universal approximations of continuous fuzzy-valued functions by multi-layer regular fuzzy neural...
TRANSCRIPT
Fuzzy Sets and Systems 119 (2001) 313–320www.elsevier.com/locate/fss
Universal approximations of continuous fuzzy-valued functionsbymulti-layer regular fuzzy neural networks(
Puyin LiuDepartment of System Engineering and Mathematics, National University of Defence Technology, Changsha, Hunan 410073,
People’s Republic of China
Received 23 July 1998; received in revised form 10 March 1999; accepted 1 July 1999
Abstract
The fact that four-layer feedforward regular fuzzy neural networks with sigmoid function in the �rst hidden layer arecapable of approximately representing continuous fuzzy valued functions on any compact set of R is shown. At �rst, Bernsteinpolynomials associated with fuzzy valued functions are employed to approximate continuous fuzzy valued function de�ned ona compact set. Secondly, by the conclusions related to standard feedforward networks, universal approximations of continuousfuzzy valued functions by regular fuzzy neural networks are obtained. c© 2001 Elsevier Science B.V. All rights reserved.
Keywords: Regular fuzzy neural networks; Fuzzy-valued polynomials; Universal approximations; Universal approximator
1. Introduction
It is shown that a three-layer feedforward neu-ral network with nonlinear activation function in thehidden layer is capable of approximating generic classof functions, including continuous and integrable ones[13]. These achievements have not only solved theapproximation representation of some multi-variablefunctions by the combination of �nite compositions ofone-variable functions, but are also found to be usefulin many real �elds, such as approximation of struc-tural synthesis [11], system identi�cation [3], patternclassi�cation [4], and adaptive �ltering [7], etc.Recently, some authors have begun to pay atten-
tions to the similar approximation problems in fuzzy
( This work was supported by a grant from National NaturalScience Foundation.
E-mail address: [email protected] (P. Liu).
environments [1,2,9,10]. In [1], Buckley shows thatregular fuzzy neural networks (RFNNs) are notcapable of approximating continuous fuzzy functionsto any degree of accuracy on the compact sets ofF0(R). Based on these conclusions, the authors in[9,10] show some fuzzy functions that can arbitrarilybe closely approximated by RFNNs. Using �-cuts andinterval arithmetic, Buckley and Hayashi [2] demon-strate the fact that neural networks with non-negativeor non-positive fuzzy number inputs are universalapproximators for a class of fuzzy functions ob-tained by the extension principle. Further problems,such as, what are the conditions for fuzzy functionsthat can be approximated to any degree of accu-racy by RFNNs? How can we study approximationcapabilities of RFNNs? etc., are very important andmeaningful mathematical problems in theoretical andapplied areas.
0165-0114/01/$ - see front matter c© 2001 Elsevier Science B.V. All rights reserved.PII: S 0165 -0114(99)00132 -3
314 P. Liu / Fuzzy Sets and Systems 119 (2001) 313–320
In this paper, we obtain the universal approximationof continuous fuzzy-valued function F :R→F0(R)by the RFNNs with real input and fuzzy valued output.In addition, we give some equivalent conditions forthe fact that fuzzy-valued functions can abitrarily beclosely approximated by a class of RFNNs. Finally,some important problems related to the subject forfuture research are put forward.
2. Preliminary
Let F(R) be the set of all fuzzy sets on the realnumber set R; N the set of all natural numbers, andF0(R) the set of all fuzzy numbers, i.e. the followingconditions hold for A∈F0(R):(i) ∀�∈ (0; 1]; A�,{x∈R | A(x)¿�} is the closed
interval of R;(ii) Supp( A),{x∈R | A(x)¿0}⊂R is the bound-
edly closed subset of R;(iii) A is normal, i.e. {x∈R | A(x)= 1} 6= ∅:For simplicity, let A0; zero-cut of A, denote thesupport Supp(A) of A. Obviously, A0 is a boundedlyclosed interval of R. Write A�• ,{x∈R | A(x)¿�}.De�ne the metric on F0(R) as follows: for
A; B∈F0(R);
D(A; B)=∨
�∈[0;1](dH(A�; B�))=
∨�∈[0;1]
(dH(A�• ; B�• ));
where dH means Hausdor� metric, i.e. if A; B⊂R;
dH(A; B)
= max
∨x∈A
∧y∈B
|x−y|;∨
y∈B
(∧x∈A
|x−y|)
and∨;∧denote, respectively, the supremum oper-
ator sup(max) and in�mum operator inf(min). It iswell known that (F0(R); D) is a completely separablemetric space [6].Assume A∈F0(R); if A�= [a1�; a2�] for each �∈
[0; 1]; henceforth we write
|A|,∨
�∈[0;1](|a1�| ∨ |a2�|):
By the extension principle [12], each functionf :Rn→Rmay be extended to oneF0(R)n→F(R);
which is still denoted by f for simplicity:
∀A1; : : : ; An ∈F0(R);
f(A1; : : : ; An)(y) =∨
f(x1 ;:::; xn)=y
(n∧i=1
{Ai(xi)}): (1)
For A; B∈F0(R); and c∈R; addition A + B; multi-plication A · B and multiplication by a scalar c; c · Aare de�ned by the extension principle. The followingconclusion follows from [12].
Remark 1. Suppose A; B∈F0(R); c∈R; then theequalities
(A+ B)� = A� + B�; (A · B)� = A� · B�;(c · A)�= c · A�
(2)
hold.
3. Universal approximations of fuzzy-valuedfunctions
Let C denote a subset of the set of all fuzzy-valuedfunctions that R→F0(R): C is said to be continu-ous if given arbitrarily as F ∈C; F is continuous onR. C is called the additive class if H1 + H2 ∈C forarbitrary H1; H2 ∈C; where (H1 + H2)(x)=H1(x) +H2(x):
De�nition 1. Suppose fuzzy-valued function F :R→F0(R): For arbitrary compact set U ⊂R and ”¿0;there is H ∈C; such that ∀x∈U; D(F(x); H (x))¡”:Then C is called the universal approximator for F; Cis also called the universal approximation of F:
The following remark is trivial.
Remark 2. Let C1;C2 be subsets of the set of func-tions that R→F0(R); and ∀H ∈C1;C2 is universalapproximator for H . Then, if C1 is a universal approx-imator for F :R→F0(R); so is C2:
Lemma 1. Assume that I is an arbitrary in-dex set; {ai | i∈ I}; {bi | i∈ I}⊂ [0; 1]; and h¿0;∀i∈ I; |ai − bi |6h: Then the following hold:(i) |∨i∈I ai −∨i∈I bi |6h;(ii) |∧i∈I ai −∧i∈I bi |6h:
P. Liu / Fuzzy Sets and Systems 119 (2001) 313–320 315
Proof. (i) By the hypotheses, if �,∨i∈I ai; �,∨
i∈I bi, then �; �∈ [0; 1]: It su�ces to prove |�−�|6h:∀”¿0; there are i0; j0 ∈ I; such that ai06�¡ai0 + ”;bj06�¡bj0 + ”: Consequently,
ai0 − bj0 − ”¡� − �¡ai0 − bj0 + ”: (3)
If i0 = j0; we obtain by the assumptions and (3) that
− h− ”6ai0 − bj0 − ”¡�− �¡ai0 − bj0 + ”6h+ ”;
i.e. |�− �|¡h+ ”: So |�− �|6h: If i0 6= j0; it suf-�ces to show |�− �|6h in the following cases:(i) ai0¿aj0 ; bi0¿bj0 ; (ii) ai0¿aj0 ; bi0¡bj0 ; (iii) ai0¡aj0 ; bi0¿bj0 ; (iv) ai0¡aj0 ; bi0¡bj0 .In case (i), it is obvious that
− h6aj0 − bj06ai0 − bj0¡ai0 − (�− ”)6ai0 − bi0 + ”6h+ ”:
So (3) implies the fact that − h − ”¡� − �¡h + 2”holds, i.e. |� − �|¡h+ 2”⇒|� − �|6h:In case (ii), we may easily prove that
− h6aj0 − bj06ai0 − bj0¡ai0 − bi06h:
Hence, (3) implies that − h − ”¡� − �¡h + ”⇒|� − �|6h:For the same reasons,we have |� − �|6h in cases
(iii) and (iv). Thus (i) holds.(ii) If ∀i∈ I; a′i =1− ai; b′i =1− bi; we have
a′i ; b′i ∈ [0; 1]; and |a′i − b′i |6h (i∈ I): So by (i),
|∨i∈I a′i −∨i∈I b′i |6h holds. Therefore,∣∣∣∣∣∧i∈Iai −
∧i∈Ibi
∣∣∣∣∣=∣∣∣∣∣(1−
∧i∈Iai
)−(1−
∧i∈Ibi
)∣∣∣∣∣=
∣∣∣∣∣∨i∈Ia′i −
∨i∈Ib′i
∣∣∣∣∣6hi.e. (ii) holds.
By the fact that a∨ b= 12(a+ b+ |a− b|) for
a; b∈R; the following lemma is obvious.
Lemma 2. Suppose n∈N; and a1; : : : ; an; b1; : : : ; bn∈R; then (∑n
i=1 ai) ∨ (∑n
i=1 bi)6∑n
i=1(ai ∨ bi):
Lemma 3. Assume A; A1; A2 ∈F0(R); n∈N; and{Wk | k =1; : : : ; n}; {Vk | k =1; : : : ; n}⊂F0(R):Then;(i) D(A · A1; A · A2)6| A| ·D(A1; A2);(ii) D(
∑nk=1 Wk ;
∑nk=1 Vk)6
∑nk=1D(Wk ; Vk):
Proof. ∀�∈ [0; 1]; set
(Ai)�= [a1i�; a2i�] (i=1; 2); A�= [a1�; a
2�];
(Wk)�= [w1k�; w2k�]; (Vk)�= [v1k�; v
2k�]
(k =1; : : : ; n):
By the assumptions, the following
dH((A1)�; (A2)�) = dH([a11�; a21�]; [a
12�; a
22�])
= |a21� − a22�| ∨ |a11� − a12�|
hold. What one has to do is to show (i), (ii),respectively.(i) Since by Remark 1, ∀�∈[0; 1]; (A · A1)�=
A� · (A1)�; so if (A · A1)�= [�1; �2]; we have
�1 = (a1� · a11�)∧ (a2� · a11�)∧ (a1� · a21�)∧ (a2� · a21�);�2 = (a1� · a11�)∨ (a2� · a11�)∨ (a1� · a21�)∨ (a2� · a21�):
Similarly, if set (A · A2)�= [�1; �2]; the following
�1 = (a1� · a12�)∧ (a2� · a12�)∧ (a1� · a22�)∧ (a2� · a22�);�2 = (a1� · a12�)∨ (a2� · a12�)∨ (a1� · a22�)∨ (a2� · a22�)
hold, and we may easily obtain
|a1� · a11�− a1� · a12�)|6|a11�− a12�| · |A|6|A| ·D(A1; A2);|a2� · a11�− a2� · a12�)|6|a11�− a12�| · |A|6|A| ·D(A1; A2);|a1� · a21�− a1� · a22�)|6|a21�− a22�| · |A|6|A| ·D(A1; A2);|a2� · a21�− a2� · a22�)|6|a21�− a22�| · |A|6|A| ·D(A1; A2):
(4)
So, by Lemma 1 and (4), the following
dH((A · A1)�; (A · A2)�) = dH([�1; �2]; [�1; �2])= |�1 − �1| ∨ |�2 − �2|6 |A| ·D(A1; A2)
316 P. Liu / Fuzzy Sets and Systems 119 (2001) 313–320
hold. Thus, D(A · A1; A · A2)6|A| ·D(A1; A2): Hence,(i) holds.(ii) ∀�∈ [0; 1]; the following hold by Lemma 2:
dH
((n∑k=1
Wk
)�
;
(n∑k=1
Vk
)�
)
= dH
([n∑k=1
w1k�;n∑k=1
w2k�
];
[n∑k=1
v1k�;n∑k=1
v2k�
])
=
∣∣∣∣∣n∑k=1
(w1k� − v1k�)∣∣∣∣∣ ∨
∣∣∣∣∣n∑k=1
(w2k� − v2k�)∣∣∣∣∣
6
(n∑k=1
|w1k� − v1k�|)
∨(
n∑k=1
|w2k� − v2k�|)
6n∑k=1
(|w1k� − v1k�| ∨ |w2k� − v2k�|)
=n∑k=1
dH((Wk)�; (Vk)�):
So
D
(n∑k=1
Wk ;n∑k=1
Vk
)
6∨
�∈[0;1]
(n∑k=1
dH((Wk)�; (Vk)�)
)
6n∑k=1
D(Wk ;Vk);
hence (ii) holds.
By Lemma 3, if f :R→R is continuous, A∈F0(R); and A ·f :R→F0(R) is de�ned as follows:(A ·f)(x)=f(x) · A (x∈R); then A ·f is continuouson R:
Theorem 1. Assume that Fi :R→F0(R) (i=1; 2;: : : ; n) are fuzzy valued functions; and C is the uni-versal approximator respectively for F1; : : : ; Fn: Then;if C is the additive class; C is also the universalapproximator for F1 + · · ·+ Fn:
Proof. By the inductive method, it su�ces to showthe conclusion when n=2: Given an arbitrarily com-pact set U ⊂R and ”¿0; by the assumptions that C isthe universal approximator for F1 and F2; respectively,there are H1; H2 ∈C; such that
∀x∈U; D(F1(x); H1(x))¡”2;
D(F2(x); H2(x))¡”2:
(5)
Set H =H1 + H2; then H ∈C, and by Lemma 3,∀x∈U; the following hold:
D((F1 + F2)(x); H (x))
= D(F1(x) + F2(x); H1(x) + H2(x))
6D(F1(x); H1(x)) + D(F2(x); H2(x))
¡”2+”2= ”:
So, C is the universal approximator for F1 + F2:
By Theorem 1, it is obvious that if C is theuniversal approximator, respectively, for fuzzy func-tions F1; : : : ; Fn; then C is also the universal approx-imator for the linear combination c1 ·F1 + · · · +cn ·Fn (c1; : : : ; cn ∈R):If C is the approximator for F; the properties of F
are similar to those of C:
Theorem 2. Let F :R→F0(R) be a fuzzy valuedfunction; and C be the universal approximator for F;then if C is continuous; F is continuous on R:
Proof. Suppose the conclusion is false, then there arex0 ∈R and ”0¿0; such that ∀�¿0; there is x′ ∈R:|x0− x′|¡�;D(F(x′); F(x0))¿”0: Select, respectively,�=1; 1=2; : : : ; 1=n; : : : ; correspondingly, we obtain thesequence {xn | n∈N}⊂R; such that ∀n∈N; |xn − x0|¡1=n; D(F(xn); F(x0))¿”0: Set U={x0; x1; : : : ; xn;: : :}:Obviously,U ⊂R is a compact set by the fact thatxn→ x0 (n→+∞): The assumptions imply that thereis an H ∈C; such that ∀x∈U; D(F(x); H (x))¡”0=4:On the other hand, H is continuous at x0 by the conti-nuity of C; so there is �0¿0 satisfying the followingfacts:
∀x∈U; |x − x0|¡�0⇒D(H (x); H (x0))¡”04:
P. Liu / Fuzzy Sets and Systems 119 (2001) 313–320 317
Let n0 ∈N satisfying 1=n0¡�0; thusD(H (xn0 ); H (x0))¡”0=4: Therefore,
”06D(F(xn0 ); F(x0))
6D(F(x0); H (x0)) + D(H (x0); H (xn0 ))
+D(H (xn0 ); F(xn0 ))
¡”04+”04+”04¡”0:
This is a contradiction. So, F is continuous on R:
Let
P=
{P |P(x)=
n∑i=1
pi(x) · Ai; n∈N; A1; : : : ; An ∈
F0(R); pi is a polynomial with respect to x}:
Each element P in P is called the fuzzy-valuedpolynomial. Obviously, P is continuous.To show universal approximations of continuous
fuzzy-valued functions by P; we at �rst prove thefollowing lemma.
Lemma 4. Suppose closed interval [a; b]⊂R; andF : [a; b]→F0(R) is continuous. De�ne the realfunction f : [a; b]× [a; b]→R as follows: ∀x; y∈[a; b]; f(x; y)=D(F(x); F(y)): Then f is continuouson [a; b]× [a; b]:
Proof. Given arbitrarily (x0; y0)∈ [a; b] × [a; b]; and”¿0: By the continuities of F at x0 and y0; respec-tively, there is a �¿0; such that
∀x; y∈ [a; b]; |x − x0|¡�⇒D(F(x); F(x0))¡”2;
|y − y0|¡�⇒D(F(y); F(y0))¡”2:
So, ∀(x; y)∈ [a; b]× [a; b]; if |x− x0|¡�; |y−y0|¡�;we have by the triangle inequality of metric:
D(F(x); F(y))
6D(F(x0); F(y)) + D(F(x0); F(x))
6D(F(x0); F(y0)) + D(F(y0); F(y))
+D(F(x0); F(x));
which implies D(F(x); F(y))−D(F(x0); F(y0))6D(F(x0); F(x))+D(F(y0); F(y)): For the samereason, the following holds:
D(F(x0); F(y0))
6D(F(x0); F(y)) + D(F(y); F(y0))
6D(F(x); F(x0)) + D(F(x); F(y))
+D(F(y); F(y0));
which implies D(F(x0); F(y0))−D(F(x); F(y))6D(F(x0); F(x)) + D(F(y0); F(y)): Therefore,
|D(F(x); F(y))− D(F(x0); F(y0))|6D(F(y0); F(y)) + D(F(x0); F(x));
i.e.
|D(F(x); F(y))− D(F(x0); F(y0))|¡”2+”2= ”:
So, f is continuous at (x0; y0); which proves thelemma.
Theorem 3. Suppose F :R→F0(R) is continuousfuzzy valued function; then for an arbitrary bound-edly closed interval [a; b] and ”¿0; there is a P ∈P;such that ∀x∈ [a; b]; D(F(x); P(x))¡”:
Proof. It will not harm to assume [a; b] = [0; 1]:Since F is continuous on [a; b]; F is uniformlycontinuous on [a; b]: Consequently, there is a �¿0;such that ∀x; y∈ [0; 1]; if |x−y|¡�; we haveD(F(x); F(y))¡”=2: So if there is an n∈N; thefollowing
∀k =1; : : : ; n; ∀x; y∈[k − 1n;kn
];
D(F(x); F(y))¡”2
holds. Let
P(x)=n∑i=0
(ni
)xi(1− x)n−i ·F
(in
)(x∈ [0; 1]):
Then P ∈P: By Lemma 4, f(x; y),D(F(x); F(y)) iscontinuous on [0; 1]× [0; 1]; hence there is an M¿0;such that ∀x; y ∈ [0; 1]; D(F(x); F(y))6M:We may
318 P. Liu / Fuzzy Sets and Systems 119 (2001) 313–320
easily show that
n∑i=0
(ni
)(i − nx)2xi(1− x)n−i
= nx(1− x)6n2
(x ∈ [0; 1]):
So, considering the fact that
n∑i=0
(ni
)xi(1− x)n−i=1;
n∑|x−i=n|¡�
(ni
)xi(1− x)n−i61
and Lemma 3, we obtain for x∈ [0; 1] that
D(F(x); P(x))
6n∑i=0
D((n
i
)xi(1− x)n−i ·F(x);
(ni
)xi(1− x)n−i ·F
(in
))
6∑
|x−i=n|¡�
(ni
)xi(1− x)n−iD
(F(x); F
(in
))
+∑
|x−i=n|¿�
(ni
)xi(1− x)n−iD
(F(x); F
(in
))
6”2+
∑|x−i=n|¿�
(x − i=n�
)2(ni
)xi(1− x)n−i
×D(F(x); F
(in
))6”2+
M2n�2
: (6)
Let n¿M=�2”, then ∀x∈ [0; 1]; D(F(x); P(x))¡”:
Obviously, Theorem 3 holds if we substitute a com-pact set U ⊂R for [a; b]:
4. Approximation capability of a class of RFNNs
In this section, we introduce four-layer feedforwardRFNNs, by which universal approximations of con-tinuous fuzzy valued functions may be investigated.De�ne
Fig. 1. Two hidden layer feedforward RFNN.
H=
H |H (x)=
n∑i=1
Wi
m∑j=1
Vij · �(x · Uj + �j)
n; m∈N; Wi; Vij ; Uj; �j ∈F0(R)
;
where � is the extended function of the given function� :R→R:
Remark 3. ∀H ∈H; H is a four-layer feedforwardRFNN with activation function �; threshlod vector(�1; : : : ; �m) in the �rst hidden layer. The topologicalarchitecture is as shown in Fig. 1.
Restricting fuzzy numbers Uj; Vij ; �j, respectively,to be uj; vij; �j ∈R; we obtain the subsetH0 ofH:
H0 =
H |H (x)=
n∑i=1
Wi
m∑j=1
vij · �(uj · x + �j)
n; m∈N; Wi ∈F0(R); vij; uj; �j ∈R :
Obviously, both H and H0 are additive classes,and if � :R→R is continuous on R; H is continuousand so isH0:
Theorem 4. Let one of the following conditions holdfor � :R→R:(i) � is a sigmoid function; i.e. � is bounded and
limx→−∞ �(x)= 0; limx→+∞ �(x)= 1;
P. Liu / Fuzzy Sets and Systems 119 (2001) 313–320 319
(ii) � is a bounded; continuous and nonconstantfunction on R:
Then; given arbitrarily that P ∈P; H0 is the uni-versal approximator for P; consequently; so isH:
Proof. Given arbitrarily that ”¿0; and compact setU ⊂R; let P(x)= ∑n
i=1 pi(x) · Ai (x∈R): Since fori=1; : : : ; n; pi is a polynomial, so pi is continuous onR: Therefore, by the conclusions in [5,8], there areK ∈N and vj; uj; �j ∈R (j=1; : : : ; K); such that eithercondition (i) or (ii) holds, the following hold:
∀x∈U;∣∣∣∣∣∣pi(x)−
K∑j=1
vj · �(uj · x + �j)∣∣∣∣∣∣¡
”
|Ai|:
(7)
Set Hi(x)=(∑K
j=1 vj · �(uj · x + �j))· Ai (x∈R):
Then Hi ∈H0; and the following hold for each x∈Uby Lemma 3 and (7):
D(pi(x) · Ai; Hi(x))
= D
pi(x) · Ai;
K∑j=1
vj · �(uj · x + �j) · Ai
6|Ai| ·∣∣∣∣∣∣pi(x)−
K∑j=1
vj · �(uj · x + �j)∣∣∣∣∣∣¡”:
So, H0 is the universal approximator for Fi: Fi(x)=pi(x) · Ai (i=1; : : : ; n): By the fact that H0 is anadditive class and Theorem 1, H0 is also the uni-versal approximator for P: P(x)=
∑ni=1 Fi(x)=∑n
i=1 pi(x) · Ai; and so isH by the fact thatH0⊂H:
Theorem 5. Let � :R→R be bounded; continuousand nonconstant; and F :R→F0(R) be continuous.ThenH0 is the universal approximator for F; so isH:
Proof. By Theorem 3, P is the universal approxima-tor for F: Theorem 4 implies that for each P ∈P; H0
is the universal approximator to P: Consequently, thetheorem holds by Remark 2.
Theorem 6. Suppose � :R→R is a continuous;bounded and nonconstant function; and F :R→
F0(R) is a fuzzy valued function; then the followingare equivalent:(i) F is continuous;(ii) H0 is the universal approximator for F ;(iii) H is the universal approximator for F ;(iv) P is the universal approximator for F:
Proof. Theorem 5 implies that (i)⇒ (ii); (ii)⇒ (iii)is trivial; Theorem 2, and the continuity of H implythat (iii)⇒ (i). By the continuity ofP and Theorems 2and 3, we may prove that (i)⇔ (iv).
Considering the fuzzy-valued functions F1; F2 :R→F0(R):
F1(x)= A1 sin(x · B1) + · · ·+ An sin(x ·Bn);F2(x)= A1 cos(x ·B1) + · · ·+ An cos(x ·Bn)for A1; : : : ; An; B1; : : : ; Bn ∈F0(R) are continuous onR, we obtain by Theorem 6 thatH0; as well asH isthe universal approximator for F1; F2, respectively.
5. Conclusions
This paper discusses the approximation capabilitiesof multi-layer feedforward RFNNs to fuzzy valuedfunctions. If the activation function in the �rst hiddenlayer is continuous, bounded and non-constant, someequivalent conditions for fuzzy valued functions thatcan be arbitrarily closely approximated by the RFNNsare shown. The conclusions obtained here general-ize the ones in [1,2,10]. For the same reasons, wemay obtain similar conclusions for the multi-variablefuzzy valued functions that Rd→F0(R): The fur-ther problems for future research are the universal ap-proximations of fuzzy functions thatF0(R)d→F0(R)by RFNNs. Can some equivalent conditions be con-structed for such functions?
References
[1] J.J. Buckley, Y. Hayashi, Can fuzzy neural nets approximatecontinuous fuzzy functions, Fuzzy Sets and Systems 61 (1)(1994) 43–51.
[2] J.J. Buckley, Y. Hayashi, Can neural nets be universalapproximators for fuzzy functions, Fuzzy Sets and Systems101 (1999) 323–330.
320 P. Liu / Fuzzy Sets and Systems 119 (2001) 313–320
[3] T.P. Chen, Neural network and its approximation problems insystem identi�cation, Sci. China Ser. A 24 (1) (1994) 1–7.
[4] T. Chen, H. Chen, Approximation of continuous functionalsby neural networks with applications to dynamic systems,IEEE Trans. Neural Networks 4 (6) (1993) 910–918.
[5] T.P. Chen, H. Chen, R.W. Liu, Approximation capabilityin C(Rn) by multilayer feedforward networks and relatedproblems, IEEE Trans. Neural Networks 6 (1) (1995) 25–30.
[6] P. Diamond, P. Kloeden, Metric Spaces of Fuzzy Sets, WorldScienti�c, Singapore, 1994.
[7] Y. Ito, Representation of functions by superposition of a stepor sigmoidal function and their applications to neural networktheory, Neural Networks 4 (1991) 385–394.
[8] H. Kurt, Approximation capabilities of multilayer feedforwardnetworks, Neural Networks 4 (1991) 251–257.
[9] P.Y. Liu, Analyses of regular fuzzy neural networks forapproximation capabilities, Fuzzy Sets and Systems 114(2000) 329–338.
[10] P.Y. Liu, H.X. Wang, Universal approximation of a class ofcontinuous fuzzy valued functions by regular fuzzy neuralnetworks, Electron. Sinca. 25 (11) (1997) 41–45.
[11] J.G. Luo, J. Yu, H. Wang et al., Research on analysismethods of structural approximations based on arti�cial neuralnetworks, Sci. China Ser. A 24 (6) (1994) 653–658
[12] H.T. Nguyen, A note on the extension principle for fuzzysets, J. Math. Anal. Appl. 64 (2) (1976) 369–380.
[13] F. Scarselli, A.C. Tsoi, Universal approximation using feed-forward neural networks: a survey of some existing methods,and some new results, Neural Networks 11 (1) (1998) 15–37.