Զѧ
ACTA AUTOMATICA SINICA
1997ꡡ234ڡVol.23No.41997



񾭼任
ģ(CMAC)ķ1
ŷ»ܡƼ
ժҪ񾭼ķһǳҪ.ļСģ(CMACCerebellar Model Areiculation Controller)ԭѧϰ㷨÷淽ڻʹõ任ϵ(ֱֵֵĹؽڽǶ)CMACķܣΪ1100ʱCMAC.ϵͳľҪǽһȴܵ.ĻӰ쾫ȵĸؼܵĸĽ.
ؼʡܣСģ(CMAC)任.
GENERALIZATION OF NEURAL NETWORK MODEL(CMAC) FOR
COORDINATE TRANSFORMATION IN NEURAL COMPUTATION
OUYANG KAI CHEN HUI ZHOU PING ZHOU CHEN
(Dept. of BME., Capital University of Medical Science, Beijing 100054)
AbstractGeneralization of neural network is a very important topic for coordinate transformation in neural computation. In this paper, we describe the principle of Cerebellar Model Articulation Controller(CMAC) including its learning algorithm, and discusse the generalization of CMAC through simulation of coordinate transformation (the input is position coordinate values and the output is articulation degrees of robot). The CMAC may still run well at generalization rate 1100. Several factors affecting the accuracy are also discussed.
Key words Generalization, coordinate transformation, cerebellar model articulation controller (CMAC)
1
MarrӾ۳񾭿ѧѾΪ񾭿ѧһҪ֧.Ȼ񾭿ѧоĽֻϵͳϢӹһֿԣҪʵ֤Ŀǰϵͳṩ֪ʶȫ£ַԲʧΪһֿѧ̽رĳЩƲѳΪʵоĳϢѧһչ¹۵㡢·ԴȪ.˿ѧҵļעΪǰѧǰ֮һ.
񾭼任ǳһ.ѧϵͳǲģ緽ƽеɾͲô.ԭ任ʱҪԸѵҪȫƽ϶еĵ㶼ѵʹ֮һķֱ´ﵽȣѵ൱.оӸϵͳе񾭼ʱתŵĤ밼ӳİת̶ϵУѵﵽ5184ֱʽΪ1/2562).һҪ⡪⣬ýٵѵʹڸڴﵽָľ.
зоAmari1ݣΪǽ뼯ĸӳ䵽ӳ(Ӧ)ĳһ.ɴ˿ԿɾȾ⣬ȡӳ䷽ʽ.Զ֪ķǼ޵ģʵҲ֤ʵһ.
оһַ磬СԵ񾭽ģAlbus֮ΪCMAC(Cerebellar Model Articulation Controller)2.ķӳϵ֪ͬ.ڼ򵥽CMACĹɺо䷺.
2ģͼ
2.1ģ͵Ľṹ
CMAC֮ӳ䷽뼤CԪCԪڴ洢ۼӹ.CԪǰԭģڵ(ֵȽϽӽ)ӦCԪͬĽ϶࣬ͬĽ٣ɴ˿ԿCѡȡԽ󣬶ӳϵӰԽ󣬴ӶγɽϺõķ.
򵥵CMACģͼ1ʾ.Կ뼯SPӳΪ֣SA(AΪȨصĵַռ)AP.


ͼ1򵥵CMACģ
2.2Ԫȷ
2.2.1ݵ
SΪNάģS=(S1,S2,SN),CMACЩҪʹ֮Ϊɢ.磬NάÿһSiΪ1RiΧڵһSi

СSiminSimaxֱΪSiֵСֵSiΪֱ.
2.2.2ӳ䵽洢Aĵַ
ӦȷCе׳֮ΪеĳΪۺ(generalization).ʵϴ໥Ӱ.CֵԽCMACķԽǿѧϰʱƱؼӳ.ȷCֵǺҪģǽϸ۷CĹϵ.
SA ӳ䷽ǶSÿSiֵĴСһ򣬽ӳCⵥԪַVi1Vi2ViCǴSiĵԪ.ЩⵥԪϳC,,(ΪA)Ϊ洢ȨصĵַΪV1iV2iVNiĳ.Ҫע⣺
(1)SiӳⵥԪַVijķ(ƹ)һսûӰ죬һȷӳķӦһµ.
(2)ҪֵԭȷⵥԪĵiS'iSiһλӦVi1Vi1һԪ໥һԪ.
2.2.3Aӳ䵽P
ڴ洢AУÿԪжLȨ(LΪά)CԪP԰ʽ㣺

(1)
2.2.4ѧϰ㷨
2.2.4.1Ȩصȷмලѧϰ
Ϊ(S*P*)CMACΪPһަjpjܦj(j=1,2,L)ӦȨ޸ģʽȨ

(2)
ЦΪѧϰʣtΪѧϰ.˹ʽζȨһξ̯CĵԪϣҲǵ.
2.2.4.2ѧϰ
S1,S2,SMѵ˳ѧϰ.ÿΰ(2)ʽѵһйַԪڶʱͻƻһ.ҪжѭѵԴﵽѵȶõ.
3о
3.1ϵ
񾭼е任к֣ܶһ.Ϊоһԣѡ˽ϸӵģС˶еĶɶȵĹؽڿʾ任.X(x,y,z)ʾ֫ҪﵽĿռλ꣬Ԧ(1N)ʾؽӦתĽǶȣX=f(ȣ.˶Ƶ任֪X(֪X).ռĳΪ(x,y,z),ɶֱ۵ĽǶΪ(123)ֱ۳ΪL

(3)
CMAC任֪(x,y,z), (123).
3.2ȷ
任뼯ӦȫάĿռ䣬ΪzΪƽxy.ʧһԣȡƽڵһ.ֱ۳Ϊ30ֱΪ0.1xyȡ300㣬뼯ڵĸΪ70686(1/4Բڵĵ).ѵȡ13(߼нΪ7.5)ϵĵ㣬ÿȡ60㣬Ϊ1360=780.Щϵĵ㾭һҪͬģʵΪ703.ô703ѵ£Ҫϵͳ70686뼯УʶΪ703706861100.
3.3
ΪоڱĵķоУ任(X,)Ĺϵ(3)ʦźŦ(3)=f-1

(4)
뼯ѵCMACXȣɦȰ(3)ʽX͵ĲΪ1ʾ.
1(xΪ)

Сƽ
2.2120.0010.6050.457
뼯3.6490.0010.6770.577

ԲΪ۲ֻѵ뼯1%ϵͳķ(ͼ2).


ͼ2ѵCMAC
(ע⻬Ϊߣ⻬ΪCMAC)
3.4Ӱ쾫ȵ
ͼ2ԿϵͳһķǱȽϴֲ.Ϊ߾ȣҪоӰأӶҵ߷ȵİ취.
1)ѵ
CMAC֪ͬѧϰ㷨С˷µŻ㣬Բϸ֤.ĽCϴ󡢸нϴ󽻵£ѵڴﵽһֵʱټС1ڴг.
CMACѧϰ㷨нĴԣĿǰнΪϸۣģ34е۲Cϴ󡢷Ϻõ.ݷĽָ䲻0ǿԼСһ̶.
2)
ݱCMACģͼ淽Ϊ¸֣
(1)ʽ(3)(4)ǷԼ㣬ʴһڱѡ˫ȸʱԺ(ԼС106).
(2)ѵʱ.ܶйأCĴСķֲĶӦϵԼѧϰ㷨.
(3).CMACøԪص÷ܵģʵ໥ϵԻȻȫǡʹ˷.ڱ1뼯Ϊԭ.
(4)ʼ.CMACѧϰ㷨ǰųѵѵʱıǰĵԪĵԪֵһѵʱǰ漸ͽϴʼ.ǰ˳ԷΪλXļнǴ0仯90㡢90仯0㹹ɲͬУֲǰϴ(ǰϵУѵʱδﵽѵȵֱռ40.3%48.9%).
Ƕȱ仯45߷ֱﵽ090ʱֵı仯ϴѧϰ񵴣.
(5).CMACģҪȻӳ䣬˼ʹѵ÷ǳãҲ.
3)ÿĵԪCԷӰ
CCMACĻCĴС뷺кܴϵCСϵͣCӷӣѵʱҪӣѵԼʼҲӣҲCԽ󷺻Խ(2).
2ͬϵͳ30(ԲxΪ)

ֵСֵƽֵ
CĴС1520.0010.0763.7120.905
302.8000.0030.8870.756
502.3730.0040.6250.579
801.9840.0050.6960.555
1202.6570.0271.2070.747
7031.9840.0050.6960.555
5532.1780.0080.6460.564
3853.5510.0521.1860.872
ֲ1.9840.0050.6960.555
2.1490.0130.6990.516
Բ1.5320.0180.6610.412

ǵķо
(1)CһֵΧСã籾ʵC5080֮Ϊˣ
(2)ӾϽȷCĴСӰķΧܸǷ㡢ܷӳĹϵΪԭ.2гCֵͬ(C=80)ͬʱϵͳ.
4)ѡ
÷߷ֲ(ҪǵΪǶ)ȻҲֲ緽ֲԲֲֲܷ.ֲϵͳȵӰ2.
4
CMACԽϸӵķϵӳķопԿCMAC϶֪иķһҪġֵӵ.ѧϰ㷨ΪֲڣϵͳӳȽϴֲڣӳľȲߣֻ˵Ͼһķӳ.
ڱ֤CMACٶȺͷǰ£һ侫ȾͳΪCMACоһս.Ϊʦʵʣ¼ĽķⷽĹҲڽ.
1) ϱǷ.׵˶ϵͳ׵еıǾĳ̶ֹķĳһСķΧڸĽϵͳľ.ǵĳʵЧ.
2) СԽCMACģ͵ĸĽ.˵СԶ˶ƾнϸӵĽṹͻɽΪϸġѧϰ˶ƣʦʵʽҪĸĽ.
3)Ľѧϰ㷨Ϣ洢.
4)ӳϵȷĴСֲҲǰѾ֪ʶѧϰЧʩ֮һ.ͬʱCĽṹֵķҲֵо.
1)ܹȻѧ𡢡ƻʵǼƻ֧.
2)֣鳤Ӿ˶֪񾭼о۲ʿģ.йѧԺ1995.52-56.
߼:ŷУ 62꣬ ڣ 1959ҵ廪ѧ.ĿǰǱҽѧѧ£ȫҽѧѧᡢȫѧᡢȫԶѧѧϢƷֿѧίԱҽѧѧҽѧ̡˹͸־ıί.ĿǰеʵǼƻ863ƻ.ר ʮƪ.
¡ܡŮ 29꣬ 1990ҵйѧΪ׶ҽƴѧҽѧϵʦ.ĿǰμʵǼƻ863ƻ⼰һЩĿҪ˹Ӧüϵͳģо.ڹʻϷһƪѧϷƪ.
ߵλ:׶ҽƴѧҽѧϵ100054
ο
1ݡAmari S L Mathematical foundations of neurocomputing. Technical Reports Faculty Engineering University of Tokyo, Japan, 1989.148.
2ݡAlbus J S. A new approach to manipulator control: Cerebellar Model Articulation Controller (CMAC). IEEE. Trans. ASME. 1975.9.220227.
3ݡ.˹ģͼӦ.Ϻѧ磬1993.170184.
4ݡWang Yin fai. A Sideris learning convergence in the Cerebellar Model Articulation Controller (CMAC). IEEE Trans. Neural Network, 1992,3(1):115121.
ոڡ1995-09-29
