о뷢չ
JOURNAL OF COMPUTER RESEARCH AND DEVELOPMENT
1999Vol.36No.7P.794-799




㡡졡ƽ
ժҪоŷǳҪ.ձΪ:ֻҪĽ½,ĵΪĵƽ̬,½,ȶƽ̬ΪļС.ȡ˵ܱ֤ȶԣȡ˵ʹȶҲܱ֤ƽ̬ΪļС.֤ȶĳֱҪĽн.ƽ̬ļСĹϵ.һķ.ΪӦ,ϸ֤HopfieldԣһڼʵԳƾֵӦȫ.
ؼʣ磬  ȶ
ţTP18
ON THE ENERGY FUNCTIONS OF NEURAL NETWORKS
ZHANG Yi
(Institute of Computer Science and Engineering, University of Electronic Science and 
Technology of China, Chengdu 610054)
ZHOU Ming-Tian
(Institute of Computer Science and Engineering, University of Electronic Science and 
Technology of China, Chengdu 610054)
WANG Ping-An
(Department of Computer Science and Engineering, The Chinese University of Hong Kong, Hong Kong)
AbstractEnergy functions are used widely in the study of neural networks. It is known that if a neural network has a decreasing energy function bounded below and the equilibrium points of the network are identical with the zero points of the derivative of the energy function, then the network is stable and the equilibrium points are the local minimum points of the energy function. Examples are given in this paper to show that these results are actually not correct. It is proved under the above conditions for the energy functions that a neural network is stable if and only if all the solutions of the network are bounded. Relationships between the equilibrium points of neural networks and the local minimum points of energy functions are discussed. As applications, the stability of Hopfield neural network is proved rigorously and a neural network is given for finding out all eigenvectors of real symmetric matrix corresponding to the largest eigenvalue.
Key wordsneural networks, energy functions, stability
1
ѧHopfield80ķбȶԣ13. HopfieldеһHamilton,ɹعһбHopfieldȶԣͨоŻ,˺ܴɾ.νȶָκεطĽƽ̬еĳ.Lyapunovµĳȶ.HopfieldԺ,ѧķо.ձΪ19ȶԭ:(1) Ľ½; (2) ĵΪĵƽ̬; (3) ½.Щѧ߻Щ.ף4ݡ7.ʵ,һѧǲ.һά
(1)
.׼,ʽ(1)ƽֻ̬0.

Ȼ,E(x)½磬

Լ

,x00,ʽ(1)x0Ľx(t,x0)=x0et+.ɼʽ(1)ȶ.
ף73оŻ.ͨHopfield繹.

ƽ̬һ.,ƽ̬,Ȼ̸ȶ.
۱ǰǲбȶԵ.,Ӧõ·ģʱʹΣ.,ΪʲôHopfieldȶ?ΪHopfield⻹ص,һ.ӦŻ֤Ҫ.ͨLyapunov бϵͳȶģף10ݶԴо.ķҪ.ĵĿľҪһķ.ǽбȶԵ,ƽ̬ǷӦСһ.ΪӦ,ϸ֤HopfieldȶԲһڼʵԳƾֵӦȫ.
2ȶԷ
΢ַ
(2)
.xRn,f㹻⻬Ա֤ڣ0,+)ϵĴΨһ.x0Rn,x(t,x0)ʾʽ(2)Ĵx0Ľ.ǳ(2)ȶ,ƽ̬Ƿǿյһй߾ƽ̬еĵ.
1. ERnRʹöһt0,

(2)ȶĳֱҪΪһнн.
֤. Ҫ.ȶ,ƽ̬Ϸǿÿһƽ̬еĵ.,ÿһн.
.x0Rnx(t,x0)н.ǣĳнռDʹöһt0,x(t,x0)λD.(x0)x(t,x0)Ħ-޼(x0)ǿҦ(x0)D.E(x(t,x0))½.ֶһt0,x(t,x0)λDУǣ.ȡyʦ(x0)x(t,y)(2)yĹߣһt0x(t,y)ʦ(x0).ɺEE(x(t,y))=E0һt0.Ӷһt0Уһt0.˵x(t,y)yƽ̬.ǣƽ̬Ϸǿ.ɦ(x0)֪ͨ(x0)ʵֻһyӶ.(2)ȶ.
֤.
Ľϸ֤Hopfieldȶ.
оµHopfield
(3)
T=[Tij]nnȨϵCi>0,Ri>0IiΪgi(.)Ϊ΢ϸ񵥵н纯
ǳΪ(3)ƽ̬

2. ȨϵTǶԳƵģ(3)ȶ.
֤. ֤һнн.ʽ(3)

ǣһt0

Hopfield

(4)
.㺯E(3)Ĺߵĵ


Ȼ

ɶ1֪(3)ƽ̬Ϸǿһнƽ̬еĵ.ǣȶ.
֤.
3. ڶԽԪΪĶԽǾ=diag(1,2,,n)ʹþTǶԳƵģ(3)ȶ.
֤. (3)дΪµȼʽ

ɶ2.
1. ֮һ(3)ȶ.
 n=1;
 n=2T12=T21=0T12=T210;
 n=3,T31T23T12=T13T21T32Tij=Tji=0TijTji0,(ij;i,j=1,2,3).
3ƽ̬С
һHopfield



3ƽ̬0,ln2-ln2.ɶ23֪,ȶ.ҪעȶԲζÿƽ̬Ե.ʵ,ƽ̬0ǲȶ.ӿԲμף10.йƽ̬ȶԷԲμף11ݡ14.ʽ(4),HopfieldΪ



׼,.,ƽ̬0ľֲ.ͬln2,-ln2ľֲС.
,ʹһȶƽ̬ҲһľֲС.ͷף19Ϊƽ̬ӦֲСĽ.Ķƽ̬ΪֲС.
4. (2)һннҴERnRʹöһt0Լҽ=0.x*(2)һƽ̬.x*EһֲСĳֱҪǴx*ĳСD(x*)ʹöx0D(x*),x(t,x0)x*0(t+)E(x*0)E(x*).
֤. Ҫ.x*(2)һƽ̬EһֲС.۲,xkx*(k=1,2,3,)ʹE(xk)E(x*)x(t,xk)x*k(t+)ԼE(x*k)<E(x*).Ĺ½Ĳע⵽E(xk)E(x*)(k+)شtk0(k+)ʹx(tk,xk)x*(k+)E(x(tk,xk))<E(x*).Ȼʱx*EľֲС.ìܱҪԳ.
.x0D(x*),ɶ1֪(2)ȶ,x(t,x0)(2)ĳƽ̬x.ע⵽E(x(x0,t))½,Ӷһt0

x0D(x*)֪x*EһֲС.
֤.
2. 趨4,(2)һֲƽ̬ΪľֲС.
Żʱ,ΪȤȫС.
5. (2)һннҴERnRʹöһt0Լҽ.x*(2)һƽ̬x*EһȫСĳֱҪǶһƽ̬x;E(x)E(x*).
֤. Ҫ.x*(2)һƽ̬EһȫС.۲,һƽ̬xʹE(x)<E(x*).Ȼ,x*ȫСì.
.x0Rn,ɶ1֪(2)ȶ,x(t,x0)(2)ĳƽ̬x.ע⵽ʱE(x(x0,t))½,Ӷһt0

x0Rn֪x*EһȫС.
֤.
3. 趨5ֻһƽ̬,ƽ̬ΪȫС.
ǹһԽһ.AһnnʵԳƾ, Aֵ.΢ַ
(5)
xRn.AΪȨǿ,xRnΪ״̬,ʽ(5)һ͵ȫ˹.ɵ·֪ʶ֪,ʽ(5)ɵ·ģ.A-Iֵ,ʽ(5)Ľн.



ݶ1,(5)ȶ.׼,ƽ̬ǡAĶӦ˵ռV.x*V,E(x*)=0.ɶ5֪VеĵE(x)ȫС.AĶӦ˵,ʿڼAĶӦ˵ȫ.ۿԿ,ȡȫС㲻һҪֻһƽ̬.
4
Ľһķ.ǸӦбȶһϸ׼ƽ̬ǷӦС.ΪӦ,ϸ֤HopfieldȶԲһڼʵԳƾֵӦȫ.еĽ˹Ȼָ.
ĿùȻѧ(Ŀ69871005)Ĵʡѧ.
߼飺,У19634,ʿ,оΪ˹硢ܼݡھ򡢷΢ϵͳ.
,У19393,ʿʦ,оΪ硢˹.
ƽ,У196111,ʿʦ, оΪʵҽѧеӦá ѧӻ ͼѧ .
ߵλ(ӿƼѧѧ빤ѧԺɶ610054)
(ӿƼѧѧ빤ѧԺɶ610054)
ƽĴѧѧ빤ѧϵۣ
οף
1Hopfield J J. Neural networks and physical systems with emergent collective computational abilities. Proceedings of National Academic Science, 1982,79(7): 25542558
2Hopfield J J. Neurons with graded response have collective computational properties like those of two-state neurons. Proceedings of National Academic Science, 1984,81(10): 30883092
3Hopfield J J, Tank D W. Simple neural optimization networks: An A/D converter, signal decision circuit, and a linear programming circuit, IEEE Trans on CAS, 1986, 33(5): 533541
4޷, ܴ. СֵӦʸ. ѧ, 1994, 22(4): 1319
(Luo Falong, Li Yanda. Neural network approach to computing the eigenvectors corresponding to the largest and smallest eigenvalues of a positive matrix, Acta Electronica Sinica(in Chinese), 1994, 22(4): 1319)
5޷, ܴ, ʵʱTLS. йѧ(A), 1993, 23(6): 666672
(Luo Falong Li Yanda. Solving the TLS problem by neural networks. Science in China (Series A)(in Chinese), 1993, 23(6): 666672)
6޷, . һֶάģͼӦ. ѧ, 1992, 20(10): 3338
(Luo Falong Bao Zheng. A 2-D neural network model with the applications. Acta Electronica Sinica(in Chinese), 1992, 20(10): 3338)
7, , ˵. ŻеһЩо. ѧ, 1992, 20(10): 9499
(Liu Jun, Wang Zhaoming Gu Deren. Research on the properties of the neural optimization, Acta Electronica Sinica(in Chinese), 1992, 20(10): 9499)
8о, ֣. ˹. : ߵȽ, 1992
(Yang Xingjun Zheng Junli. Artificial Neural Networks(in Chinese). Beijing: Advanced Education Publishing House, 1992)
9. ˹ģͼӦ. Ϻѧ, 1993
(Zhang Liming, Models and Applications of Artificial Neural Networks(in Chinese). Shanghai: Fudan University Publishing House, 1993)
10Yang H, Dillon T S. Exponential stability and oscillation of Hopfield graded response neural networks. IEEE Trans on Neural Networks, 1994, 5(5): 719729
11Zhang Yi. Global exponential stability and periodic solutions of delay Hopfield neural networks. International Journal of System Science, 1996, 27(2): 227231
12Zhang Yi, Zhong Shouming, Li Zhengliang. Periodic solutions and global stability of delay Hopfield neural networks. International Journal of System Science, 1996, 27(9): 895901
13, , . ʱȶ. Ӧ, 1998, 15(2): 197220
(Zhang Yi, Zhong Shouming, Wang Li. Global stability of neural networks with infinite delay. Control Theory and Applications(in Chinese), 1998, 15(2): 197220)
14. ʱ͵˫ĶԷ. о뷢չ, 1999, 36(2): 150155
(Zhang Yi. Qualitative analysis of bidirectional associative memory neural networks with delays. Journal of Computer Research and Development(in Chinese), 1999, 36(2): 150155
ոڣ1998-10-30
޸ڣ1999-04-12
