Զѧ
ACTA AUTOMATICA SINICA
1997 23 3 vol.23 No.3 1997



֪ľֲ1

ժҪǰȫֻƺ;ֲƵĸص㣬һɶȶֹͬ繹ɵľֲ֪.ԸӷϵͳõѧϰܣѧϰƵֱģͷԶѧϰ.
ؼʡֲ֪ƣCMAC.
A LOCAL DESIGN FOR MULTI-LAYER PERCEPTRON
XU LI
(Department of Electrical Engineering, Zhejiang University, Hangzhou 310027)
AbstractBy syncretizing the global design and local design of feedforward neural networks, this paper proposes a localized multi-layer Perceptron (LMLP), which is constructed by a set of distributively associated subnets with the same structure. Simulation results show that this network has strong learning capacity for complex nonlinear systems, and is applicable for real-time learning control, by using either direct inverse modeling or distal learning strategy.
Key words Multilayer perceptron, local design, CMAC network.
1
ȫֻƺ;ֲǰֲͬƷ.Ϊȫĵʹ֪(MLPֳBP)ȨֵӰ죬ѧϰʺڶԹ̶ѧϰֲCMAC2,3Ϊκʱֻ벿ȨֵһأнϿѧϰٶȣʺʵʱƣ1.һ棬MLPܱƽBorelɲ⺯7ݣھֲڽϴʱѵͨ÷ǳѣ8ݣCMACܽϺõشѧϰľֲ⣬ֻȷԪĵֱֵַӹϵ.CMACÿӿռеźŽСൽһӳ䣬νֱʡ10.ͬʱӳʹԷ򴫲ˣԲԶѧϰƣ6.ϽһֵƷĻͽṹǴڵģʵдо1.
ľֲ֪(Localized Multi-Layer Perceptron, LMLP)ںMLPCMACĸص㣬Ƕף12ĳһ.оڸӷϵͳѧϰеӦ.
2ֲĶ֪
2.1LMLPĽṹԭ
ƵĻ˼ǣN(x1,x2,xN)ɵռSʷΪnӿռ(S1,S2,Sn)ͬӿռòͬÿMLPʽ.n໥Զ(SN1SN2SNn).
ͼ12ʾԭͼУLMLPN뵥ԪmԪMԪ.ÿԪֻܴ״̬֮һ򡰽ֹ.ÿռm*ԪͬĲmԪеλòͬλЭȷ.

ͼ1LMLPԭͼ

ͼ2Эԭͼ
ÿֳRj(j=1,2,N)ͬNnӿռ䣬
n=R1R2RN (1)
ÿӿռӦһ磬ÿӦm*Ԫ(ͼ2m=3).VʾSim(i=1,2,,n)˳ӳλñPijʾSNiеjԪλãPi1,Pi2Pim*VͱʾSNiԪVеλüϣΪSNiV.ͼ2УSN1V=123VSN2V=234VSNnV=mV-2,mV-1,mVVmV VλõԪ.
вͬԪΪ(ͼ2С=1),mVӦ
mV(n-1)+m*.(2)
ȻֵܻǳΪʹʵHĵԪm̫CMACгɹӦõӴձ(hash-coding)5ʵִmVmġൽ١ӳ.ǣSNiԪʵеλãSNiH=Pi1Pi2,PimHӳȷ
SiSNiVSNiHi=1,2,,n.
SNiɣSNiHȷԪͬΪ繫õ㹹.磬ͼ2УSN1H=24m-2HSN2H=4m-2,mHSNnH=35m-2H.ȻSN1SN2жԪSN2SNnһԪ.ͼ2ȷSN1Ԫλʱ.
κʱֻ̣m*ԪЭֻһڹԪֹ.Զûȫһûһȫģһֲ״̬ѧϰķֶ(dichotomization)ͼ(generalization).йزȡֵȡm10 m*m*ֵӾ.
ĺӴӳܲӳͻйⷽģ29.ԪһҪͬƪڴ˲.⣬ѿȡֵֻǦ=1.
2.2LMLPʽ
LMLPMLPʽ￼ּΣ(a)LMLP-IóĶ(MIMO)MLP(b)LMLP-ö뵥(MISO)MLPṹ.
MLPָõֵΪSigmoidṹ㷨н.
LMLP-.һĳMLPȨֱֵ--㣬Խϴ.ͬʱʵϵͳMISOģMIMOϵͳҲöMISOϵͳ.ԣLMLP-ֻһɵȨֵMISOṹ-Ȩֵ.ݶȷοCMACĹԭɵ㷨£
ǰ㣬


j=(y*y)/m*,
wij=wij+jxi
j=j+¦j,

СxiΪiyjΪjԪwijΪӵi뵥Ԫ͵jԪȨֵͦѧϰ.Ȩֵֻһ飬ҸԪԼټ.
Ҫ˵ǣֻоּΣʵϣMLPʽҲͬ.
2.3LMLPѧϰ
ලʽѧϰƲԿɷΪ6ݣֱģͷ411ݺԶѧϰ(Distal Learning)6.öҪѰұض״̬ռ䵽ƿռģ.ֱģͷڡ͹ԡ(Convexity)6ݣضǹڿƿռ䵽״̬ռġൽһӳ䣬ֻܸڿƿռĳƽֵԿܻҲģͣԶѧϰõĳģ.
ֿƷУLMLPӦûκҪӦLMLPģ411еCMACģ6еBP.ʩ(1)һ򵥵ı(2)뼫ֵƣĳΧʱͲȡСĿ.
ñضǻCSTR11ݣٻѧģΪ

СDa=3.0;=1.5;=22.9;H=2.55;״̬x1 x2ֱΪӦתʺͷӦ¶ȣƱu1u2ֱȴ¶ȣu1̶ͨΪ1.0x1ӳ˲Ʒ.
3
CSTRϵͳΪоLMLPѧϰеӦ.ٵĿϡ·ֱֵΪ0.9650.935ķ.TΪ20룬ѧϰΪ200ڣϵΪ80ֵƵֵΪ0.01(x1,x2,x1,u2)ֱ𻮷Ϊ(601444).(m*m)ͳһѡΪ(30200).ȨֵʼΪСΪյMLP30Ԫ.
3.1ֱģͷ
(x1,x2,x1)ɵĿռ䱻Ϊ3360(60144)ӿռ.ͼ3ʾΪǰ5ѧϰڵĸ.ڡ͹ԡ⣬CMACϵͳڵġѧϰ·ֵӦƫѾϵֵƵҲLMLP-.LMLP-ϵͳӵڶѧϰһֱֳõĸ.

ͼ3ֱģͷ
3.2Զѧϰ
ǰֱΪ(x1,x2,x1)(x1,x2,u2)ֳ3360ӿռ.ͼ4Զѧϰܺܺõؽ͹ԡ⣬LMLPʽĿϵͳΪ𲽸Ƶĸ.лLMLP-ϵͳӵĸȫϣLMLP-.ͼ5ڳMLPĿϵͳڸٷʱ.

ͼ4Զѧϰ

ͼ5MLPѧϰ
4
LMLPCMACȷԪַķȷԪеλãMLPʽ.ԣԸӿռڵźžʵ֡һһӳ.LMLP-ʵַӳ䣬LMLP-ΪԱƽС.ϽLMLP-ķԱƽǿLMLP-򣬵ǣڸӿռԻԹϵʱLMLP-ѧϰǿLMLP-.
MLPȫƣʹ䶯̬Ӧѧϰ.CMACֿѧϰľֲ⣬нǿĶ̬Ӧѧϰֱ.ڲƵԭLMLPCMACΪƵĶ̬ӦLMLPġһһӳʹиǿѧϰ.LMLPĽṹ͹ʽһо.
1)Ȼѧ͹ίԤл.
߼飺1964ֱ19861989㽭ѧϵѧʿ˶ʿѧλΪ㽭ѧϵڡʿ.Ҫоרϵͳģơѧϰ.
ߵλ㽭ѧϵݡ310027
ο
1Werbos P J. Neurocontrol and elastic fuzzy logic:capabilities, concept, and applications. IEEE Trans. Industr. Electronics, 1993, 40(2):170-180.
2Albus J S. A new approach to the manipulator control:The cerebellar model articulation controller. Trans. ASME, J. Dyn. Syst. Meas. Contr., 1975, 97:220-227.
3Albus J S. Data storage in the cerebellar model articulation controller. Trans. ASME, J. Dyn. Sys. Meas. Contr., 1975, 97:228-233.
4Miller ,W T. Real-time application of neural networks for sensor-based control of robots with vision. IEEE Trans. SMC, 1989, 19(4):825-831.
5Knuth D. Sorting and searching. The Art of Computer Programming, Addison-Wesley, henlo Park, Calif., 1973, (3):506.
6Jordan M I, Rumelhart D E. Forward models:supervised learning with a distal teacher. Cognitive Sci., 1992, 16:307-354
7Hornik K. Multilayer feedforward networks are universal approximators. Neural Networks, 1989, (2):359-366.
8Jacobs R A, Jordan M I. Learning piecewise control strategies in a modular neural network architecture. IEEE Trans SMC, 1993, 23(2):337-345.
9Wong Y F, Sideris A. Learning convergence in the cerebellar model articulation controller. IEEE Trans. Neural Networks, 1992, 3(1):115-121.
10Chow M Y, Menozzi A. A self-organized CMAC controller. Proc. IEEE Int'l Conf. on Industr. Tech., Guangzhou, China, Dec., 1994, 68-72.
11Xu L, Jiang J P, Zhu J. Supervised learning control of a nonlinear polymerisation reactor using the CMAC neural network for knowledge storage. IEE Proc.-Control Theory Appl., 1994, 141(1):33-38.
12.һ־ֲķ򴫲.ߣ199510(2)148-152.
ոڡ1994-06-06
