о뷢չ
JOURNAL OF COMPUTER RESEARCH AND DEVELOPMENT
1999ꡡ3611ڡVol.36No.111999



ξȫֱƽ
춶ʿͬС
ժҪһֲξHRBFN (hierarchical radial basis function network)֤HRBFNһȫֱƽ.HRBFNʺھвνṹӦмΪҪӦ.
ؼʣȫֱƽ磬
ͼţTP183
UNIVERSAL APPROXIMATION THEORY OF HIERARCHICAL
RADIAL BASIS FUNCTION NETWORK
YU Dong-Jun, WANG Shi-Tong, and WU Xiao-Jun
(Department of Electronics and Information. East China Shipbuil ding Institute, Zhenjiang 212003)
AbstractIn this paper, a new hierarchical radial basis function network HRBFN is presented, and its universal approximation property is proved. HRBFN is very suitable for application domains with hierarchical structures, and therefores has very important theoretical and practical significance.
Key wordsuniversal approximator, hierarchical network, radial basis function network
1
ѧ߶RBFN(radial basis function network)о˽ܳ.ף1ף3ݶRBFNıƽԼӦоڵѡȡζԲϵͳӦRBFN沢δ漰.RBFNĹۣҸ˹켸ڵRBFN㷨4.HRBFNּڽζԲϵͳӦRBFN.
(RBFN)õ˹㷺Ӧãȡ˺ܺõЧ. ݲþʱʹõľ벻̫࣬ΪĿӣΪʹܴﵽϺõӳ䣬ڵĿͨ輱.ΪҪǣȻдһӦֲϵͳڲιϵĳ(ĳЩ)ҪĳЩһӳõٺʹһϵͳ.ʹͨľƽϵͳͻϵͳеĲԣʵϷǳ.
ǿܷϵͳĲԹӦĲ;HRBFNȱ˾ŵ㣬Ϻܺõطϵͳ.ĵĹǿ϶.ǲξǷȫֱƽHRBFNֻܹƽһ޵Իôʹ÷ΧȻС.οܹ֤HRBFNһȫֱƽ˵ʵʹòξ(HRBFN)ƽȫе.͵HRBFNͼ1ʾ.


ͼ1͵Ĳξ(HRBFN)
ͨRBFNѵʱһþʽSOMȷļȣ㵽 ȨϵС˷ȷпٵŵ.ͬHRBFNеÿһRBFNþʽSOMȷģڱĵĵ3Уһ򵥵˵Ȩϵȷ̣ͨ󷽳ȷͬпټ򵥵.
ڵ2֣ǹһξ粢оһЩʣڵ3֣ʹ ù췽֤˲ξȫֱƽԣ˽.
2HRBFNĹ켰
ξĻ˼õάάľ .ΪˣһRBFN.
ͨľṹͼ2ʾ.ʧһԣ롢.=(x1,x2)ٶڵĿΪm2,ôΪ
(1)


ͼ2ṹ
зֱΪiڵx1x2ļ.
(2)
(3)
ڣ
(4)
ʽ(1)дΪ
(5)
ٿmmipͬʱֻҪ㹻СĶô СĦ1>0
Zpi(mip)=1(6)
Zpi(miq)<1,(pq)(7)
֤.ʽ(7)ȡ

ֻҪ

Ϳʹʽ(7).֤.
1㹻СʱΪZpi(miq)=0,(pq).
ͼ3ʾi=1x1ʽ(7)mͼ.


ͼ3x1mͼ
֪ʶ,ͿԹ;.
ٶnx1,x2,,xnfi(i)ʾHRBFNĵiRBFNiΪiRBFN.ô; Ĳ£
(1) һΪn1x1,x2,xn1RBFNҲ˵fi(i)еi=1һRBFNΪyi=f1(1)1=(x1,x2,,xn1);
(2)i=i+1i(i>1)RBFNni+1(ni1)Ϊ
fi(xNi+1,,xN\-i+n\-i,yi-1)
(3)ת(2).=n. 
ǾHRBFNĹ.һn1=2,ni=1ͼ1.ǽ֤HRBFNȫֱƽ.
3HRBFNȫֱƽ
ڱУǸHBFNȻȫֱƽ֤ǹԵģҲǟһH RBFNȻ֤HRBFNȫֱƽ.Ϊ˱ϵļ㣬.֤̿Ժܷؽƹ.һҪ⣺
1g()ΪһܼU =ۦ1,1ݡۦ2,2ݡۦ3 ,3ݪR3ϵʵ.ٶg()ʽδ֪ǶĳЩU ǿԵõg()ֵ.ǵҪһHRBFNȥƽg().
.
һxi(i=1,2,3)ķΧΪۦi,iݣy1Ϊ01ݣ ҲһRBFNΪx1,x2Ϊy1y1ʣ0,1ݣڵ Ϊm2.xiۦi,iݻΪm-1ȷݣRBFNļ ˹Zji(xi)ҪZji(xi)ʽ(7).Ϳɶxim Ϊ
(8)
(9)
(10)
 
ƵأǿԽ01ݻΪm-1ȷݣΪy1mҲ ø˹Ci(y1)= 1im, miy1mͬļ.ͬģĦ10ֻҪ㹻С
Ci(mi)=1(11)
Ci(mj)<1, (ij)(12)
1㹻СʱΪCi(mj)=0,(ij).
Ϳʽ(8)(9)(10)Ci(y1)ʽ(12)ʱΪy1mϸʽ׸.
ͼ4m=4ʱ.


ͼ4ĿΪ3ʱʾͼ(m=4)
ڶ峣p,q=1,2,,m. ôһRBFNΪ
(13)
0wpq1,y1ΪwpqļȨƽ,Աضy1ʣ0,1.
ڶRBNϵͳΪ
(14)
ǾҪȷwij2,£Zj3(m3j)=1, kjZk3(m3j)=0, Ҷy1ʣ0,1,Ci(y1)Ϊֵ.ʽ(14)ǿ֪ʽ
(15)
iΪ{1,2,,m-1}еĳ.
f2(wpq1,m3j)=g(m1p,m2p,m3j)õʽ
(16)
ǹ̶jȡpqʹ֮1仯m͵õʽ(16)ṃдɾʽ
(17)
ȡp1仯mqֲ.Ϊһm mľÿһֻΪԪpΪУi,i+1Ϊ.ͨ
(18)
ͿԵmwij2j1仯m󣬾ܹȷеm2wij2.
Ĳ͵õHRBFNΪ
f(x1,x2,x3)f2(f1(x1,x2),x3)(19)
ڣҪ֤ʽ(19)ʾHRBFNܹ⾫ƽδ֪g.
1. f(x1,x2,x3)Ϊʽ(19)ʾHRBFN Ҵƽg(x1,x2,x3)ΪܼU=ۦ1,1ݡۦ2,2 ݡۦ3,3R3ϵʵôʽ
(20)
С*maxΪ
֤. ֤,q,j=1,2,,mf(m1p,m2q,m3j)=g(m1p,m2q,m3j).
ʽ(13)֪p,q=1,2,,m,f1(m1p,m2q)=wpq1.ʽ (18)õwij2֤f2(wpq1,m3j)=g(m1p,m2q,m3j)ˣ
f(m1p,m2q,m3j)=f2(f1(m1p,m2q),m3j)=f2(wpq1 ,m3j)=g(m1p,m2q,m3j). 

(21)
ζŶġUضUpqjʹáUpqj.
ΪUһ㣬ضڶӦpx,qx,jx{1,2,,m-1} ʹáUpqj.
ȡ=(m1px,m2qx,m3jx)Tf()=g().ֵ
mixΪеĵiԪ.ڡUpqj
|xi-mix|bi.
(22)
ΪUڵһ㣬ʽ(22)ܹõʽ(20).֤.
2. ĶUR3ϵʵ Լ20ʽ
(19)HRBFNʽ
(23)
֤. U䣬ԾΪ.ˣֻҪʹbi㹻Сضʹ
(24)
ֻȡm㹻ʹbi㹻Сǿ԰쵽.֤.
֤һҪļʽ(7)Ҳ˵ʽ(6)еĦ1Ҫ㹻Сǿ԰쵽.
4ᡡ
ͨ3ڵ֤Ѿ֪HRBFNΪȫֱƽʵʵĲϵͳӦHRB FNȫе.⣬HRBFNܹʹRBFNڵĿӶ.һҪоУHRBFNķʼֵHRBFNȶӰ켰ǱĿǰĹ.
*õȻѧ ߼Ŀ(69843002)ҽίʦĿͽʡ˲Ż Ŀ.
߼飺춶У197410˶ʿоҪоΪ 硢˹.
ʿͬУ196410ڣʿʦҪ оΪģѧ硢˹.
СУ196712 ʿоҪоΪģʽʶϵͳģϵͳ.
ߵλҵѧԺϵ򽭡212003
ο
1Lin C T et al. Neural Fuzzy Systems. England Prentice-Hall Press, 1997
2Powell M J D. Approximation heory and Methods. CambridgeCambridge University Press, 1981
3Barron A. Universal approximation bounds for superpositions of sigmid funtion, IEEE Trans Information Theory, 1993, 39
4Wang Shitong, Yu Dongjun. Constructive theory of radial basis function network. Accepted by Journal of Advanced Software Research, 1999
5Wang Lixin. Universal approximation by hierarchical fuzzy systems. Fuzzy Sets and Systems. 1998, 93(2): 142148
6ʿ. ģơ񾭿ƺܿ. :ҵѧ, 1996
(Li Shiyong. Fuzzy Control, Neural Control and Intelligent Cybernetics( in Chinese)Harbin: Harbin Institute of Technology Press, 1996)
7. ˹ģͼӦ. Ϻ:ѧ磬1992
(Zhang Liming. Models and Applications of Artificial Neural Networks(in Chinese). Shanghai:Fudan University Press, 1992)
ԭյڣ1998-12-24޸ĸյڣ1999-08-20.
