Զѧ
ACTA AUTOMATICA SINICA
1998 24 1  Vol.24  No.1 1998



ǰѧϰ㷨
乤ҵӦ

ؼʡǰ磬ѧϰ㷨߻ѻװ. 
A NOVEL RANDOM LEARNING ALGORITHM FOR MULTILAYER
NEURAL NETWORKS AND ITS APPLICATION
DAI LIANKUIMA LONGHUALI XIAODONG
(Institute of Industrial Process Control,Zhejiang University,Hangzhou 310027)
Key wordsMultilayer networks,learning algorithm,fluidized catalytic cracking units.
1
ิӵĹҵ̽ģ⣬ϵͳʶ뺯ƽȣתΪķ.
ϣx(k),y(k)k=1,MϺ
(1)
ʹ
(2)
Сx(k)=(x1(k),xn(k))TΪMΪy(k)R1Ϊʵf()ΪϺ(k)ΪӦֵ.
ԪΪһ͵ķϵͳͨԼ򵥷Ժĸʵָӵĺģͽṹ򵥡ѧϰǿص㣬Ŀǰѹ㷺Ӧڸֹҵ.Զǰ練򴫲ѧϰ㷨ڵĲ㣬һµĶǰѧϰ㷨Ӧĳ߻ѻװô͸ɵʵʱ.ҵӦýѧϰ㷨нϿѧϰٶȺͽϸߵϾȣܹѧϰֲС㣬ɹ㷺ڹҵ̵Ľģ.
2ǰѧϰ㷨
򴫲ѧϰ㷨1ݽ˶ǰѧϰ⣬ʹǰΪĿǰӦ㷺֮һҪ⣺
1)ѧϰ㷨ٶȹ
2)Ȩϵ϶࣬ͨݶ½Ż㷨ֲС㣬ʹ޷һ.
Ϊˣģ2ö̬滮˼ѵǰķWitts˻Ż˼ǰѧϰ㷨3.ĽĽṹص㣬Ӧѧϰ㷨.
ǰĴڶ4֤һ任Ķ뵥ƽнϵһf(x),ȼ·Ż.
(3)
СV=(v0,v1,vm)TΪȨϵ
W=wiji=1,,m;j=0,1,,nΪȨϵ
m Ϊڵ nΪڵ(k)Ϊֵ
hi(k),zi(k)ֱΪiڵi=1,m;
g(x)ΪSigmoid
Դľص㣺wijһvi(i=0,1,,m)Ĺ͵С⣬ɲõС㷨⣻ȨϵɲѰŷãɴѧϰ㷨.
ǰѧϰ㷨
һ.һ
(4)
 xi,max,xi,minΪxi(k)k=1,,MСֵi=1,n;ymax,yminΪy(k)k=1,MСֵ.
ڶ.ţqmaxڵm,q=0;ʼ EoptȨϵwijֵwoptij
wij=t,woptij=wij,i=1,,m,j=0,1,,n,
ЦtΪ(-11)֮ľȷֲ.
.ÿԪ

Ĳ.ڣz1(k),zm(k);y(k)k=1,MõС㷨Ȩϵvii=0,1,,m.
岽.

.EEopt,q=0.Eţת.Ϊ N(02)ĸ˹ֲΪƽ.
߲.EEopt,qqmax,̽wijת.ЦwijΪN(02)ĸ˹ֲ.
3ҵӦá߻ѻװô͸ɵʵʱƣ5
ĳҵ߻ѻװõֳɼݣǰԪѧϰ㷨˴͸ɵĹģͣʵ˸ָԤ.ڹҵֳݱģ;нϸߵľȣҿӦı仯ԭʵĲ.ıջԤϵͳõĵǿ³ԣϵͳ˷ƽԣʹ͸ɵĲΧ٣ʹĲڽϼѵˮ׼ϣԵľЧ.
ҵӦýڸѧϰ㷨˷˷򴫲ѧϰ㷨ٶֲСĲ㣬ʹöǰʵʱƳΪ.ͬʱڸΪԺʱΪʵʱԣɲõС㷨Ȩϵ͸ǰĹҵӦô˱.
ߵλ㽭ѧҵƼоݡ310027
ο
1ݡRimelhart D E,McClelland J L.Parellel distributed processing.MIT Press,1986.
2ݡSaratchandram P.Dynamic programming approach to optimal weight selection in multilayer neural networks.IEEE Trans.Neural Network,1991,2:465-467.
3ݡWitts M J et al.Artificial neural networks in process engineering.IEEE Processings-D,1991,138:256-266.
4ݡRobert H N.Theory of the backpropagation neural network.In:Proc.of IJCNN,1989,1,593-598.
5ݡ.͸ɵʵʱ.ԶǱ19962323-27.
ոڡ1995-03-31
