(1.重慶教育學院 計算機與現(xiàn)代教育技術系, 重慶 400067; 2.重慶大學 自動化學院, 重慶 400044)
摘 要:
采用It’s微分公式和不等式分析技巧,研究了一類不確定隨機離散分布時滯神經(jīng)網(wǎng)絡的魯棒穩(wěn)定性問題。該模型同時考慮了神經(jīng)網(wǎng)絡模型的兩種擾動因素,即隨機擾動與不確定性擾動。通過構造適當?shù)腖yapunov泛函,以線性矩陣不等式形式給出了系統(tǒng)在均方根意義下的全局魯棒穩(wěn)定性判據(jù),能夠利用LMI工具箱很容易地進行檢驗。此外,仿真結果進一步證明了結論的有效性。
關鍵詞:魯棒穩(wěn)定性; 隨機神經(jīng)網(wǎng)絡; 分布時滯; 不確定性; 線性矩陣不等式
中圖分類號:TP183文獻標志碼:A
文章編號:10013695(2009)04122203
Robust stability of uncertain stochastic neural networks with discrete and distributed delays
FENG Wei1,2, ZHANG Wei1, WU Haixia1
(1.Dept. of Computer Modern Education Technology, Chongqing Education College, Chongqing 400067, China;
2. College of Automation, Chongqing University, Chongqing 400044, China)
Abstract:
By It’s differential formula and combining the method of inequality analysis, this paper investigated the problem of stochastic asymptotical stability of a class of uncertain stochastic neural networks with discrete and distributed delays. There were two kinds of disturbances which were unavoidable to be considered in neural networks. The parameter uncertainties are timevarying and normbounded. The timedelay factors are unknown and timevarying with known bounds. Based on LyapunovKrasovskii functional and stochastic analysis approaches, presented some new stability criteria in terms of linear matrix inequalities (LMIs) to guarantee the delayed neural network to be robustly stable in the mean square for all admissible uncertainties. It gave a numerical examples to demonstrate the usefulness of the proposed asymptotical stability criteria.
Key words:robust stability; stochastic neural networks; distributed delays; uncertainties; LMIs
0 引言
人工神經(jīng)網(wǎng)絡在模式識別、并行計算和非線性最優(yōu)問題等方面的廣泛應用而受到極大的關注。然而,所有這些成功應用都極大地依賴于神經(jīng)網(wǎng)絡的動態(tài)特性。毫無疑問,穩(wěn)定性是其中神經(jīng)網(wǎng)絡的主要性能之一。在生物和人工神經(jīng)網(wǎng)絡系統(tǒng)中,時滯是不可避免的且時滯的存在也是導致網(wǎng)絡不穩(wěn)定的關鍵因素之一。到目前為止,時滯神經(jīng)網(wǎng)絡的穩(wěn)定性分析問題已經(jīng)取得了大量的研究成果,得到了很多保證時滯神經(jīng)網(wǎng)絡的漸近或指數(shù)穩(wěn)定性的充分條件。其中時滯類型包括常時滯、變時滯與分布時滯等[1~6]。
近年來,引入?yún)?shù)不確定性和隨機擾動的神經(jīng)網(wǎng)絡穩(wěn)定性分析問題已經(jīng)引起國內(nèi)外學者廣泛重視。其中原因來自兩方面:a)神經(jīng)元的連接權依賴于人工神經(jīng)網(wǎng)絡模型中的電阻和電容值,其中包括不確定性(建模誤差); b)在真實神經(jīng)系統(tǒng)中,突觸傳輸是一個由釋放神經(jīng)遞質(zhì)與其他一些隨機原因所帶來的隨機擾動過程。目前,該領域的研究已取得了一些進展。文獻[7]指出可通過確定隨機輸入使神經(jīng)網(wǎng)絡穩(wěn)定或不穩(wěn)定。文獻[8]研究了一類隨機時滯Hopfield神經(jīng)網(wǎng)絡在均方根意義下的指數(shù)穩(wěn)定性。文獻[9]研究了一類不確定離散與分布時滯Hopfield神經(jīng)網(wǎng)絡的隨機穩(wěn)定性。文獻[10]研究了一類不確定隨機時滯神經(jīng)網(wǎng)絡的穩(wěn)定性。文獻[11]研究了一類不確定離散脈沖系統(tǒng)的魯棒穩(wěn)定性。文獻[12]給出一類不確定隨機時滯Hopfield神經(jīng)網(wǎng)絡全局魯棒穩(wěn)定的新判據(jù)。然而,同時考慮參數(shù)不確定性擾動與外部隨機擾動的因素,對于區(qū)間時滯的不確定隨機離散分布神經(jīng)網(wǎng)絡的魯棒穩(wěn)定性的研究還較少。
1 系統(tǒng)描述與引理
其中:“*”代表對稱矩陣的對稱項,且
Ξ11=-2PA+2PW0L+Q+Z1+Z2+LRL+LEL+D0+(h2-h1)T
Ξ22=-(1-hd)Q+D1+bLL
Ξ55=-(1-hd)R-bI,Ξ66=-(h2-h1)-1T
證明 利用如下LyapunovKrasovskii泛函定義導出穩(wěn)定性結果:
根據(jù)It’s微分公式,沿著式(4)的任意軌跡,V(x(t),t)對時間t的隨機導數(shù)為
LV(x(t),t)≤2xT(t)P[-Ax(t)+W0f(x(t))+W1f(x(t-τ(t))+
D∫t-∞U(t-s)f(x(s))ds]+trace[δT(t,x(t), x(t-τ(t)))Pδ(t, x(t), x(t-τ(t)))]+
(xT(t)Qx(t)-(1-hd)xT(t-τ(t))Qx(t-τ(t))+
xT(t)Z1x(t)-xT(t-h1)Z1x(t-h1)+
xT(t)Z2x(t)-xT(t-h2)Z2x(t-h2)+
xT(t)LRLx(t)-(1-hd)fT(x(t-τ(t)))Rf(x(t-τ(t)))+
(h2-h1)xT(t)Tx(t)-∫t-h 1t-h2xT(s)Tx(s)ds+
∑ni=1ei∫∞0ui(s)f 2i(xi(t))ds-∑ni=1ei∫∞0ui(s)f 2i(xi(t-s))ds(5)
其中:
∑ni=1ei∫∞0ui(s)f2i(xi(t))ds-∑ni=1ei∫∞0ui(s)f2i(xi(t-s))ds=
fT(x(t))Ef(x(t))-∑ni=1ei∫∞0ui(s)ds∫∞0ui(s)f 2i(xi(t-s))ds≤
xT(t)LELx(t)-∑ni=1ei(∫∞0ui(s)fi(xi(t-s))ds)2=
xT(t)LELx(t)-(∫t-∞U(t-s)f(x(s))ds)TE(∫t-∞U(t-s)f(x(s))ds)(6)
考慮對任意正常數(shù) b,由式(3)得
-bfT(x(t-τ(t)))f(x(t-τ(t)))+bxT(t-τ(t))LTLx(t-τ(t))≥0(7)
由式(5)~(7)和引理3得
LV(x(t),t)≤xT(t)[-2PA+2PW0L+Q+Z1+Z2+LRL+LEL+
D0+(h2-h1)T]x(t)+xT(t)[2PW1]f(x(t-τ(t)))+
xT(t)[2PD](∫t-∞U(t-s)g(x(s))ds)+
x T(t-τ(t))[-(1-hd)Q+D1+bLL]x(t-τ(t))+
xT(t-h 1)[-Z1]x(t-h1)+xT(t-h2)[-Z2]x(t-h2)+
fT(x(t-τ(t))[-(1-hd)R-bI]f(x(t-τ(t)))+
(∫t-h 1t-h2x(s)ds)T[-(h2-h1)-1T](∫t-h 1t-h2x(s)ds)-
(∫t-∞U(t-s)f(x(s))ds)TE(∫t-∞U(t-s)f(x(s))ds)(8)
由式(8)可得LV(x(t),t)≤ξT(t)Ξξ(t)。其中:
ξT(t)=[xT(t) xT(t-τ(t)) xT(t-h1)xT(t-h2) f T(x(t-τ(t)))
(∫t-h 1t-h2x(s)ds) T(∫t-∞U(t-s)f(x(s))ds)T]
因此,對于式(4)在任何狀態(tài)下保證LV(x(t),t)為負定的充分條件是 Ξ<0。這就意味著式(4)在均方根意義下是全局魯棒穩(wěn)定的。證明完畢。
為了使式(4)更加符合實際,可在其基礎上引入不確定性參數(shù),作進一步的推廣??紤]如下不確定隨機變時滯神經(jīng)網(wǎng)絡模型:
dx(t)=[-(A+ΔA(t))x(t)+(W0+ΔW0(t))f(x(t))+(W1+ΔW1(t))f(x(t-τ(t)))+(D+ΔD(t))∫t-∞U(t-
s)f(x(s))ds]dt+σ(t,x(t),x(t-τ(t)))dω(t)(9)
其中:參數(shù)和變量與式(4)相同,不確定項ΔA(t)、ΔW0(t)、ΔW1(t)、ΔD(t)滿足
ΔA(t)=M1F(t)N1,ΔW0(t)=M2F(t)N2ΔW1(t)=M3F(t)N3,ΔD(t)=M4F(t)N4(10)
M1、M2、M3、M4、N1、N2,N3、N4是已知常矩陣;F(t)是未知有界的適維變時滯函數(shù),滿足
FT(t)F(t)≤I,t≥0(11)
定理2 如果存在矩陣P>0,D0≥0與D1≥0使得trace[σT(t,x(t),x(t-τ(t)))Pσ(t,x(t),x(t-τ(t)))]≤xT(t)D0x(t)+xT(t-τ(t))D1x(t-τ(t))成立,則式(9)在均方根意義下是全局魯棒穩(wěn)定的。如果存在正定對稱矩陣Q、R、T、Z1、Z2和對角陣E>0,正常數(shù)ε1、ε2、ε3、ε4與b,使得下列LMI成立:
Π=(1,1)PM1PM2PM3PM4*-ε1I000**-ε2I00***-ε3I0****-ε4I<0
(1,1)=Ξ11000PW10PD
*Ξ2200000
**-Z10000
***-Z2000
****Ξ5500
*****Ξ660
******Ξ77
其中:“*”代表對稱矩陣的對稱項,且
Ξ11=-2PA+2PW0L+Q+Z1+Z2+LRL+LEL+
D0+(h2-h1)T+ε1NT1N1+ε2NT2N2
Ξ22=-(1-hd)Q+D1+bLL,Ξ55=-(1-hd)R-bI+ε3NT3N3
Ξ66=-(h2-h1)-1T,Ξ77=-E+ε4NT4N4
證明利用定理1中同樣的LyapunovKrasovskii泛函導出穩(wěn)定性結果。根據(jù)It’s微分公式,沿著式(9)的任意軌跡,對V(x(t),t)求時間t的隨機導數(shù)。利用引理1與2,可得
LV(x(t),t)≤Ξ+ε-11Ω1ΩT1+ε1ΩT2Ω2+ε-12Ω3ΩT3+ε2ΩT4Ω4+
ε-13Ω5ΩT5+ε3ΩT6Ω6+ε-14Ω7ΩT7+ε4ΩT8Ω8<0
Ω1=[MT1PT 0 0 0 0 0 0]T,Ω2=[-N1 0 0 0 0 0 0]
Ω3=[MT2PT 0 0 0 0 0 0]T,Ω4=[N2 0 0 0 0 0 0]
Ω5=[MT3PT 0 0 0 0 0 0]T,Ω6=[0 0 0 0 N3 0 0]
Ω7=[MT4PT 0 0 0 0 0 0]T,Ω8=[0 0 0 0 0 0 N4]
因此,在滿足定理2中LMI成立的條件下,式(9)是全局魯棒穩(wěn)定的。
3 仿真示例
示例1 考慮一個二階隨機離散區(qū)間分布時滯神經(jīng)網(wǎng)絡式(4),其參數(shù)如下:
A=2002,W0=0.3-0.2-0.20.1,W1=0.10.10.10.2
D=0.60.50.4-0.6,h2=2.2,h1=0.4,τd=0.8
L=I,D0=D1=diag{0.5,0.32}
由定理1,應用LMI工具箱求解線性矩陣不等式可知,式(4)在均方根意義下是全局魯棒穩(wěn)定的??尚薪馊缦拢?/p>
P=359.1908-383.5376-383.5376342.7975,Q=171.19678.18668.1866179.6002
R=95.081534.965034.9650134.5515,T=65.88775.12355.123573.8860
Z1=81.30154.25034.250387.9166,Z2=81.30154.25034.250387.9166
E=272.056700272.0567,b=19.0310
示例2 考慮一個二階不確定隨機離散區(qū)間分布時滯神經(jīng)網(wǎng)絡式(9),其參數(shù)如下:
A=2.5002,W0=-0.50.10.1-0.1,W1=-0.10.20.20.1
D=0.60.4-0.50.6,h2=2.5,h1=1,τd=0.8
L=I,D0=D1=diag{0.4,0.2}
N1=N2=N3=N4=0.2I,M1=M2=M3=M4=0.1I
由定理2,應用LMI工具箱求解線性矩陣不等式可知,式(9)在均方根意義下是全局魯棒穩(wěn)定的。可行解如下:
P=51.61272.17662.176663.7344,Q=58.7580-0.3669-0.366937.0565
R=39.69140.59550.595525.3821,T=28.8978-0.2342-0.234220.2474
Z1=27.6890-0.1566-0.156621.7824,Z2=27.6890-0.1566-0.156621.7824
E=45.62000045.6200,b=5.3377
ε1=31.0144,ε2=34.0144,ε3=31.0558,ε4=33.0484
4 結束語
本文研究了一類不確定隨機離散分布區(qū)間時滯神經(jīng)網(wǎng)絡模型的全局魯棒穩(wěn)定性問題?;诰€性矩陣不等式(LMI)技術,給出了對于任意的有界區(qū)間時滯0<h1≤τ(t)≤h2,該類神經(jīng)網(wǎng)絡在其平衡點為全局魯棒穩(wěn)定的充分條件。文中的結論對于神經(jīng)網(wǎng)絡的分析及設計是很實用的。最后,通過仿真示例進一步驗證了結論的有效性。
參考文獻:
[1]ARIK S. Stability analysis of delayed neural networks[J]. IEEE Trans on Circuits Systems I, 2000, 47(7):10891092.
[2]CAO Jinde, WANG Jun. Global asymptotic and robust stability of recurrent neural networks with time delays[J]. IEEE Trans on Circuits Systems I, 2005, 52(2):417426.
[3]LIAO Xiaofeng, CHEN Guanrong, SANCHEZ E N. Delaydependent exponential stability analysis of delayed neural networks: an LMI approach[J]. Neural Networks, 2002, 15(7):855866.
[4]ZHANG Jiye, JIN Xuesong. Global stability analysis in delayed Hopfield neural network models[J]. Neural Networks, 2000, 13(7):745753.
[5]WANG Zidong, LIU Yurong, LIU Xiaohui. On global asymptotic stability of neural networks with discrete and distributed delays[J]. Physics Letters A, 2005, 345(46):299308.
[6]CAO Jinde, WANG Jun. Global exponential stability and periodicity of recurrent neural networks with time delays[J]. IEEE Trans on Circuits Systems I, 2005, 52(5):920931.
[7]BLYTHE S, MAO Xuerong, LIAO Xiaoxin. Stability of stochastic delay neural networks[J]. Journal of the Franklin Institute, 2001, 338(4):481495.
[8]WAN Li, SUN Jianhua. Mean square exponential stability of stochastic delayed Hopfield neural networks[J]. Physics Letters A, 2005, 343(4):306318.
[9]WANG Zidong, LIU Yurong, FRASER K, et al. Stochastic stability of uncertain Hopfield neural networks with discrete and distributed delays[J]. Physics Letters A, 2006, 354(4):288297.
[10]HUANG He, FENG Gang. Delaydependent stability for uncertain stochastic neutral networks with timevarying delay[J]. Physica A, 2007, 381:93103.
[11]LIU Bin, LIU Xinzhi. Robust stability of uncertain discrete impulsive systems[J]. IEEE Trans on Circuits Systems Ⅱ, 2007, 54(5):455459.
[12]ZHANG Jinhui, SHI Peng, QIU Jiqing. Novel robust stability criteria for uncertain stochastic Hopfield neural networks with timevarying delays[J]. Nonlinear Analysis: Real World Applications, 2007, 8(4):13491357.
[13]MAO Xuerong. Stochastic differential equations and their applications[M]. Chichester: Horwood Publishers, 1997.