孫云霞 程培 李殿強 尚蕾
摘 要:研究帶有脈沖的隨機Cohen-Grossberg神經(jīng)網(wǎng)絡的幾乎必然指數(shù)穩(wěn)定性問題, 基于Lyapunov穩(wěn)定性理論,利用隨機分析技巧和線性矩陣不等式工具,得到系統(tǒng)基于矩陣不等式的幾乎必然指數(shù)穩(wěn)定性充分條件, 并通過一個例子來驗證結論的有效性.
關鍵詞:Cohen-Grossberg神經(jīng)網(wǎng)絡;脈沖;線性矩陣不等式;Lyapunov函數(shù);幾乎必然指數(shù)穩(wěn)定
中圖分類號:O231 文獻標志碼:A
0 引言
Cohen-Grossberg神經(jīng)網(wǎng)絡(簡稱CGNN)是由Cohen等[1]于1983年首次提出的一種神經(jīng)網(wǎng)絡模型,它包括種群生物學、神經(jīng)生物學、進化理論等學科中許多著名模型作為其特例. 作為一類廣義的神經(jīng)網(wǎng)絡模型,CGNN模型在模式識別、系統(tǒng)辨識、信號處理、圖像處理、最優(yōu)化、機器學習以及控制等方面都有廣泛的應用.
在實際神經(jīng)網(wǎng)絡中,突觸之間信息的傳遞往往會受到由神經(jīng)遞質或其它隨機因素的釋放而導致的隨機噪聲的影響. 隨機噪聲的存在可能使得神經(jīng)網(wǎng)絡產(chǎn)生振蕩行為或其它失穩(wěn)現(xiàn)象甚至出現(xiàn)混沌現(xiàn)象,從而影響神經(jīng)網(wǎng)絡的整體性能[2]. 另外, 在電子網(wǎng)絡的運行過程中,其狀態(tài)可能會受到由切換、頻率變化或其它突發(fā)噪聲引起的瞬時擾動,從而在特定時刻經(jīng)歷瞬時突變,這種瞬時突變現(xiàn)象稱為脈沖擾動. 近年來, 神經(jīng)網(wǎng)絡[3-4]、隨機CGNN[5-9]以及帶有脈沖的隨機CGNN [10-12]的穩(wěn)定性問題吸引了大批學者的關注,并取得了很多的研究成果. 但是,目前關于帶有脈沖的隨機CGNN穩(wěn)定性方面的研究主要都集中于均方穩(wěn)定性分析方面,而關于幾乎必然(a.s.)指數(shù)穩(wěn)定性方面的研究結果不多見.
本文將基于Lyapunov穩(wěn)定性理論,利用It?觝公式、指數(shù)鞅不等式和Borel-Cantelli 引理等隨機分析技巧,結合線性矩陣不等式(LMI)工具,研究帶有脈沖的隨機CGNN的幾乎必然指數(shù)穩(wěn)定性,建立系統(tǒng)幾乎必然指數(shù)穩(wěn)定的充分條件,并通過一個數(shù)值例子及仿真模擬來驗證所獲結果的有效性.
1 準備知識
本文采用以下記號:記(Ω,{Ft}t≥0,P)為帶有σ代數(shù)流{Ft}t≥0的完備概率空間,w(t)=(w1(t),…,wm(t))T為定義于該空間上的m維標準布朗運動,N為正整數(shù)集,Rn為n維歐氏空間,Rn×m為n×m實矩陣,I為合適維數(shù)的單位矩陣,符號diag表示對角矩陣.上標T表示向量或矩陣的轉置,符號*表示矩陣中由對稱性得到的元素.
考慮帶有脈沖的隨機CGNN:
dx(t)=-a(x(t))[b(x(t))-Ag(x(t))]dt+σ(x(t))dw(t),t≠tkx(tk)= (1)
其中:x(t)=[x1(t),…,xn(t)]T為n維神經(jīng)元狀態(tài)向量,矩陣a(x(t))=diag(a1(x1(t)),…,an(xn(t)))為放大函數(shù),b(x(t))=
[b1(x1(t)),…,bn(xn(t))]T為神經(jīng)元形為函數(shù),A∈Rn×n為連接權矩陣,g(x(t))=[g1(x1(t)),…,gn(xn(t))]T為神經(jīng)元激勵函
數(shù),σ(x(t))∈Rn×m為噪聲強度矩陣函數(shù).x(tk)= 表示系統(tǒng)在脈沖時刻tk的狀態(tài)跳變,Ck為脈沖強度矩陣. 假設脈沖時刻tk滿足0=t0假設1 存在正常數(shù)hi和li(i=1,2,…,n),使得:0
證:由假設1可知,函數(shù)a(x(t))滿足:
a(x(t))·a(x(t))≤l2I
線性矩陣不等式(3)兩邊分別乘以diag(a(x(t)), I)得:
Φ=-2hQ△+γD-(k0-■-α)Q a(x(t))QA+a(x(t))K?撰 * -2h?撰<0 (7)
定義Lyapunov函數(shù)V(x(t))=xT(t)Qx(t). 由系統(tǒng)(1)的第二個方程及不等式(6)可知:
(8)
對任意t≠tk, 由It?觝公式可得:
dV(x(t))=LV(x(t))dt+HV(x(t))dw(t) (9)
其中:
LV(x(t))=-2xT(t)Qa(x(t))[b(x(t))-Ag(x(t))]+trace[σT(x(t))Qσ(x(t))]≤
-2xT(t)Qa(x(t))b(x(t))+2xT(t)a(x(t))QAg(x(t))+γxT(t)Dx(t),
HV(x(t))=2xT(t)Qσ(x(t)).
由假設3知:
-■λiai(xi(t))gi(xi(t))(gi(xi(t))-kixi)≥0 (10)
由假設1和假設2知:
-2xT(t)Qa(x(t))b(x(t))≤-2hxT(t)Q△x(t) (11)
將式(10)和式(11)代入式(9)可得:
LV(x(t))≤xT(t)-2hQ△+γD-(k0-■-α)Qx(t)+2xT(t)a(x(t))QAg(x(t))-
2■λiai(xi(t))gi(xi(t))(gi(xi(t))-kixi)+(k0-■-α)xT(t)Qx(t)≤ξT(t)Φξ(t)+(k0-■-α)V(x(t)),
其中:ξT(t)=[xT(t), gT(x(t))].
由Φ<0, 知:
LV(x(t))≤(k0-■-α)V(x(t)) (12)
由式(5)知:
HV(x(t))■=2xT(t)Qσ(x(t))■=4xT(t)Qσ(x(t))■≥2k0V2(x(t)) (13)
對任意t∈[tk-1, tk), k∈N,由It?觝公式可得:
lnV(x(t))=lnV(x(tk-1))+■■ds-■■■ds+■■dw(s) (14)
對于t=tk,由式(8)可知:
(15)
因此,對于任意t≥t0, 由式(12)~式(15)并利用迭代技巧可得:
lnV(x(t))≤lnV(x0)+■lnμ+(k0-■-α)(t-t0)-■■■ds+M(t) (16)
其中:
M(t)=■■dw(s)
是一個連續(xù)鞅并且滿足初值M(t0)=0, 其二次變差過程=■■ds.
任意取定ε∈(0,1),由指數(shù)鞅不等式,對k∈N
由Borel-Cantelli 引理可得對幾乎所有的樣本點ω∈Ω,都存在一個整數(shù)k0=k0(ω),當k≥k0時有:
因此,對任意t0≤t≤k都有:
M(t)≤■lnk+■
證:由假設1及式(23)知
Φ1=-2hQ△+BTQB-(k0-■-α)Q a(x(t))QA+a(x(t))K?撰 * -2h?撰<0.
定義Lyapunov函數(shù)V(x(t))=xT(t)Qx(t). 由條件(24)可知:
對任意t≠tk,由It?觝公式可得:
dV(x(t))=LV(x(t))dt+HV(x(t))dw(t),
其中:
2■λiai(xi(t))gi(xi(t))(gi(xi(t))-kixi)+(k0-■-α)xT(t)Qx(t)≤ξ1T(t)Φ1ξ1(t)+(k0-■-α)V(x(t)),
. 由Φ1<0,知:
LV(x(t))≤(k0-■-α)V(x(t))
由式(25)知:
以下證明過程與定理1相同,故省略. 證畢.
3 數(shù)值例子
考慮如下脈沖隨機CGNN:
dx(t)=-a(x(t))[b(x(t))-Ag(x(t))]dt+σ(x(t))dw(t), t≠tkx(tk)=Cx(tk-),k∈N (26)
其中,a(x(t))=3+sinx1(t) 0 0 3+sinx2(t),b(x(t))=2x1(t)2x2(t),A=1 00 1,σ(x(t))=4x1(t)4x2(t),g(x(t))=tanh(x1(t))tanh(x2(t)),C=1.2I,
tk=0.1k.
通過簡單計算可選取:
h=2, l=4, k0=32, Δ=diag(2,1), K=0.5I, D=16I.
令ρ=0.386 7, α=0.01, 利用MATLAB工具箱求解可得滿足線性矩陣不等式(3)和式(4)的可行解:
γ=25.930 3, Q=19.212 3 0 0 20.443 3,?撰=11.943 5 0 0 12.922 3.
由脈沖強度矩陣C=1.2I知,脈沖對網(wǎng)絡的穩(wěn)定性起擾動作用,選取μ=1.44,則矩陣不等式(6)成立.
由定理1知神經(jīng)網(wǎng)絡(26) 在lmin(0.386 7)上一致幾乎必然指數(shù)穩(wěn)定. 選取脈沖時間間隔tk-tk-1=0.4, 初始值x0=[0.5, 0.1]T,神經(jīng)網(wǎng)絡(26)的單個樣本軌跡如圖1所示:
4 結論
本文對一類隨機脈沖Cohen-Grossberg神經(jīng)網(wǎng)絡的幾乎必然指數(shù)穩(wěn)定性進行了研究.通過選取適當?shù)腖yapunov函數(shù)和利用線性矩陣不等式工具,得到判定其幾乎必然指數(shù)穩(wěn)定的充分性條件.最后,通過一個數(shù)值例子和仿真模擬驗證了結論的有效性.
參考文獻
[1] COHEN M A, GROSSBERG S, GROSSBER. Absolute stability of global pattern formation and parallel memory storage by competitive neural networks[J]. IEEE Transactions on Systems, Man, and Cybernetics, 1983, 13(5):815-826.
[2] 劉少寶,吳瑩,郝忠文, 等. 鈉離子和鉀離子通道噪聲擾動對神經(jīng)網(wǎng)絡時空模式的影響[J]. 物理學報,2012,61(2):1-7.
[3] 李向陽,曾文波. 基于BP神經(jīng)網(wǎng)絡矩形微帶天線諧振頻率預測[J]. 廣西科技大學學報,2014,25(3):26-31.
[4] 陳偉,孔峰,陶金. 神經(jīng)網(wǎng)絡在網(wǎng)絡檢測中的應用[J]. 廣西工學院學報,2011,22(1):78-81.
[5] WANG C H , KAO Y G , YANG G W. Exponential stability of impulsive stochastic fuzzy reaction-diffusion Cohen-Grossberg neural networks with mixed delays[J]. Neurocomputing, 2012, 89:55-63.
[6] ZHU Q X, CAO J , RAKKIYAPPAN R. Exponential input-to-state stability of stochastic Cohen-Grossberg neural networks with mixed delays[J]. Nonlinear Dynamics,2015,79(2):1085-1098.
[7] LI X D, FU X L. Global asymptotic stability of stochastic Cohen-Grossberg-type BAM neural network with mixed delays: an LMI approach[J]. Journal of Computational and Applied Mathematics,2011,235(12):3385-3394.
[8] ZHU Q X , LI X D. Exponential and almost sure exponential stability of stochastic fuzzy delayed Cohen-Grossberg neural networks[J]. Fuzzy Sets and Systems, 2012, 203 :74-94.
[9] YANG Y Q, LIANG T, XU X Y. Almost sure exponential stability of stochastic Cohen-Grossberg neural networks with continuous distributed delays of neutral type[J]. International Journal for Light and Electron Optics,2015,126(23):4628-4635.
[10] DONG M, ZhANG H G, WANG Y C. Dynamics analysis of impulsive stochastic Cohen-Grossberg neural networks with Markovian jumping and mixed time delays[J]. Neurocomputing,2009,72(7-9):1999-2004.
[11] RAKKIYAPPAN R, CHANDRASEKAR A, LAKSHMANAN S, el at. Park. Exponential stability of Markovian jumping stochastic Cohen-Grossberg neural networks with mode-dependent probabilistic time-varying delays and impulses[J]. Neurocomputing,2014,
131:265-277.
[12] ZHU Q X, CAO J D.Robust exponential stability of Markovian jumping impulsive stochastic Cohen-Grossberg neural networks with mixed time delays[J]. IEEE Transactions on Neural Networks,2010,21(8):1314-1324.
Abstract: This paper focuses on the analysis of almost sure exponential stability of stochastic Cohen-Grossberg neural networks with impulse. Based on the Lyapunov stability theory, by using some stochastic analysis techniques and linear matrix inequality tool, we establish a set of sufficient conditions of almost sure exponential stability of system in terms of matrix inequalities. Finally, we give a numerical example to illustrate the effectiveness of the results obtained.
Key words: Cohen-Grossberg neural networks; impulse; linear matrix inequality; Lyapunov function; almost sure exponential stability
(學科編輯:張玉鳳)