Զѧ
ACTA AUTOMATICA SINICA
1997ꡡ234ڡVol.23No.41997



СģĽ㷨о1
ۡ
ժҪĽСģĻԭڷAlbus1㷨Ļϣָ˸㷨ѧϰʱȱ.ѧϰӦĸĽ㷨֤˸㷨ԣ˸øĽ㷨ٶȿص.
ؼʡ磬 Ӵձ룬 磬 CMAC.
AN IMPROVED CMAC NEURAL NETWORK ALGORITHM
LIU HUI XU XIAOMING 
(Department of Automation , Shanghai Jiaotong University, Shanghai 200030)
AbstractThe basic principle of CMAC (Cerebelllar Model Articulation Controller) is introduced. Based on a detail analysis of Albus algorithm1, the paper points out its drawback in batch learning. An improved algorithm is proposed and theoretical proofs are also given. Simulation results show that the improved method has higher speed and better convergence than the original.
Key words Neural network, hash-coding, associative network, CMAC.
1
1982Hopfield˹ڷԼRumelhart˷רPDP緶Χоȳ.˹Ԫ˹ڿƽеӦоչرгֱƽ⸴ӷԺΪӵķ⿪һƵľ빦ܵCMAC(Cerebellar Model Articulation Controller).
CMACJ.S.Albus21975.PerceptronƣȻÿԪϵһԹϵӽ忴CMACģʺڷԵӳϵ.ͬʱ㷨ʮּ򵥵Ħ㷨ٶȺܿ.һά״̬ռеӳ䵽һȽС.ֻҪԶά״̬ռвѧϰͿɴﵽ켣ѧϰͿƵĽ⣬رʺڻ˵ĹѧϰƣʵʱѧϰƣԺӳ䣬Լģʽʶ.CMACӦãӲʵ.
2CMACĻԭ
CMACļ򵥽ṹģͼ1ʾռSпܵSiɣCMAC罫ܵκ룬֪ͨMӳ䵽һܴ洢AеcԪ.ռڽڴ洢AвصĵԪԽصԽ.֮ռԶAвص.һʵϵͳռеǺܴģ磬ĳϵͳ10룬ÿȡ100ֵռ10010AĴ洢Ҫ10010Ԫ.ھѧϰⲢռе״̬Ĵ洢ռͨõļ洢ѹӴձ(Hash-coding)ӳ䵽һСöʵֵĴ洢A.κCMAC뼤cʵ洢λãЩλõֵӵõP.


ͼ1CMACģͽṹ
3AlbusCMAC㷨
蹲PѵaiΪMAӳ䣬CMACڵiѵΪf(Si=wTaiΪdiźŦi=di-f(si)CMACǰѧϰɵȨֵݶȷĻϲLMS㷨õģ2

(1)
ЦΪѧϰ.
ڵlѭʱԵkѵʱΪ/cΪk.CMACڵķռʵʴ洢AصʸƱӰʹiΪ

(2)
cikΪi,kAеصԪȻcik=cki.Iѭk䱾ۻΪ֮صÿȨΪkEkckΪۻȨ.ۻȨԵiΪcikk.CMAC㷨0kconst.ɦiwi㷨пɻһǶȽۻȨiȨwiΪҪѧϰıCMACҲתΪi.ʼȨΪ㣬ÿsiΪ

(3)
ѧϰĿʹf(si)=diдɾʽ
CD
(4)
A=a1 a2apT
AW=D
(5)
ͨȨѧϰһַѵΪs1,d1),(s2,d2),(sp,dp).һѧϰÿһȨ.ȶ(s1,d1)ѧϰȨʹĶӦϵf(s1)=d1ʱȨΪw1.Ȼw1Ϊ(s2,d2)ѧϰֱȨֵw2f(s2)=d2.һ£w1w2ʴʱδf(s1)=d1.һѧϰķѧµģ˾ɵġ.Ϊ˿˷ȱ÷ѭѧϰֻ.ǽѧϰķȨи£һһ룬ΪһܵȨַԿ˷ȱ㣬нϿٶ.
Albus˲δCMAC㷨Ե֤Wong3ݺParks4ݴԷĽǶȣWong5ݻƵĽǶȲAlbus㷨.ǵоһѧϰģѧϰδо.ѧϰһЩָ̽Albus㷨һѧϰʱãѧϰʱ׷ɢȱݣӦĸĽ㷨.
4ѧϰʱAlbus㷨ȱݼĽ㷨
1.ڶĿеһѵռ䱻󲻴ͬѵͬһԪAlbus㷨ѧϰʱΪ
֤.Albus㷨ѧϰʱۻȨʽΪ

(6)


(7)
ΪռΪһάѵֵڣ0RϣԪΪR+c-1.
*ʱʽ(6)

(8)


(9)
eΪһ޳N1УN1R+2c-2.أeдFourierʽ

(10)
(10)(9)(7)ɵ

(11)


(12)
cһȡΪ32256sin(c)/cɺ.ء2У

(13)
ҽH(ej)1ʱ㷨һƵ(2)ڣH(ej)1

(14)
=2ʱءKer(A)Ӱwͦ.Albus㷨ѧϰʱΪ
1.ڶĿеһѵѵֵڣ0RϣҰΪc.ռ䱻󲻴ͬѵͬһԪAlbus㷨ѧϰʱΪ

֤.ʽ(14)


(15)
Fourier(10)֪kȡֵΪ0,N,-1ݼAlbus㷨ѧϰʱΪkںȡֵʱ㷨.N1Խ󣬼ȡֵR͸ҰcԽ󣬲ĵԽ.
Ϊ˿˷Albus㷨ѧϰ̫ʺϵȱݣһָĽCMAC㷨ʽ(1)޸Ϊ

(16)
qΪѧϰһÿȨƽµĴ.
2.ڶĿеһѵռ䱻󲻴ͬѵͬһԪĽCMAC㷨ѧϰʱΪ2ctg-1qأ2-ctg-1q).
֤.ĽCMAC㷨ȨиʱӦ㷨֤еĹʽ(13)Ϊ

(17)
ҽH(ej)1ʱ㷨һƵ(2УڣH(ej)1
2ctg1qأ2(-ctg-1q).
(18)
=2ʱءKer(A)Ӱwͦ.ԸĽCMAC㷨ѧϰʱΪ2ctg-1qأ2(-ctg-1q).
.CMACеľ򷺻ÿʹӦֵΧc֪ͬʱӦأһѧϰһÿȨҲֹһα.minq=2.qȡСֵ2ʱ(18)øĽ㷨Ϊ0.93أ5.35ȣĽ㷨ΧAlbus㷨(14)qԽĽ㷨ΧԽ.
2.ڶĿеһѵѵֵڣ0RϣҰΪc.ռ䱻󲻴ͬѵͬһԪĽCMAC㷨ѧϰʱΪ
֤.ʽ(18)ɵøĽ㷨ѧϰʱΪ

(19)
.(19)һչΪ

(20)


(21)
qȡСֵ2ʱctg-1q=0.46ȣк0Уǵݼģctg-1q0.46Ӷb10,b20.ɼĽ㷨ΧAlbus㷨.qԽb1b2Խ󣬸Ľ㷨ΧҲԽ.
ĽCMAC㷨һѧϰҲõģ֤ɲμף4ݣۻȨJocobiԷ.
5
ĳCMACʵһάԺӳ䣬

(22)
siʣ0180ݣֱΪ1ѵֵڣ0RϣR=180.CMACĸҰc=32,ѧϰһÿȨƽµĴqȡΪ6.N1=242ctg-1q=0.17.s=5ļȡ㣬ʽ(22)ۼ㣬ɵ37ݣЩݷֱAlbus㷨ͱĵĸĽ㷨CMACѵ.Ϊʾķʱ˲ͬѵʱݦs=6.ʽ(15)Albus㷨ѵʱΪ60k182;ʽ(19)֪12k230ʱĵĸĽ㷨ѵʱ׼Ľ㷨Albus㷨.
ͼ2ʾΪһѧϰʱߣ㷨ȡýϺõЧĸĽٶȸһЩ.ͼ3ѧϰʱߣɸͼɼAlbus㷨׵·ɢñĵĸĽ㷨ԻϿЧ.


ͼ2һѧϰ


ͼ3ѧϰ
6
ڶAlbus㷨оĻϣʺѧϰĸĽ㷨֤˸㷨ԣ˸㷨һѧϰʱнϺõԺͿ.
1)ѧعԱл.
߼:ۡ19681996ϺͨѧԶۼӦרҵʿѧλ.ΪϺͨѧԶϵʦ.ҪоȤܿƣѧϰƼڿеӦ.
1957ϺͨѧԶϵڡʿʦ.ҪоȤܿƺ͸ӹҵϵͳԤƵȷо.
ߵλ:ϺͨѧԶϵϺ200030
ο
1ݡAlbus J S. Data storage in the cerebellar model articulation controller (CM.AC). Trans. ASME, J.Dyn. Syst. Meas. Contr., 1975,97:228233.
2ݡAlbus J S. A new approach to manipulator control: The cerebellar model articulation Controller (CMAC). Trans. ASME, J.Dyn. Syst. Meas. Contr., 1975, 97:220227.
3ݡWong Y F,Sideris A. Learning convergence in the cerebellar model articulation controller .IEEE Trans. on Neural Networks, 1992, 3(1):115121.
4ݡParks P C, Militzer J. Convergence properties of associative memory storage for learning control system. Automation and Remote Control, 1989, 50(2):254286.
5ݡWong Y F. CMAC learning is governed by a single parameter. IEEE Int. Conf. on NN, 1993, 14391443.
ոڡ1995-04-10
