and the cart
In Figure 7, we observe that GGMP improves the velocity of the form: “The more the pole deviates
performance in terms of smoothness of monotonic from the upright position, the higher velocity the
behavior compared to Mamdani’s inference. In fact, cart must gain”.
when the observation value evolves around the ker- Thus, we propose an approach, that we call Grad-
nel of the premise of a rule there is no great im- ual GMP, which is able to integrate the gradual-
pact in the conclusion. This can be furthermore ity, when this underlying hypothesis exists, into the

793
future works which will concentrate on a general
formulation, able to handle any type of fuzzy obser-
vation, for GGMP-based FIS.

References

[1] T. Yamakawa. Stabilization of an inverted pen-


dulum by a high-speed fuzzy logic controller
hardware system. Fuzzy Sets and Systems,
32(2):161–180, 1989.
[2] P. Lee and L. Wang. Collision avoidance by
Figure 8: Smooth monotonic behavior of the pro- fuzzy logic control for automated guided ve-
posed method in case of system with two inputs and hicle navigation. Journal of Robotic Systems,
the rule base consisting of three fuzzy rules 11(8):743–760, 1994.
[3] K. M. Tay and C. P. Lim. An evolutionary-
based similarity reasoning scheme for mono-
tonic multi-input fuzzy inference systems.
IEEE International Conference on Fuzzy Sys-
tems, pages 442 – 447, 2011.
[4] K. M. Tay, C. P. Lim, and T. L. Jee. Building
monotonicity-preserving fuzzy inference mod-
els with optimization-based similarity reason-
ing and a monotonicity index. IEEE Interna-
tional Conference on Fuzzy Systems, pages 1–8,
2012.
[5] H. Seki and H. Ishii. On the monotonic-
ity of functional type SIRMs connected fuzzy
Figure 9: Smoother response of the fuzzy system reasoning method and T-S reasoning method.
based on GGMP in case of a complex system (2 IEEE International Conference on Fuzzy Sys-
inputs and 19 fuzzy rules) tems, pages 58–63, 2008.
[6] H. Seki and M. Mizumoto. SIRMs connected
fuzzy inference method adopting emphasis and
generalized modus ponens. We show that the mech- suppression. Fuzzy Sets and System, pages
anism preserves the ordering property of the con- 112–126, 2012.
clusion with respect to the observation, a property [7] E. H. Mamdani. Application of fuzzy algo-
closely related to monotonicity. rithms for control of simple dynamic plant. In
In this paper, we studied in a simple case (crisp ob- Proceedings of the Institution of Electrical En-
servations) the monotonic response of what could be gineers, volume 121, 1974.
the adaptation to this case of GGMP-based FIS. We [8] E. V. Broekhoven and B. D. Baets. Monotone
showed that, for a single rule FIS, GGMP induces Mamdani-Assilian models under mean of max-
strict monotonocity, to be compared with static be- ima defuzzification. Fuzzy Sets and System,
havior in the reference case (Mamdani). Further ex- 159(21):2819–2844, 2008.
perimental results, on single and double input FIS, [9] E. V. Broekhoven and B. D. Baets. Only
show that the proposed method implies a smooth smooth rule bases can generate monotone
monotone behavior even when in presence of sev- Mamdani-Assilian models under center-of-
eral linguistic values for each variable. gravity defuzzification. IEEE Transactions on
They also reveal that, for all classical FIS, mono- Fuzzy Systems, 17(5), 2009.
tonicity is due to rule overlap, and therefore when [10] M. Štěpnička and B. D. Baets. Monotonicity of
observations are handled by only one rule (even implicative fuzzy models. IEEE International
when inside a FIS), the monotone behavior can not Conference on Fuzzy Systems, pages 1–7, 2010.
be insured. The only way to guarantee strict mono- [11] L. A. Zadeh. The concept of a linguistic vari-
tonicity is to have an inference that has a gradual able and its application to approximate rea-
behavior. soning III. Information Sciences, 9(1):43–80,
In systems with multiple input variables, the anal- 1975.
ysis becomes more complex because of the inter- [12] D. Dubois and H. Prade. Gradual inference
action of system parameters, as has been studied rules in approximate reasoning. Information
by other authors. Nevertheless, our experiments Sciences, 61(1-2):103–122, 1992.
point out that introducing graduality at inference [13] B. Bouchon-Meunier. Stability of linguistic
level smoothes the monotone response of the sys- modifiers compatible with a fuzzy logic. In
tem. This encouraging result opens up the way for 2nd International Conference on Information

794
Processing and Management of Uncertainty in
Knowledge-Based Systems, pages 63–70, 1988.
[14] J. Ramík and J. Římánek. Inequality relation
between fuzzy numbers and its use in fuzzy op-
timization. Fuzzy Sets and System, 16(2):123–
138, 1985.
[15] T. Takagi and M. Sugeno. Fuzzy identification
of systems and its applications to modeling and
control. IEEE Transactions on Systems, Man
and Cybernetics, 15(1):116–132, 1985.
[16] N. Yubazaki, J. Yi, and K. Hirota. SIRMs (Sin-
gle Input Rule Modules) connected fuzzy in-
ference model. Journal of Advanced Computa-
tional Intelligence and Intelligent Informatics,
1(1):23–30, 1997.
[17] M. Kurano, M. Yasuda, J. Nakagami, and
Y. Yoshida. Ordering of convex fuzzy sets -
A brief survey and new results. Journal of the
Operations Research Society of Japan, 43(1),
2000.
[18] J. M. Won, S. Y. Park, and Jin S. Lee. Param-
eter conditions for monotonic Takagi-Sugeno-
Kang fuzzy system. Fuzzy Sets and System,
132(2):135–146, 2002.

795

You might also like