о뷢չ
JOURNAL OF COMPUTER RESEARCH AND DEVELOPMENT
1999 36 5 Vol.36 No.5 1999



BĹ
춶ʿͬ
ժҪBԣڴ˻ϲùԵķ֤Bܹ⾫ȱƽⶨϵʵ.˹㷨ʹô㷨Ҫ£СB.
ؼʡB磬
ͼšTP18
CONSTRUCTIVE THEORY FOR B-SPLINE NEURAL NETWORK
YU Dong-Jun and WANG Shi-Tong
(Department of Computer Science,East China Shipbuilding Institute,Zhenjiang212003)
AbstractThe characteristics of B-spline basis function is first discussed, based on which B-spline neural network is proved to be a universal approximator. And then a constructive algorithm is presented. It is proved that this algorithm can be used to build a B-spline neural network with minimum hidden units to approximate any continuous function defined on compact set to a prescribed accuracy.
Key wordsB-spline, neural network, constructive theory
1
һķԺǶ֪ľƽԭȫֱƽ1.ʹõ3.3ԣӳʵͨļȨʵֵģȡͬʱвͬʣҼ㸴ӳ̶Ȳͬ.õĻSigmoid˹͵ȵȣȡBʱƴΪB.
ͼ1ʾ3磬ӳͨi()ļȨʵֵģy==(x1,x2,,xn)T.ȡΪBʱƴ3ΪB.Bмõʣ2ݣԣһԣԣBõ˼Ϊ㷺Ӧ.


ͼ1B˽ṹ
ȻBźܶŵ㣬ǵһʵӦʱڵĿѡȡȴǸ⣬δָ.ľּڽ⣬ҵСڵĿ㷨.
2B
BǹϳõһֲֵϹʵҪӦ÷Χܹ.
ٶΪ=(x1,x2,,xn)TxiUiUiΪ䣬ԶΪ
Ui={xi|xminixixmaxi}
Եiάзָ
xmini<i,1<i,2<<i,mi<xmaxi
i,jΪxiĵjڲ.
<i,-1<i,0=xmini; xmaxi=i,mi+1<i,mi+2<
Ϊ.
1 
ΪU=xmin,xmaxݣڴmڲ㣬֮߷ֱk-1.kΪBĽ.ɶ嵥Ϊ
j(x)=Nkj(x)
У

ͼ2ʾk=3ʱB.


ͼ23B
B3Ҫʣ2ݣ
 ԣNkj(x)>0xʣۦj-k,j),1jm+k.
 ԣNkj(x)=0xۦj-k,j)1jm+k.
 һԣxU.
ǰѼϦ={Nkj(x),j=1,2,,m+k}ΪkB.
(2) 
Ϊ=(x1,x2,,xn)TRnNkii,ji(xi)ʾxi(i=1,2,,n)ĵji(ji=1,2,,mi+ki)ҵxiĻΪki.ΪҦj1j2jn()ͬԣԣһ.
 ԣj1j2jn()>0,ڪСi,xiʣۦi,ji-mi,i,ji),1jimi+ki.
 ԣj1j2jn()=0,ڪi,xiʣۦi,ji-mi,i,ji),1jimi+ki.
 һԣ =(x1,x2,,xn)TU1U2Un.
֤ף2.
3BĹ
ڱУǽһöͨķõһB磬ԸľȱƽĶϵĶʵ.
ȣ뼸4.URΪA(x)ΪUϵĺ.x*UʹöxU,A(x)A(x*)A(x)Ϊֵx*Ϊֵ.
1. Ϧ={Ai(x),i=1,2,,m}еÿAi(x)ΪֵҶxUг.ƦΪֵ.
ȻB㶨1.
2. ڸĺf:URԼһֵ={Ai(x),i=1,2,,m}Ƕ庯:URΪx*iΪAi(x)ķֵ.Ϊڦ϶fĲֵ.
Ai(x)ȡBʱΪڦ϶fBֵ.ʱͶӦһB磬ͼ3.


ͼ3B
3. i(xi)={Aij(xi),j=1,2,,mi}ΪUiRϵķֵi=1,2,,nô={A1i1(x1)A2i2(x2)Anin(xn),1ijmi,1jn,xiUi},С ΪU1U2Unϵĳ˻ֵ.
ɵ2ڵ˵B϶3ǳ֮ΪB˻ֵ.
ڣԸ1.
1. f:RΪRnϵʵڦ>0ϵһB˻ֵ={A1i1(x1)A2i2(x2)Anin(xn)}ʹó()Ϊڦ϶fBֵ.
֤. Ϊ˱ϵļ㣬ǼٶάΪ2=0,12άϵBĽΪ3.
fģܵģfضһ.˶ڦ>0ضڦ>0ʹõʱг(x1,y1),(x2,y2).
xὫ[0,1]ȻNȷݣڲ㣬߸Ӧ㣨ΪΪ3N㹻ʱ.ͿxᶨB={Aj(x),j=1,2,,N+2}Aj(x)=N3j(x)ͼ4ʾ.


ͼ4xᶨB
ͬyᶨB={Bj(y),j=1,2,,N+2}Bj(y)=Aj(y),j=1,2,,N+2.ȻԹ˻B
={Cij(x,y)|Cij(x,y)=Ai(x)Bj(x),i,j=1,2,,N+2}
ݶ2ע⵽Ai(x)ķֵΪڦ϶fBֵдΪ(x,y)=.
һ(x,y)ʣܿԽ䶨λڣ

0i,jN-1.
Ai(x),Bj(y)Ϊ3Bÿһά3Ϊ㣬ʽдΪ









УBĹһ
a1+a2++a9=Ai+1(x)+Ai+2(x)+Ai+3(x)ݣBi+1(x)+Bi+2(x)+Bi+3(x)=11=1.
ʧһԣֱΪʽеСֵʽ
ҡ
ʽУ




ڵ͵֮ľܦģ




ڵ(x,y)֮ľܦģУ

ۺʽУ

ϵ(x,y)ԣ֤˶1.
1ܹ֤ⶨϵʵBЩʵֵBľȱƽԭûиûķʵӦú׵».һ㷨ʹø㷨ҳСĻ.
4㷨
ͬΪ˼㷨Ƕά.
f:Ϊʵһţǿҵһģʹõʱ.ڦģá(x0)ʾx.NɽΪN2.
㷨1. ʼ񻮷֣ǳʼνļֵS0
FOR 0<i,j<N
{

 


}
tʾt(x,y)ʾڵtεʱ.ʼʱt=00(x,y)=f(x,y).
1. t=t+1,


ٶStxStyֱp,qԪأôܿԽЩԪΪ
StxУ0x1<x2<<xp1,(1pN-1)
StyУ0y1<y2<<yq1,(1qN-1)
Stx[01]߷ֱʵx-2,x-1xp+2,xp+3(ȡ,)0Ϊx01Ϊxp+1.ʹЩ㣬Ϳxᶨ3B
t={Ati(x)|Ati(x)=N3i(x),1ip+3}
ʽN3i(x)ǰʽݹ鶨ģ

1ip+3.
ͬyᶨ3B
t={Btj(y)|Btj(y)=N3j(y),1jq+3}
ʽеN3j(y)ͬǵݹ鶨ģ

1jq+3.
2. æt,t,Ϊt(x,y)ֵt

xi,yjֱΪAti(x),Btj(y)ķֵ.
æt+1(x,y)=t(x,y)-t(x,y)St
FOR 0i,jN(1)
AND 

(2)
OR
(3)
 
 StEt+1St+1;
 }
3. IF Et+1ΪռTHEN 
(x,y)=0(x,y)+1(x,y)++t(x,y)
㷨.
ELSE ת1.
ǾҪ֤㷨BҪţҲȻҪ㶨1Ľ.
2. 㷨BȻ㶨1Ľ.
֤. 㷨Ĳ3֪ڵʽ(1)ضʽ(2)ʽ(3)㷨ܼ.Et+1ΪռҲζŶĵʣ㣺


Ҳ˵ʽ

ĵ(x,y)ʣܿȷ䷶Χ

NִʱضУ

֤.
5ᡡ
Ĳ˹Եķ֤BȻȫֱƽҸ켸СڵĿ㷨ʹΪBʹṩݣԵʵӦ.
õȻѧ(Ŀ 69673001).
߼飺춶У197410˶ʿооΪ硢˹.ʿͬУ196410ڣӢBristolѧʿҪоΪ˹ܡ硢ģѧ.
ߵλҵѧԺѧϵ򽭡212003
ο
1.˹ģͼӦ.Ϻѧ磬1993
(Zhang Liming. Models and Applications of Artificial Neural Networks(in Chinese). Shanghai:Fudan University Press,1993)
2ߣˣ־.ܿ뼼.廪ѧ磬1991
(Sun Zengqi,Zhang Zaixing,Deng Zhidong.Intelligent control theory and its applications.Peking:Tsinghua University press,1997)
3Lin C T. Fuzzy Neural Networks. England: Prentice-Hall Press, 1997
4Wang Peizhuang et al.Constructive theory for fuzzy systems.Fuzzy Sets and Systems,1997,88(2):180185
5ʵ.絼.ɳƼѧ磬1993
(Hu Shouren et al. Introduction to Neural Networks(in Chinese). Changsha:National University of Defence Technology Press, 1993)
ԭյڣ1998-07-16
޸ĸյڣ1998-12-04
