Զѧ
ACTA AUTOMATICA SINICA
1997 23 1 Vol.23 No.1 1997



Ŀѧϰ㷨о
־»
ժҪһڵ(SNN)ѵӱĹSNNµĿѧϰ㷨(FLA). һһֹϵͳʶģͣFLAԽ۷. ʵиFLAKarayiannisLFAиٶ. 
ؼʡ磬BP㷨ͳƷ.
STUDY ON THE FAST LEARNING ALGORITHM
OF SINGLE-LAYER NEURAL NETWORKS
FENG ZHAOZHI HUANG ZAILU YANG SHUZI
(Dept. of Electronic & Information Eng., Huazhong University of science and Technology, Wuhan 430074)
Abstract This paper proposes a new generalized criterion for the training of single-layer neural networks, which leads to a novel fast learning algorithm for single-layer neural network. In order to analyse the convergent properties of the fast algorithm we developed, a new generalized system identificaton model is also presented. Experiment results show that the fast algorithm proposed in this paper performs the training of neural nework faster than the corresponding learning algorithm given by Karayiannis.
Key words Single-layer NN, fast BP algorithm, statistical analysis.
Karayiannis1˵(SNN)һϵпѧϰ㷨(FLA)ϺõظSNN. ȻKarayiannisûԪ״̬ΪҲδFLA. ȻBershad2ݵSNNBP㷨ȡͻƣϵͳʶģͲڶFLA۷. Щ⣬ĽͼSNNFLAԷһµ̽. 
1SNNĿѧϰ㷨
(SNN)Ľṹͼ1ʾ. ͼУNinNoutֱԪwijjԪiԪ֮Ȩ.x*k=[x0,kx1,kxNin,k,yk=[y1,ky2,kyNout,k]ֱΪSNNk=1,2,,NpNpΪģʽx0,k=1.0SNNĻ̵ϵͳѧΪ
(1)
ʽw*i=[wi0wi1wiNin];f(x)=sigmoid(x);*Ϊת.
ģʽx*kѵĿģʽtk[t-1,k t2,ktNout,k],SNNѧϰ̾ǵ޸ȨֵӶʹ()ֵΪС. Ϊ˼SNṆµĹ
(2)

һ̡(0,1).
ݶ½ƵSNNµĿѧϰ㷨ĵݹ㹫ʽ. Ȩֵw(k)ijͷУźŦi,k()ѧʽΪ

(3)
(4)
Ȼڹʽ(4)У(ti,k-yi,k)ΪiԪ(1-y2i,k)Ϊͳĵ. ˣΪFLA(1-y2i,k)KarayiannisFLAеĵ(1-yi,k)yi,k(f(x)=sigmoid(x)). һǵʽ(1-yi,k)yi,k1-y2i,k,exp(-y2i,k)cos(yi,k)ǿɵõSNNһg(yi,k)={(1-yi,k)yi,k,1-y2i,k,exp(-y2i,k),cos(yi,k)}. SNNµĿѧϰ㷨ϸ£


ͼ1ṹ
Algorithm Fast BP algorithm
Initialize Randomly W
Select E0,,,,;=1;num=0
: num=num+1; k=0;E=0
: k=k+1 
yi,k=
i,k()=g(yi,k)(ti,k-yi,k)
wijwij+Ǧi,k()xj,k
yi,k=

if km; then: go to 
=exp(-/E2); if EE0; then: go to  End
2㷨Է
Ϊ˶FLAзƹBershad˵ϵͳʶģͣ˹ϵͳʶģ(ͼ2). ͼ2УX(n)Ǹ˹FW(n)ֱϵͳʶģȨֵSNNȨֵe1(n),e(n)ֱǹϵͳʶģ͵ڲϵͳf(x),(x)ΪԺ. Ϊ۷㣬ȡf(x)=(йطŵĶף2ݵĶͬԺ˵).


ͼ2Ӧģ͵ĵ
(ϵͳʶģ)
SNNFLAĵ㹫ʽΪ



(5)
ף2ݣƵȨֵW(n+1)ѧ



W(n)ƽȨֵΪ,


(6)
Ц=(E)=exp(-/E2). ʽϺõSNNȨķ. (0)=0()ʱ

һε֮(n)űƽFķ仯Ӷѵǳɹ.
ڹʽ(6)УnNָұƽ1. ȻڴͳEBP㷨У2ݣ=Եرƽ1. ˣΪFLAȴͳEBP㷨ٶȿĻ.
3ʵ
ñĿѧϰ㷨ǳɹ˶ĳ;ڽʽṹģʽʵ. ʵвõΪNin=120,Nout=4=0.3,=0.1. ĳʼȨֵΪ(-0.5,0.5)֮ѵE0=0.01. ȡڽʵĳزźŵ120Ϊһζ;ڽʵĲģʽΪѵģʽڶλȡĽʽṹΪʶģʽ. Ը;ľڽʽṹЧܴﵽ100%ȷ. Ӧg(y) (g(y)={(1-y)y,1-y2,e-y2,cos(y)})Ӧѧϰͼ3ʾ. ȻµĿѧϰ㷨бKarayiannisBP㷨ٶȣCPUʱҲ. ʵ˵ ѧϰ㷨ЧĺͿе. 


ͼ3ʯͲ⾮ӦеӦѧϰ
*Karayiannis FBP
ߵλѧϢϵ人430074
ο
[1]Karayiannis N B, Venetsanopoulos A N. Fast learning algorithm for neural networks. IEEE Trans. Cas-, 1992, 39(7)453474.
[2]Bershad N J, Shynk J J, Feintuch P L. Statistical analysis of the single-layer backpropagation algorithm:Part -mean weight behavior. IEEE Trans. on Signal Processing, 1993, 41 (2)573582.
ոڡ1996-07-22


