Զѧ
ACTA AUTOMATICA SINICA
1997ꡡ234ڡVol.23No.41997



CMACѧϰԵо
ҡлع
ժҪCMACѧϰ̵ȼԷGauss-Seidel һʵоѧϰ̵.þƳٶ.Ϊʡ洢ռʩhashĲӰ졪ƻ̬.Ϸڵԭ.
ؼʡCMACԣGauss-Seidel, hash.
A STUDY OF THE CONVERGENCE OF
THE CMAC LEARNING PROCESS
LUO ZHONG
(Dept. Image Processing,IRSA,CAS,P.O.Box 9718, Beijing 100101)
XIE YONBIN
(Dept.Automatic Control,School of Electronic and Informatioin Engineering,Xian Jiaotong University,Xian 710049)
ZHU CHONGGUANG
(Dept. Image Processing, IRSA, CAS, P.O.Box 9781, Beijing 100084)
AbstractBased on the fact that the CMAC learning process is equivalent to the Gauss-Seidel iteration for solving a linear system of equations,this paper addresses the convergence of the CMAC learning process.By means of matrix analysis,the convergence speed is estimated.The negative effect on convergence of hash coding,i.e.,it deteriorates the convergence performance is considered and the reason for its existence is also theoretically analyzed.
Key wordsCMAC,convergence,Gauss-Seidel iteration,hash conding.
1
CMACСģ͹ؽڿAlbus1975״һ˹12.CMACԽṹѧϰ㷨򵥣뷺.ѳɹӦڻ˿ƣ3ݡģƣ4ݡʱз5ݵ.CMACһṹͼ1ʾ.


ͼ1CMACһṹ
CMACռΪXRd,XdΪӦɢռ.뵥Ԫеǿ.뼤һA*ǿȣǵĴͼΪӦ֮.xkXd伤ǿΪΪ

(1)
dkΪѧϰdk-ykԸA*ǿ.

(2)
Ȼ뵥ԪڴռǾ޴.Ϊʡ洢ռ.Albushash.ǿȴһ뵥ԪνhashԪУ뵥ԪֻhashԪɢеַ.
2CMACѧϰ̵ʵ
XdһӼXtΪѵѵȡXt.ѧϰɶڹ.ÿһڣһ˳Xtѵ.ѧϰС磬ѧϰ.Ϊֻһά.
輤ǿΪwk+1,,wR+A*,ǿΪ M=N+A*-1.I={i|0iN-1,iZ},S={s|1sM,sZ}, :IS{0,1}.

(3)
Xt=NtIt={i|0iNt-1, iZ}ͦt:ItS{0,1}.

(4)

(5)

(6)
D=[d0,d1,,dNt-1]T,
(7)
W=[w1,w2,,wm]T,
(8)
E(k)=[E0(k),E1(k),,ENt-1(k)]T,
(9)
A=[t(i,j)]NtM,iIt,jS,
(10)
C=AAT,
(11)

(12)
LRΪC(ԽԪ)(ԽԪ).
CMACѧϰĿԷ
AW=D
(13)
NtN<M(13)Ϊ.WongSideris֤CMACѧϰ̵ȼԷGauss-Seidel6.
CE=D,
(14)

LE(k)=D-RE(k-1).
(15)
һѧϰڶӦһε.W()E()ڣ
W ()=ATE()=AT(AAT)-1D
(16)
(13)СLS7.
3ԷһЩ
1.AΪȾ.
֤.
2.CԳƾ.
֤.ԳΪȻ.AȣֵΪCֵΪ.C.֤.
3.RΪnnʵԳQΪһnnʵQ+QT-RΪ-Q-1(R-Q)֮ȫֵλԲG|z-(1-)|=ڲ(10)֮Բ.

Ϊڻ.
֤.Q+QT-Rʶu0uRn
(Q+QT-R)u,u=Qu,u+QTu,u-Ru,u>0,
r>1.Ϊ-Q-1(R-Q)֮һֵvΪӦ֮λ
-Q-1(R-Q)v=v,

(1-)Qv=Rv,

(1-)Qv,v=Rv,v,
Hilbertռڻ
(1-)QTv,v=Rv,v,
=+j£



Ȼ



дɸģʽΪ|-(1-)|,Re()<1.֤.
L+LT-C=A*I.ɶ3֪-L-1RȫֵλԲGڲͳ(10)Բ.ԲGİ뾶С1뵥λԲ(10)㣬ಿλڵλԲ.-L-1RȫֵλڵλԲ.ɣ8ݿ֪
E(k)=-L-1RE(k-1)+L-1D
(17)
.Ӷɵȼ֪CMACѧϰ.Ͻhashµõ.½hash.
1.3CLRQУЦmax(C)=max(C).(C)ʾCֵ.
֤.


4.L-1RֵСģС

֤.Աʾ{z|z-(1-)|Re(z)<1},-L-1Rֵȫ.ֻ֤m=|z|.(00)ʣȻz=0.ʱΪı߽ʵύ(1)ȻΪm=0m=|z|.(00)ɶԳ֪ģСĵһ߽ʵĽ.Ӷ=.ʱ0m=m=z|.m=|z|.֤.
CMACѧϰٶ(17)ʽͬ.ȡ-L-1Rֵ.ݶ4֪-L-1RֵСģΪ0.˵ֵнСģ..
2.(ԭ)8RΪnnHermiten>m. R1ΪRǰn-mǰn-mཻɵ.RֵΪ1ݦ2ݡݦnR1ֵΪ1ݦ2ݡݦn-m
iݦiݦi+m,1in-m.
5.max(C0)=max(C0),
max(C0)ݦmax(C).
֤.CC0ıΪn1,n2,,nNtNtкNtཻɵ.ûPʹC0ǰNtкNtཻɵ0=PC0PT.P󣬹ʦ(0)=(C0)Ԧmax(C0)=max(0)ݦmax(C).֤.
϶˵ȫΪѵɴﵽٶ.
4hashԵӰ
hashһǻνײ1ݣ뵥ԪͬʱָͬһhashԪ.ǰںʱײǲɱ.
һײCMACѧϰԻֿܵܵӰ.
Aʧȥ.뿴.Nt=3M=5,A*=3뵥ԪΪ5.hashǰ

ٶײʹ뵥Ԫ14ָhashԪ1ӶAC

A.

Ȼѧϰ.
Σѧϰײήٶ.ײʹAس̶󣬴ӶʹAֵСʹæmax(C)С.ӶʹٶȽ.
һѧϰڣײʹúѵǰѵѧϰЧ.Ӷʹѧϰ仯񵴣̬仵.
5
ǰķȫڶά.ֻ֤ʽ.һά֤ǵ.Ϊѧϰº.
f(x,y)=sin2x.sin2y,(x,y)[0,1]2.
ɢ
fd(i,j)=sin(2.(i)/(30)).sin(2.),0i,j30.
ȡXt={(5k,5l)|0k,l6},A*=10.ͼ2ͼ3ͼ4ʾ.


ͼ2f(x,y)ͼ


ͼ3CMACѧϰ(hash)


ͼ4CMACѧϰ(hash)
5
ͨ۷ͷʵ飬.
(1)hashCMACѧϰ̱.
(2)hashCMACѧϰײӰ3ֿܽ1)2)3)̬仵.
߼飺ҡС196911996ͨѧԶۼӦרҵʿѧλ.йѧԺңвʿо.оΪۡźŴͼ.ѷ10ƪ.
ߵλ:ҡйѧԺңͼҡ100101
лҵѧԺϵ710032
ع⡡йѧԺңͼҡ100101
ο
1ݡAlbus J S.A new approach to manipulator control: The cerebellar model articulation.Controller(CMAC).Trans of ASME Joural of Dynamic Systems,Measurements,and Control,1975,9:220227.
[2]Albus J S.Data Storage in the Cerebellar Model Articulation Controller(CMAC).Trans of ASME Journal of Dynamic Systems,Measurement,and Control,1975,9:228233.
[3]Miller W T.Real-time Application of Neural Networks for Sensor-based Control of Robots.with Vision,IEEE Trans.Syst.,Man,Cybern.,1989,19:825831.
[4]Nie J Linkens D A.FCMAC:A Fuzzified Cerebellar Model Articulation Controller with Self-organizing Capacity,Antomatica,1994,30(4):655664.
[5]Touretzky D S.Ed.Neural Information Processing Systems 1,Los Altos,CA:Morgan Kaufmann,1989.2939.
[6]Wong Y F,Sideris A.learning convergence in the cerebellar model articulation controller.IEEE Trans.Neural Networks,1992,3(1):115120.
[7]Franklin J S.Matrix theory.Englewool Cliffs,N.J.:Pretice-Hall,Inc.,1968.261268.
[8]ͣڱ..ҵѧ磬1991.2465.
ոڡ1995-08-03
