Զѧ
ACTA AUTOMATICA SINICA
1998 24 5  Vol.24  No.5 1998



ʱعȶԷ
ᡡƽһ
ؼʡع磬ȶȶ.
STABILITY ANALYSIS FOR CONTINUOUS-TIME
RECURRENT NEURAL NETWORKS
ZHANG LANLINGLIU HEPINGSUN YIKANG
(Department of Automation,Beijing University of Science and Technology, Beijing100083)
Key wordsContinuous-time recurrent neural network, asymptotic stability, absolute stability.
1
عȶԵķǲŵֱӷ1ݣַĹؼŵѡĿǰûϵͳعŵķֱӷȶǱȽѵ.ĻGersgorinһعΪ罥ȶ;ȶĳַͨļжȶ.
2ƽĴԼ任
оʽع
(1)
y=y1(t),y2(t),,yn(t)TRnΪԪ״̬ʸ(.)tanhA=diaga1,a2,,anݡRnnǳֵ(ai0,i=1,,n)i=I1,I2,,InTRnΪԪⲿֵʸW=wi,jnnRnnȨ.
yeΪϵͳ(1)һƽ㣬
ye=A1(Wye)+A1i.(2)
ӳ䪢(y)=A1(Wy)+A1iyΪӳ䪢յһ㣬
y=A1(Wy)+A1i.(3)
Ƚ(2)(3)ʽɼӳ䪢յĲ㼴Ϊϵͳ(1)ƽ.
1.(Τ㶨)2Hn=a,bnΪRnıռf:HnHnһfHnһ.
1.綯̬ϵͳ(1)ⲿiֵAӾWRnйռK=y:y-A1iޡܣA1.TϵͳKһƽ.
֤.ӳ䪢ռTĶɵ
(y)-A1iޡA1(Wy)ޡܣA1.(Wy)ޡܣA1.T.   (4)
ɦ(Wy)ԡKĶ弰1֪Kһ㣬ϵͳ(1)Kһƽ.
ϵͳȶʱԲΪ0ƽ㣬Ҫб任.
ye=y1ey2eyneTΪϵͳ(1)һƽ㣬z=y-yeg(Wz)=(Wy)-(Wye)o=Wz=o1,o2,,onT
=f(z)=-Az+g(Wz)=-Az+g(o)(5)
ϵͳ(5)ƽz=0ϵͳ(1)ƽyeӦ.
3عĽȶ
2.(Gersgoring)3B=bijݡMn(MnΪnnĸ󼯺)ʾBĸȥľкͣBֵsλnԲ̵ĲУ(CΪ).bij+Ri(B)1(1in)BֵsλڵλԲ.
3.4ݶڷϵͳf(0)=0f΢D=Dֵиʵƽ0ǽȶ.
2.綯̬ϵͳ(1)yeһƽ㣬Ȩƽyeȶ.
֤.ȿ綯̬ϵͳ(5)ſɱȾJΪ

r(s)Ϊֵsʵ2

3Ӧϵ֪ϵͳ(1)ƽYeȶ.
4عľȶ
11.ϵͳ(1)ⲿiȷĳֵAӾWϵͳֻΨһĽȶƽ㣬ϵͳȶ.
4.5Hn=a,bnΪRnıռf:HnHnһf(x)ſɱȾֵڵλԲڣfѹӳ䣬ʱf(x)HnֻΨһĲ.
3.ϵͳ(1)ⲿIֵAӾWKĶͬ1yeΪϵͳһƽ㣬yeK.
(3)ʽKĶ֪۳.
4.ϵͳ(1)ⲿiȷĳֵAȨ㣬ϵͳֻΨһһƽ.
֤.ӳ䪢յſɱȾLΪ


Lii+Ri(L)1(i=1,2,n).
24ê(y)Kֻһ㣬ϵͳ(1)Kֻһƽ㣬϶3֪ϵͳֻΨһһƽ.
5.ϵͳ(1)ⲿiȷĳֵAȨ㣬ϵͳȶ.
ɶ241֤۳.
5
ĲGersgorinع罥ȶȶĳ.÷ͨȨм򵥼жȶԣһͨõļ򵥡ķʵӦüֵ.
ߵλ(ƼѧԶϵ100083)
ο
1Matsuoka K. Stability conditions for nonlinear continuous neural networks with asymmetric connection weights.Neural Networks19925(3)495500
2Devaney R L. An Introduction to Chaotic Dynamic Systems. Addison-Wesley publishing company, 1984
3϶R AԼѷ C A. .ѧ磬1989
4Vidyasagar M. Nonlinear Systems Analysis.Englewood Cliffs. Newjersey: Prentice-Hall Inc, 1978
5Jin L, Nikiforuk P N, Gapta M M. Absolute stability conditions for discrete-time recurrent neural networks. IEEE Trans. Neural Networks, 1994,5(6): 954963
ոڡ19960926
