Abstract
This paper studies the robust stability of uncertain neural networks with multiple time delays with respect to the class of nondecreasing activation functions. By using the Lyapunov functional and homeomorphism mapping theorems, we derive a new delayindependent sufficient condition the existence, uniqueness, and global asymptotic stability of the equilibrium point for delayed neural networks with uncertain network parameters. The condition obtained for the robust stability establishes a matrixnorm relationship between the network parameters of the neural system, and therefore it can easily be verified. We also present some constructive numerical examples to compare the proposed result with results in the previously published corresponding literature. These comparative examples show that our new condition can be considered as an alternative result to the previous corresponding literature results as it defines a new set of network parameters ensuring the robust stability of delayed neural networks.
Keywords:
stability analysis; delayed neural networks; interval matrices; Lyapunov functionals1 Introduction
Dynamical neural networks have recently received a great deal of attention due to their potential applications in image and signal processing, combinatorial optimization problems, pattern recognition, control engineering, and some other related areas. In the electronic implementation of analog neural networks, during the processing and transmission of signals in the network, due to the finite switching speed of amplifiers, some time delays occur which may change the dynamical behavior of the network from stable to unstable. Therefore, it is important to take into account the effects of the time delays in the dynamical analysis of neural networks. On the other hand, it is well known that unavoidably some disturbances are to be considered in the modeling and stability analysis neural networks. The major disturbances occur within the network, which is mainly due to the deviations in the values of the electronic components during the process of implementation. Therefore, in recent years, many papers have focused on studying the existence, uniqueness, and global robust asymptotic stability of the equilibrium point in the presence of time delays and parameter uncertainties for various classes of nonlinear neural networks, and one reported some robust stability results [143].
In the current paper, we aim to study the robust stability of a class of uncertain neural networks with multiple time delays. By using the Lyapunov functional and homeomorphism mapping theorems, a new delayindependent sufficient condition for global robust asymptotic stability of the equilibrium point for this class of neural networks is derived. Meanwhile, three numerical examples are presented to demonstrate the applicability of the condition and to show the advantages of our result over the previously published robust stability results.
We use the following notation. Throughout this paper, the superscript T represents the transpose. I stands for the identity matrix of appropriate dimension. For the vector , will denote . For any real matrix , will denote , and and will denote the minimum and maximum eigenvalues of Q, respectively. If is a symmetric matrix, then will imply that Q is positive definite, i.e., Q has all eigenvalues real and positive. Let and be two symmetric matrices. Then, will imply that for any real vector . A real matrix is said to be a nonnegative matrix if , . Let and be two real matrices. Then, will imply that , . We also recall the following vector and matrix norms:
2 Preliminaries
The delayed neural network model we consider in this paper is described by the set of nonlinear differential equations of the form
where , n is the number of the neurons, denotes the state of the neuron i at time t, denote activation functions, and denote the strengths of connectivity between neurons j and i at time t and , respectively; represents the time delay required in transmitting a signal from the neuron j to the neuron i, is the constant input to the neuron i, is the charging rate for the neuron i.
In order to accomplish the objectives of this paper in the sense of robust stability of dynamical neural networks, we will first define the class of the activation functions that we will employ in the neural network model (2.1) and the parametric uncertainties of the system matrices A, B, and C.
The activation functions are assumed to be nondecreasing and slopebounded, that is, there exist some positive constants such that the following conditions hold:
This class of functions will be denoted by .
We will set intervals for the system matrices , and in (2.1) as follows:
In what follows, we will give some basic definitions and lemmas that will play an important role in the proof of our robust stability results.
Definition 2.1 (See [28])
Let be an equilibrium point of neural system (2.1). The neural network model (2.1) with the parameter ranges defined by (2.2) is globally asymptotically robust stable if is a unique and globally asymptotically stable equilibrium point of system (2.1) for all , , and .
Lemma 2.1 (See [29])
Ifsatisfies the conditionsfor allandas, thenis a homeomorphism of.
Lemma 2.2 (See [1])
then, for any positive diagonal matrix P and a nonnegative diagonal matrix ϒ, the following inequality holds:
3 Robust stability analysis
In this section, we will present a new sufficient condition that guarantees the global robust asymptotic stability of the equilibrium point of the neural network model (2.1), which is stated in the following theorem.
Theorem 3.1For the neural network model (2.1), assume thatand the network parameters satisfy (2.2). Then the neural network model (2.1) has a unique and globally robust asymptotically stable equilibrium point for eachu, if there exist a positive diagonal matrixPand a nonnegative diagonal matrix ϒ such that the following condition holds:
where, , , , withwith, andare positive constants such thatand.
Proof We will first prove the existence and uniqueness of the equilibrium point of system (2.1) by making use of the homeomorphism mapping theorem defined in Lemma 2.1. To this end, we define the mapping associated with system (2.1) as
We point out here that if is an equilibrium point of the neural network model (2.1), then, by definition, satisfies the following equilibrium equation:
Therefore, every solution of the equation is an equilibrium point of system (2.1). Hence, if we shaw that is homeomorphism of , then we will conclude that has a unique solution for each u. In order to prove that is a homeomorphism of , we choose two real vectors and such that . In this case, we can write the following equation for given by (3.1):
Let . If when , then (3.2) takes the form
from which it follows that if as C is a positive diagonal matrix. Now assume that when . In this case, if we multiply both sides of (3.2) by , we obtain
where is a positive diagonal matrix.
In the light of Lemma 2.2, we can write
where ϒ is a nonnegative diagonal matrix.
We also note the following inequality:
where and are positive constants such that and , and .
Equation (3.6) can be written in the following form:
where α and β are some positive constants. Letting and in (3.7) yields
Using (3.4), (3.5), and (3.8) in (3.3) results in
which can be written in the form
For the activations functions belonging to the class , it has been shown in [27] that for the inequality in the form of (3.9), if , then , for all , and as . Hence, we have proved that the map is a homomorphism of , meaning that the condition of Theorem 3.1 implies the existence and uniqueness of the equilibrium point for neural network model (2.1).
It will be now shown that the condition obtained for the existence and uniqueness of the equilibrium point of neural network model (2.1) in Theorem 3.1 also implies the global asymptotic stability of the equilibrium point. To this end, we will shift the equilibrium point of system (2.1) to the origin. The transformation , , puts the network model (2.1) into the following form:
where , , satisfies the following property:
Note that equilibrium and stability properties of the neural network models are identical. Therefore, proving the asymptotic stability of the origin of system (3.10) will directly imply the asymptotic stability of . Now, consider the following positive definite Lyapunov functional for system (3.10):
where the , α, β, γ, and ε are positive constants to be determined later. The time derivative of the functional along the trajectories of system (3.10) is obtained as follows:
We note that the following inequalities hold:
In the light of Lemma 2.2, we can write
Using (3.12)(3.16) in (3.11) yields
where , , . We now note the following inequalities:
Using (3.18)(3.20) in (3.17) leads to
Let , and . Then (3.21) takes the form
It has been shown in [27] that, for the Lyapunov functional defined above, if its time derivative is in the form of (3.22) and with Π being positive definite, then the origin system (3.10), or equivalently the equilibrium point of system (2.1) is globally asymptotically stable. Hence, we have shown that the condition of Theorem 3.1 implies the global robust asymptotic stability of system (2.1). □
4 Comparison and examples
In this section, we present some constructive numerical materials to demonstrate the effectiveness and applicability of the proposed conditions and to show the advantages of our results over the previous corresponding robust stability result derived in the literature. In order to make a precise comparison between the results, we will first restate results from the previous literature.
Theorem 4.1 (See [27])
For the neural network model (2.1), assume thatand the network parameters satisfy (2.2). Then the neural network model (2.1) is globally asymptotically robust stable if there exists a positive diagonal matrixsuch that
whereis a positive diagonal matrix, with, for, , with, and, andandare positive constants such thatand.
Theorem 4.2 (See [27])
For the neural network model (2.1), assume thatand the network parameters satisfy (2.2). Then the neural network model (2.1) is globally asymptotically robust stable, if there exists a positive diagonal matrixsuch that
whereis a positive diagonal matrix, , with, and; andare positive constants such thatand, and, .
Theorem 4.3 (See [30])
For the neural network model (2.1), assume thatand the network parameters satisfy (2.2). Then the neural network model (2.1) is globally asymptotically robust stable if there exist positive constants, , such that
We will now consider the following examples.
Example 4.1 Consider the neural system (2.1) with the following network parameters:
from which we obtain the following matrices:
We note that and . We consider a special case of Theorem 4.1 where , in which case the matrix S in Theorem 4.1 is of the form
Then, for and , Λ in Theorem 4.1 is obtained as follows:
Note that if and only if . Hence, the robust stability condition imposed by Theorem 4.1 is . For the same network parameters of this example, we will obtain the robust stability condition imposed by Theorem 3.1. Let and
Then we obtain
Π in Theorem 3.1 takes the form
It can be calculated that if and only if . Therefore, Theorem 3.1 imposes a less restrictive stability condition on the network parameters than Theorem 4.1 does.
Example 4.2 Consider the neural system (2.1) with the following network parameters:
From the above matrices, we can obtain the following matrices:
from which we can calculate the norms and . Let and
Then we can write
For and , the matrix Π in Theorem 3.1 is in the form
The condition is satisfied if . For the parameters of this example, Γ in Theorem 4.2 is in the form
The choice implies that , which ensures the global robust stability of neural system (2.1). Hence, for the network parameters of this example, if , then the result of Theorems 4.2 does not hold, whereas the result of Theorem 3.1 is still applicable.
Example 4.3 Assume that the network parameters of neural system (2.1) are given as follows:
For the matrices given above, we can obtain the following matrices:
from which we calculate and . Let and
We have
For and , the matrix Π in Theorem 3.1 is of the form
Clearly, holds if and only if . Therefore, for this example, Theorem 3.1 ensures the global robust stability of neural system (2.1) under the condition that .
When checking the condition of Theorem 4.3 for the same network parameters of this example, we search for the existence of the positive constants , , , and such that the following conditions hold:
which can be written in form
From the properties of the nonsingular Mmatrices [44], in order to ensure the existence of , , , and the symmetric matrix in the above inequality must be positive definite, which holds if and only if . Obviously, for the interval , our condition obtained in Theorem 3.1 is satisfied, but the result of Theorem 4.3 does not hold.
5 Conclusions
In this paper, we have focused on the existence, uniqueness, and global robust stability of an equilibrium point for neural networks with multiple time delays with respect to the class of nondecreasing activation functions. We have employed a suitable Lyapunov functional and made use of the homeomorphism mapping theorem to derive a new timeindependent robust stability condition for dynamical neural networks with multiple time delays. The obtained condition basically establishes a relationship between the network parameters of the neural system and the number of neurons. We have also presented some numerical examples, which enabled us to show the advantages of our result over previously reported robust stability results. We should point here that in the neural network model we have considered, the delay parameters are constant and the stability condition we obtain is delay independent. However, it is possible to derive some delaydependent stability conditions for the same neural network model by employing different classes of Lyapunov functionals.
Competing interests
The author declares that he has no competing interests.
References

Shen, Y, Wang, J: Robustness analysis of global exponential stability of recurrent neural networks in the presence of time delays and random disturbances. IEEE Trans. Neural Netw. Learn. Syst.. 23, 87–96 (2012)

Huang, TW, Li, CD, Duan, SK, Starzyk, A: Robust exponential stability of uncertain delayed neural networks with stochastic perturbation and impulse effects. IEEE Trans. Neural Netw. Learn. Syst.. 23, 866–875 (2012)

Sakthivel, R, Raja, R, Anthoni, SM: Exponential stability for delayed stochastic bidirectional associative memory neural networks with Markovian jumping and impulses. J. Optim. Theory Appl.. 150, 166–187 (2011). Publisher Full Text

Shen, H, Huang, X, Zhou, J, Wang, Z: Global exponential estimates for uncertain Markovian jump neural networks with reactiondiffusion terms. Nonlinear Dyn.. 69, 473–486 (2012). Publisher Full Text

Wu, ZG, Park, JH, Su, H, Chu, J: Robust dissipativity analysis of neural networks with timevarying delay and randomly occurring uncertainties. Nonlinear Dyn.. 69, 1323–1332 (2012). Publisher Full Text

Wu, ZG, Park, JH, Su, H, Chu, J: Stochastic stability analysis of piecewise homogeneous Markovian jump neural networks with mixed timedelays. J. Franklin Inst.. 349, 2136–2150 (2012). Publisher Full Text

Guo, Z, Huang, L: LMI conditions for global robust stability of delayed neural networks with discontinuous neuron activations. Appl. Math. Comput.. 215, 889–900 (2009). Publisher Full Text

Zhang, Z, Zhou, D: Global robust exponential stability for secondorder CohenGrossberg neural networks with multiple delays. Neurocomputing. 73, 213–218 (2009). Publisher Full Text

Han, W, Liu, Y, Wang, L: Robust exponential stability of Markovian jumping neural networks with modedependent delay. Commun. Nonlinear Sci. Numer. Simul.. 15, 2529–2535 (2010). Publisher Full Text

Yuan, Y, Li, X: New results for global robust asymptotic stability of BAM neural networks with timevarying delays. Neurocomputing. 74, 337–342 (2010). Publisher Full Text

Wang, Z, Liu, Y, Liu, X, Shi, Y: Robust state estimation for discretetime stochastic neural networks with probabilistic measurement delays. Neurocomputing. 74, 256–264 (2010). Publisher Full Text

Zhang, Z, Yang, Y, Huang, Y: Global exponential stability of interval general BAM neural networks with reactiondiffusion terms and multiple timevarying delays. Neural Netw.. 24, 457–465 (2011). PubMed Abstract  Publisher Full Text

Kao, Y, Wang, C: Global stability analysis for stochastic coupled reactiondiffusion systems on networks. Nonlinear Anal., Real World Appl.. 14, 1457–1465 (2013). Publisher Full Text

Muralisankar, S, Gopalakrishnan, N, Balasubramaniam, P: An LMI approach for global robust dissipativity analysis of TS fuzzy neural networks with interval timevarying delays. Expert Syst. Appl.. 39, 3345–3355 (2012). Publisher Full Text

Shao, JL, Huang, TZ, Wang, XP: Further analysis on global robust exponential stability of neural networks with timevarying delays. Commun. Nonlinear Sci. Numer. Simul.. 17, 1117–1124 (2012). Publisher Full Text

Lou, X, Ye, Q, Cui, B: Parameterdependent robust stability of uncertain neural networks with timevarying delay. J. Franklin Inst.. 349, 1891–1903 (2012). Publisher Full Text

Raja, R, Samidurai, R: New delay dependent robust asymptotic stability for uncertain stochastic recurrent neural networks with multiple time varying delays. J. Franklin Inst.. 349, 2108–2123 (2012). Publisher Full Text

Kao, YG, Guo, JF, Wang, CH, Sun, XQ: Delaydependent robust exponential stability of Markovian jumping reactiondiffusion CohenGrossberg neural networks with mixed delays. J. Franklin Inst.. 349, 1972–1988 (2012). Publisher Full Text

Zhang, H, Liu, Z, Huang, GB: Novel delaydependent robust stability analysis for switched neutraltype neural networks with timevarying delays via SC technique. IEEE Trans. Syst. Man Cybern., Part B, Cybern.. 40, 1480–1491 (2010)

Huang, Z, Li, X, Mohamad, S, Lu, Z: Robust stability analysis of static neural network with Stype distributed delays. Appl. Math. Model.. 33, 760–769 (2009). Publisher Full Text

Luo, M, Zhong, S, Wang, R, Kang, W: Robust stability analysis for discretetime stochastic neural networks systems with timevarying delays. Appl. Math. Comput.. 209, 305–313 (2009). Publisher Full Text

Balasubramaniam, P, Ali, MS: Robust stability of uncertain fuzzy cellular neural networks with timevarying delays and reaction diffusion terms. Neurocomputing. 74, 439–446 (2010). Publisher Full Text

Wang, Z, Liu, Y, Liu, X, Shi, Y: Robust state estimation for discretetime stochastic neural networks with probabilistic measurement delays. Neurocomputing. 74, 256–264 (2010). Publisher Full Text

Mahmoud, MS, Ismail, A: Improved results on robust exponential stability criteria for neutraltype delayed neural networks. Appl. Math. Comput.. 217, 3011–3019 (2010). Publisher Full Text

Shao, JL, Huang, TZ, Wang, XP: Improved global robust exponential stability criteria for interval neural networks with timevarying delays. Expert Syst. Appl.. 38, 15587–15593 (2011). Publisher Full Text

Faydasicok, O, Arik, S: An analysis of stability of uncertain neural networks with multiple time delays. J. Franklin Inst.. 350, 1808–1826 (2013). Publisher Full Text

Chen, A, Cao, J, Huang, L: Global robust stability of interval cellular neural networks with timevarying delays. Chaos Solitons Fractals. 23, 787–799 (2005). Publisher Full Text

Faydasicok, O, Arik, S: Equilibrium and stability analysis of delayed neural networks under parameter uncertainties. Appl. Math. Comput.. 218, 6716–6726 (2012). Publisher Full Text

Sun, C, Feng, CB: Global robust exponential stability of interval neural networks with delays. Neural Process. Lett.. 17, 107–115 (2003). Publisher Full Text

Liu, B: Global exponential stability for BAM neural networks with timevarying delays in the leakage terms. Nonlinear Anal., Real World Appl.. 14, 559–566 (2013). Publisher Full Text

Zhang, H, Shao, J: Almost periodic solutions for cellular neural networks with timevarying delays in leakage terms. Appl. Math. Comput.. 219, 11471–11482 (2013). Publisher Full Text

Feng, Z, Lam, J: Stability and dissipativity analysis of distributed delay cellular neural networks. IEEE Trans. Neural Netw.. 22, 976–981 (2011). PubMed Abstract  Publisher Full Text

Berezansky, L, Bastinec, J, Diblik, J, Smarda, Z: On a delay population model with quadratic nonlinearity. Adv. Differ. Equ.. 2012, (2012) Article ID 230
Article ID 230
BioMed Central Full Text 
Diblik, J, Khusainov, DY, Grytsay, IV, Smarda, Z: Stability of nonlinear autonomous quadratic discrete systems in the critical case. Discrete Dyn. Nat. Soc.. 2010, (2010) Article ID 539087
Article ID 539087
Publisher Full Text 
Diblik, J, Koksch, N: Sufficient conditions for the existence of global solutions of delayed differential equations. J. Math. Anal. Appl.. 318, 611–625 (2006). Publisher Full Text

Jia, R, Yang, M: Convergence for HRNNs with unbounded activation functions and timevarying delays in the leakage terms. Neural Process. Lett.. 39, 69–79 (2014). Publisher Full Text

Liu, B, Gong, S: Periodic solution for impulsive cellar neural networks with timevarying delays in the leakage terms. Abstr. Appl. Anal.. 2013, (2013) Article ID 701087

Liu, B, Shao, J: Almost periodic solutions for SICNNs with timevarying delays in the leakage terms. J. Inequal. Appl.. 2013, (2013) Article ID 494

Peng, L, Wang, W: Antiperiodic solutions for shunting inhibitory cellular neural networks with timevarying delays in leakage terms. Neurocomputing. 111, 27–33 (2013)

Xiong, W, Meng, J: Exponential convergence for cellular neural networks with continuously distributed delays in the leakage terms. Electron. J. Qual. Theory Differ. Equ.. 10, 1–12 (2013)

Yu, Y, Jiao, W: New results on exponential convergence for HRNNs with continuously distributed delays in the leakage terms. Neural Process. Lett. (2013) 37
37
Publisher Full Text 
Zhang, H, Shao, J: Almost periodic solutions for cellular neural networks with timevarying delays in leakage terms. Appl. Math. Comput.. 219, 11471–11482 (2013). Publisher Full Text

Horn, RA, Johnson, CR: Topics in Matrix Analysis, Cambridge University Press, Cambridge (1991)